Last week was the who and the how much. What’s live, what’s piloted, what only exists as an announcement, what got switched off, and what an agent purchase costs when the bill arrives.

Both editions focused on the facts and were meant to give you a foundation so you are ready for any Agentic Commerce conversation you will likely have in the coming months. But neither handed you a procedure.

That’s this one.

The number one question I get asked when discussing agentic commerce in the workshops I run is “what’s our agentic commerce strategy?”, which most try to answer in one go. But if you ask it that way, you get one vague answer that satisfies nobody. Because underneath it sit four separate questions, each of which will give you a clean answer, a different owner, and its own deadline.

Separate them, and this stops being hard.

The habit underneath it

Both previous editions (part one and part two) rested on a single discipline, and it’s worth naming it before we build anything on top of it.

Refuse to collapse categories.

Part One refused to let a name answer a question about a job. MCP, ACP, AP2 and x402 mostly don’t compete. Sort them by the job they do and eight scary abbreviations become six boxes.

Part Two refused to let an announcement answer a question about a state. Live, piloted, announced, and withdrawn all get described as “launched.” They’re four different things, and one of them is a product that no longer exists.

The framework applies that same refusal four times, in order. Scope, Stack, State, Stake. Questions first, names second, the same way we did with the protocols.

Question one: Scope

Who is the purchase for, and how far does the agent go?

Two axes, both from Part One.

• Who it’s for: a person, or the software itself. Your agent buying a raincoat for a customer is one thing. Your agent buying the data it needs to finish a job is another, and no human will ever see that purchase.

• How far it goes: it suggests, it prepares, or it buys. Suggests means it finds three coats and your customer clicks. Prepares means filling the basket and waiting. Buys means the rules got set once, weeks ago, and the purchase happens later without anyone watching.

Six cells. Where you sit decides which rulebooks apply to you at all.

Sit in “suggests, for a person” and you need clean product data. Nothing else. No mandates, no scoped tokens, no directory listing. Sit in “buys, for itself,” and you need x402, with no chargeback to fall back on when it goes wrong.

• Your action. Write one sentence: we are in this cell today, and we intend to be in that cell by this date. Nine times out of ten, the exercise turns up money already committed to a cell the business doesn’t occupy.

Question two: Stack

Of the six jobs, which do you own, which do you rent, and which does nobody own?

The six jobs again. Use tools. Talk to other agents. Find goods and fill a basket. Show who’s asking and who said yes. Move the money. Count what was used and bill for it.

Three answers available per job, and only three.

Own means your cost, your control. Rent means someone does it for you, so write down the supplier, the switching cost, and who else could do it if that supplier raised its prices tomorrow. Nobody means one of two things: either you’re carrying that risk without having chosen to, or there’s a business sitting in the gap.

Job six comes back as nobody. For everyone. That’s the empty box from Part One, and it’s now a row on a form.

• Your action. Every rented row needs a named alternative next to it. Every “nobody” row needs a human being inside your company whose job it is, even where the honest answer is that you’re choosing to eat the risk.

Question three: State

For each thing your plan depends on, is it live, piloted, announced, or withdrawn?

Last week’s four labels, pointed at your own plan instead of somebody else’s slide. Each label comes with permission attached.

• Live. Build on it. Budget it.

• Piloted. Staff it. Don’t commit a launch date to it.

• Announced. Write it down. Budget nothing.

• Withdrawn. Study it hardest.

Withdrawn is the one people skip, and it’s the only category that tells you which constraint the press release was hiding. OpenAI had the vision right. What killed Instant Checkout was state sales tax, no live inventory feed, one-item-at-a-time processing, no baskets, and no promo codes. Walmart loaded around 200,000 products in and then took them out again because checkout inside ChatGPT converted about 3 times worse than a click-through to Walmart.com.

Numbers need the same treatment. When somebody quotes a willingness figure at you, ask who paid for the research and what the question actually said. Gartner found 11 percent. Forrester found 24. Visa’s own survey found 38, and a processor-sponsored study found 45. Visa’s own research also found that 60 percent wouldn’t allow any spending without approval first.

• Your action. Label every line in your agentic plan. Anything that wears the announced label comes out of the budget and goes on a watch list.

Question four: Stake

What does it cost, and who carries the loss?

Three buckets. This is the same three-state discipline, applied to money.

• Priced. Interchange, roughly 1.5 to 2.5 percent. Network assessment, about 0.14 percent. Processor margin, commonly quoted at 2.9 percent plus thirty cents. Then exactly one published AI platform fee, OpenAI’s four percent, which left the table when the product did. Add the build: $75,000 to $250,000 and 14 to 20 weeks for a large merchant integrating both checkout rulebooks directly, or close to nothing if you’re on Shopify, where acquisition costs through AI channels run 40 to 60 percent below paid search. Every line here is modellable.

• Unpriced. Agent directory listing. Agent registration and the identity checks around it. Agent-specific interchange, of which Visa and Mastercard have published none. Nothing published is the same as nothing charged, and your acquirer is the person to ask.

• Unassigned. Then there’s the question of who pays when the agent buys the wrong thing, which nobody has answered. Today the merchant absorbs it through the ordinary chargeback system. Ask the industry, and it splits: in Darwinium’s survey of fraud and risk professionals, 39 percent said the AI provider, 20 percent the customer, 14 percent the merchant, 11 percent the bank, and 15 percent said share it. The FCA named this a live policy question in March 2026. HM Treasury’s consultation closes on 6 October. PSD3 is still being negotiated.

• Your action. Model the priced. Ask for the unpriced. Contract for the unassigned, because that’s the only line here a spreadsheet can’t save you from.

The output: one page, five lines

A framework earns its keep when it ends in something you can put in front of a board.

Five lines. If you can’t fill them in, that is the finding.

Running it once

A mid-sized retailer, forty million a year online, sitting on Shopify.

• Scope: suggests, for a person. They want to reach payments inside a year.

• Stack: they own nothing. Jobs three through five are rented from Shopify and their processor. Job six belongs to nobody, and that costs them nothing, because they don’t sell anything metered.

• State: the plan depends on four things. Two are live, one is piloted, one is announced. The announced one leaves the budget that afternoon.

• Stake: modeled cost lands near three percent, since no platform fee applies on the surfaces they care about. Two cost lines are unpriced, so they email their acquirer. Liability is unassigned, and as the merchant, they’re the ones carrying it.

An afternoon’s work. The plan it produces matters less than the announced line item they were about to fund.

The same four questions apply to everybody else, and only the answers change. For an acquirer or PSP, question two is where the strategy actually lives: which jobs do you own on behalf of your merchants, and which do you resell with a markup? For an issuer, question one settles whether scoped tokens are this year’s problem or next year’s. For an AI platform, job six is the only row that isn’t already somebody else’s.

What it rules out

A framework that permits everything is decoration. This one says no to four things.

• Don’t buy authorization infrastructure for a cell you don’t occupy.

• Don’t count a rented job as a capability.

• Don’t budget against an announcement.

• Don’t quote a willingness number without reading the question that produced it.

Cancel one line item after reading this, and it did its job.

Where we’ve got to

Three editions, three tools. A map of the jobs. A way of reading claims. And now a procedure that turns both into decisions your finance director will actually sign.

One rule to finish on. You only get to be early in one place, so pick the square on purpose, rather than landing there because a vendor turned up with a good deck.

And keep an eye on job six. Five jobs got standards, and most of those got handed to organizations the donors don’t control. The sixth is still an empty box… empty boxes in payments have never stayed empty for long.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

P.P.S. If you want to learn about the Vision and Strategies of fast-growing payments companies, check out my podcast called Executive Conversations with the Paypers

In that newsletter, we discussed the how. What an agent is, the six jobs that have to happen for one to buy you a raincoat, the eight or so names that fill those jobs, and the three ideas underneath them all.

This time I want to discuss the who and the how much. Which of these things you can actually use today, which are still in a lab, which exist only as a press release, and which have quietly been switched off. Then what an agent purchase costs, and where the market has settled after a year and a half of building.

As you are probably just getting back to the office, don’t worry, I won’t expect you to know everything. So no prior knowledge assumed. If you missed Part One, check it out first; it’s worth ten minutes. However, if you are up to speed, let’s continue.

First, four words worth keeping straight

The biggest challenge in payments is reading announcements because everything is described in the same excited tone. A specification that exists as a document, a pilot with forty transactions, a new alliance, and a product serving millions all get the word “launched.”

So to make it easier to understand, instead of names, let’s give it four labels. I use these constantly, and they save a lot of confusion.

• Live means you can use it today, at scale, without asking permission.

• Piloted means it works, but only for selected partners in selected countries.

• Announced means a company has said it’s coming, and nothing more.

• Withdrawn means it shipped and was then taken away.

Especially the last category is interesting, as we rarely hear about it, as nobody sends out a press release when you decide to stop.

What’s actually live

Let’s start with the good news first, because there’s a lot of it.

Google’s checkout on its own surfaces is live. Merchants including Nike, Sephora, Target, Ulta, Walmart and Wayfair are named participants, and it runs on the Universal Commerce Protocol, the shop-end rulebook Google wrote with Shopify.

Microsoft’s Copilot Checkout has been live in the US since January 2026, with PayPal and Stripe handling the payments. Perplexity’s Buy with Pro has been live since November 2025 and charges merchants no commission.

Amazon’s Alexa for Shopping is live, including a feature called Buy for Me that buys from other people’s shops on your behalf. By March 2026 it covered more than a hundred million products across four hundred thousand merchants.

On the payments side, Mastercard’s Agent Pay is generally available to US cardholders through certified processors. Visa and Cloudflare’s Trusted Agent Protocol, the doorman that lets an agent prove which agent it is, shipped as an open specification with working reference code and twelve launch partners.

And x402, the machine-to-machine rail, is live and open, with no protocol fee.

Most of these were just fancy slides a year ago, and now they are considered a real stack, the Agentic Payments Stack.

What’s still in the pilot phase

Visa’s Intelligent Commerce is the big one here. It’s had hundreds of controlled live transactions, has more than thirty banks live behind it in Europe, and works at named merchants including lastminute.com and Frasers. That’s genuine, and it’s also not something you can just switch on.

Google’s Universal Cart, which carries a single basket across Search, Gemini, and eventually YouTube, is rolling out in the US, with other countries listed as coming.

American Express opened its developer kit in April 2026, but only to agents it has validated. Mastercard’s Agent Pay for Machines was announced in June 2026 with 30-odd named participants, a strong list for an early product.

What exists as an announcement

Three worth knowing, because you’ll see them quoted as though they’re finished.

Mastercard’s Agent Suite was announced in January 2026 for delivery in the second quarter. Amex’s Agent Purchase Protection, which promises to cover cardholders when an agent buys the wrong thing, is still written in the future tense on Amex’s own website, and the register of approved agents is marked as coming soon.

And Google’s Agent Payments Protocol, the permission layer we covered last time, is a finished specification that has not yet reached Google’s own consumer products. The plan is Gemini first, then others.

I don’t say this as a criticism, but rather to note that it’s a different state from live, and the difference matters if you’re planning against it.

What got switched off

The biggest surprise thus far comes from a company most of us would probably bet on to succeed.

OpenAI’s Instant Checkout, which arrived on 29 September 2025 and let you buy from Etsy sellers without leaving ChatGPT. It was switched off in early March 2026 by OpenAI, which took it away and moved ChatGPT back to being a place where you find things and then buy them elsewhere.

The reasons OpenAI gave were practical rather than philosophical. No system for handling state sales tax. No live inventory feed. One item at a time, so no baskets, and no promo codes.

Walmart had put around 200,000 products into it and pulled them out, too. Its own measurement showed checkout inside ChatGPT converting about three times worse than a normal click-through to Walmart.com. Interestingly, the same traffic brought roughly twice as many new customers, so Walmart kept the discovery and dropped the checkout.

Amazon retired its Rufus shopping assistant in May 2026 and folded it into Alexa. Rufus wasn’t a flop; Amazon credited it with nearly $12 billion in additional annualized sales; it was just absorbed.

As of this week, there’s no in-chat checkout replacement at OpenAI that you can integrate with.

Meanwhile, the front half is booming

The discovery part of Agentic Commerce is the area that hasn’t slowed down at all.

Adobe Analytics watches more than a trillion visits to US retail sites. It measured traffic arriving from AI assistants up 693 percent over the 2025 holidays and up 393 percent in the first quarter of 2026. Adobe sells analytics software, so I would take it with a grain of salt, as they benefit from reporting positive numbers to their clients.

The number I’d actually keep is the conversion one. In March 2025, visitors sent by an AI assistant converted 38 percent worse than those sent by other channels. By March 2026 they converted 42 percent better.

So people are perfectly happy to let a machine help them shop. What they haven’t done in volume is let it pay.

That’s the shape of the market right now, and it has a name in the trade: discover in the AI, buy on the merchant’s own site.

What an agent purchase costs

Now let’s discuss my favorite part, the money, which is where I see most of the interesting gaps are.

In my newsletter Agentic Economics, I provide a very in-depth breakdown of who captures value in Agentic Commerce, but here is a short overview.

Start with what you already pay on a hundred-dollar card sale. The issuing bank takes interchange, roughly 1.5 to 2.5 percent on a US credit card that isn’t present. The card network takes an assessment fee of about 0.13 to 0.15 percent. The processor takes its own margin, commonly quoted at 2.9 percent plus thirty cents.

An agent transaction rides those same rails at those same rates. Then one new layer sits on top.

OpenAI wanted to charge Shopify merchants four percent on completed Instant Checkout sales. Up until now, that is the only published AI platform fee anyone in this field has- however, not officially, but in their own CEO’s words, whose own phrasing was “a small fee.”

Everyone else currently charges nothing extra. Google adds no platform fee for AI checkout. Microsoft adds none. Perplexity takes no commission and covers the cost of free shipping, yet.

So the honest summary is that the new layer exists, it has been priced exactly once, and that price came off the table when the product did. So until now, it is still unclear how value will actually be captured, but believe me that train will always arrive on time.

The parts nobody has priced

Three costs are simply not public, and it’s worth knowing which.

There are no published agent-specific interchange categories from Visa or Mastercard. Agent transactions are described as riding existing tokenized rails at standard rates. There are no published fees for registering an agent in a network’s directory, and none for the identity checks that go with it.

Nothing published isn’t the same as nothing charged. It means the question is still open, and your acquirer is the person to ask.

What is public is the build cost. A large merchant integrating both major checkout rulebooks directly is reported to cost somewhere between $75,000 and $250,000, or roughly 14 to 20 weeks for a small engineering team. Those figures come from commerce vendors rather than audited accounts, so treat them as directional.

For the million-plus merchants on Shopify, that cost is close to zero, because the platform did it for them. That single decision is probably the most consequential thing anyone has done in agentic commerce so far.

On the upside, Shopify reports that customer acquisition costs through AI channels are 40 to 60 percent lower than paid search. Early, vendor-stated, and the main reason merchants keep showing up.

How the AI platforms are actually making money

So how are AI platforms currently making money off of shopping?

OpenAI began showing ads in ChatGPT in February 2026 in the US to logged-in adults on its free tiers. Self-serve ad buying launched in May, and in August, OpenAI confirmed that ChatGPT Ads are running in the UK, Mexico, Brazil, Japan and South Korea.

So the money at the AI layer is currently coming from the recommendation rather than the transaction. A merchant pays to be found, not to be bought from.

If that sounds familiar, it should. It’s how search has worked for twenty-five years.

The machine side, briefly

The other half of this world is software buying from software, which I covered in July, so here’s just the current state.

x402 works, costs nothing to use, and settles in seconds for fractions of a cent. Average payment size is around thirty cents, which is a ticket size cards genuinely cannot serve.

The volume, though, has fallen a long way. Daily settlement peaked near a million dollars in late 2025 and now sits around thirty thousand, down 93 percent so far this year according to analyst Jamie Coutts using Helios Analytics data. Much of the earlier activity appears to be testing rather than commerce.

The plumbing is finished, but the customers haven’t arrived yet. That’s a normal stage, and it’s worth watching rather than betting on.

Who pays when the agent gets it wrong

This one has no answer yet, and I want to be clear that I mean no answer, not a complicated answer.

Today the merchant absorbs it, through the ordinary chargeback system. No rule anywhere assigns fault between the shopper, the AI platform, and the shop.

There’s also an evidence problem. When a merchant defends a dispute, it normally submits proof a human was present: the device, the location, the path through the site. In an agent purchase, the agent produced all of that.

Ask the industry who should carry it, and you get a genuine split. In a survey of 500 fraud and risk professionals by Darwinium, 39 percent said the AI provider, 20 percent the customer, 14 percent the merchant, 11 percent the bank, and 15 percent said: "share it."

Regulators have noticed and haven’t ruled. The Financial Conduct Authority named agentic payments a live policy question in March 2026. HM Treasury has a consultation open until 6 October. PSD3, the next version of Europe’s payment rules, is still being negotiated.

If you’re a merchant, this is the line item to watch. Everything else in this edition is a cost you can model.

What people say they’ll allow

Finally, the demand question: you’ll be shown a number, and the numbers will disagree wildly.

Depending on the study, between 9 and 45 percent of consumers say they’d let an agent complete a purchase. That’s not a range, it’s a warning label.

Sort them by who paid for the research, and it settles down. Gartner, an independent firm, found that willingness to let AI make purchase decisions topped out at 11 percent. Forrester, an independent research firm, found that 24 percent trust agents with routine purchases. Visa’s own survey found 38 percent, and a study sponsored by a processor found 45 percent.

The rest of the gap is question wording. Visa’s own research found that 60 percent wouldn’t allow any spending without approval first.

Read the actual question before you quote the number. That’s the whole trick.

Where we’ve got to

Discovery is live, growing fast, and converting better than the rest of your traffic. Autonomous checkout inside an assistant is piloted, gated, or, in OpenAI’s case, switched off, and the market has settled on discovery in the AI and buy on your own site.

An agent purchase costs the same as a card purchase, plus a platform fee that is priced exactly once. The identity and directory fees haven’t been published at all.

Nobody has decided who pays when an agent buys the wrong thing. Machines can pay each other perfectly well and mostly aren’t.

That’s the state of play in August 2026. Your homework is unglamorous and genuinely worth doing: go and look at your product feed, because in a market that discovers in the AI and buys on your site, the feed is the whole shop window.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

Based on figures shared by OpenAI, Anthropic, and Google, pretty much everyone in the world has used AI in one way or another.

From typing in a question in Google and the top search result being Gemini’s AI-generated answer, to using ChatGPT or Claude. Maybe you asked it to compare two washing machines, plan a weekend, or find a birthday present for someone difficult to buy for.

Now change one thing. Give it a card.

That’s agentic commerce, and it’s what the entire payments industry has spent the past year and a half building toward.

Over the next two editions of this newsletter, I want to walk you through it from the ground up. What an agent actually is, what all the three-letter names like MCP, A2A, ACP, x402 mean, and why a system that handles trillions of dollars a year suddenly needed rebuilding.

This first edition is more about the how; the next one is the who, and the strategies that you need to deploy to ensure you are a part of one of the biggest changes in commerce and payments in decades.

No prior knowledge assumed. Let’s go.

First, what is an agent?

The word gets thrown around loosely, so here’s a working definition you can hold on to.

• A bot follows a script somebody wrote. It checks a price every morning at nine. It can’t do anything else, because nothing else was written down.

• A chatbot talks to you. It answers, suggests, explains. Then it stops, because talking is all it does.

• An agent is given a goal, not a script. You say “find me a decent raincoat under eighty dollars,” and it works out the steps itself. It searches, compares, discards, and decides. Nobody wrote those steps down in advance.

The last one is the one we will be focusing on. A bot does what it was told. An agent works out what to do.

Which is fantastic when it goes well, and is exactly why the payments industry got nervous.

Three ways this plays out

Agentic shopping isn’t one thing. It’s a spectrum, and where you sit on it changes everything about the plumbing underneath.

Most of what exists today is the first kind. The agent finds things, and you buy them. That’s already happening at scale, and it doesn't need any new payment technology at all, because you’re still the one clicking.

The second kind is where it starts getting interesting. The agent does the tedious part and hands you a finished basket. You approve. Still recognizably shopping.

The third kind is the one everything is being built for. You say once, in advance, “book it if the fare drops below four hundred,” and then you get on with your life. The purchase happens later, without you, and the shop has to somehow satisfy itself that you agreed to it.

The keyword here is Later. As it is the root cause for most of the problems I’ll discuss.

A distinction worth getting straight

It is important to note that when it comes to Agentic Commerce, there are actually two things that wear the same name, and the further we get, the more confusing it can become if you don’t distinguish between them both.

When an agent buys you a raincoat, that’s agentic commerce. A human wanted something, a machine fetched it, a shop sold it. The money ends up where it always did.

When software buys something for its own use, that’s different. An agent needs a piece of data to finish a job, so it pays a few pence for it. No human wanted that data. No human will ever see it.

The tempting mistake is to think that the difference lies in how independent the agent is. It isn’t. A fully autonomous agent buying you a raincoat is still shopping for a person. The same agent buying the computing power it runs on is not.

So it is always important to ask, who the purchase is for.

I wrote about the second kind in detail, called Machine-to-Machine Payments, a few weeks ago; however, let’s focus on the first one here.

Why the old system can’t cope

For most people who don’t work in payments, this might come as a surprise.

The card system is genuinely good at its job. It’s been hardened over decades against every kind of fraud, and the reason your card mostly works and mostly doesn’t get stolen is a large pile of unglamorous engineering.

But all of that engineering rests on one assumption, and once you see it, it’s hard to unsee.

It assumes a person is there.

Have a look at the list above. Every single one of those checks is asking, in a different accent, the same question: is there a human here, and is it the right human?

Take the human away, and each check either breaks or, worse, fires wrongly. The device isn’t recognized because it’s a data center somewhere. The location makes no sense because the computer is nowhere near you. The shopping pattern looks nothing like yours because it wasn’t made by a person.

The system doesn’t just fail to protect you. It starts blocking things you actually wanted.

And the puzzle is genuinely hard. If you can’t recognize the person, and you can’t recognize the device, and the behavior tells you nothing, what exactly are you checking?

The industry’s answer, which we’ll get to, is a proper change of philosophy. Stop trying to guess. Start asking for proof.

What a protocol is, in plain English

Before we dive into the names, let’s make sure we get at least one bit of vocabulary right, because it’s the word that does the most work in this field and it’s rarely explained.

What is a protocol?

A protocol is an agreement about how two systems talk, that neither of them owns.

Email is the example everyone can feel. Gmail and Outlook are competitors built by rivals in different decades, and a message still goes from one to the other without anybody negotiating. That works because both agreed long ago to a shared set of rules for handing over mail.

Nobody owns those rules. That’s the point of them.

If email had no protocol, every provider would need a private arrangement with every other provider. Ten providers, forty-five separate deals. It would have collapsed under its own weight before it started.

That’s the situation agentic commerce was heading for. Every AI assistant would need a private integration with every shop. So the industry started writing protocols instead.

Rather a lot of them, as it turns out.

The six jobs

If you want to be able to differentiate between the names, I suggest focusing on the job that needs to be done to remember them.

Think about what has to happen for an agent to buy you a raincoat, and you get six distinct tasks.

Every step is a separate job, which is important to know when working with Agentic Commerce. Finding a raincoat has nothing in common with proving you agreed to buy it, which in turn has nothing to do with moving the money.

Different jobs need different agreements. That’s why there isn’t one protocol, and why there was never going to be.

Now, let’s break down the names.

A lot of abbreviations, but essentially, this is where we are right now.

The most interesting part of it is probably the last one. Five jobs have at least one answer, and the third and fourth have several competing. The sixth has nothing at all. We’ll come back to why.

The names, briefly

Now they have jobs, the names are easy. One paragraph each.

• MCP lets an agent use tools. Before it existed, connecting an assistant to some outside service meant custom work every time. MCP made that a standard connection, the way a plug fits a socket. Anthropic built it and then handed it to the Linux Foundation in December 2025, a neutral home for shared technology.

• A2A lets agents find and talk to each other. One agent posts a card detailing what it can do; another reads it and passes the work over. Google built it and gave it to the Linux Foundation in June 2025.

• ACP and UCP both do the shop end: read the catalog, build a basket, run a checkout. ACP came from OpenAI and Stripe in September 2025. UCP came from Google and Shopify in January 2026. They overlap heavily, and yes, that means a shop wanting to reach everybody currently supports both. More on that next time.

• AP2 handles permission. It’s the “you said yes” layer, and it’s how a shop can tell, months later, exactly what you authorized. Google built it, then handed it to the FIDO Alliance in April 2026, with about sixty organizations signing on. Mastercard donated a similar piece of its own alongside it.

• Web Bot Auth and Visa’s Trusted Agent Protocol both answer “which agent is this?” They let an agent prove its identity at the door, so a shop can let good ones in rather than blocking everything automated.

• x402 moves money between machines. It revived a status code that the web’s designers reserved in the 1990s, labeled “payment required,” and then left unused for thirty years. Coinbase built on it and handed it to the Linux Foundation in July 2026.

For now, that is all of them. Depending on when you are reading this, it may already have been updated or changed. But as of August 2026, this was the protocol stack. So, if you remember nothing else, remember that they mostly don’t compete. They stack.

The three things underneath

Strip away the names, and there are only three real building blocks. Once these click, the rest is decoration.

The signature is the simplest. The agent has a secret key, publishes the matching public one, and signs each request it sends. The shop checks the signature against the published key. Match, and the sender is who it claims.

It’s a wax seal. It proves who sent the letter. It says nothing about what’s inside or whether they were allowed to send it.

The one-shot card number solves an obvious fear. You do not want to hand a card number to software that talks to strangers on the internet.

So nobody does. Instead, the bank issues a stand-in that works exactly once, at a named shop, for up to one amount, and then dies within minutes. If it leaks, it’s worthless. If the agent tries to spend more, it bounces. If it turns up at a different shop, it bounces.

A gift card, essentially, minted on the spot for a single purchase.

The mandate is the newest idea and the most important. It’s a signed record of what you actually agreed to.

Say you tell your agent to buy tickets for under $200. That instruction gets written down and signed with a key on your phone, before anything is bought. Later, when the agent purchases, the signed slip accompanies the transaction. Anybody in the chain can check it.

So when the question comes up months later, and it will, nobody is arguing about what you said. The answer is written down and signed.

Putting it together

One purchase, using all three.

Note what happens at the end. The shop charges the one-shot number the way it charges any card, over the same rails it has always used.

The shop stays the shop. It’s still the one who sold you the raincoat, still the one who owes you a refund, still the one your bank argues with if it goes wrong. The AI assistant is the messenger, not the merchant.

That was a deliberate choice, and a smart one. Rebuilding the world’s payment rules would have taken a decade. Slotting into the existing ones took a year.

The box nobody has filled

Back to that sixth job, because it’s the one I’m keeping an eye on for now.

Counting what an agent used, pricing it, and putting it on a bill. There’s no standard. Not a weak one, not a competing pair. Nothing.

Which is strange when you notice that the other five all got standards fast, and most got donated to neutral organizations that the donors don’t control.

I’ll leave you with the question rather than the answer, because it’s a good one to chew on between now and next time.

Why would fiercely competitive companies give away five pieces of this and not the sixth?

Where we’ve got to

An agent is software given a goal instead of a script. Agentic commerce is that software buying things for you, and it breaks the payment system because every safety check quietly assumes a person is present.

The fix was a set of shared rulebooks, one per job. Six jobs. Roughly eight names, which now, you have a reference to that includes six boxes, and an explanation of what those boxes are and what they do.

Underneath, three ideas: a seal that proves who’s asking, a card number that only works once, and a signed slip proving what you agreed to.

Next time: I’ll dive into who has actually shipped what, what’s running versus what's just a press release, and what an agent purchase costs when the bill arrives.

Enjoy the sunshine, stay hydrated, and share this with someone who needs to step up their Agentic Commerce game.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

Then he stayed (delayed) his own ruling, so that nothing would actually change.

In the twelve months since, the industry has been busy trying to buy the thing he’d just called unlawful.

Capital One closed on Discover in May 2025 and started moving its debit cards onto Pulse the month after. This July, four of the largest banks in the country were reported to be circling Fiserv’s STAR network at about $15 billion.

They’re buying the same thing, and it isn’t technology.

It’s an exemption from a rule a court has already said the Fed had no authority to write.

Let me break it down.

The difference between SMS and DMS

Let’s start with the part that made sense first.

Between 1976 and 1985, American banks had a real problem. Customers wanted cash from machines, and no single bank owned enough machines to be useful. Building a national ATM footprint alone was absurd. Sharing one was the more obvious strategy.

So they built co-operatives. Shazam in Johnston, Iowa in 1976. Pulse in Houston around 1981, formed by Texas banks. STAR in 1984, eleven banks and savings-and-loans pooling their machines. NYCE in the northeast in 1985. Each one owned by the institutions that used it, each one mutualizing a cost none of them wanted to carry alone.

Then the same switch proved effective at the checkout, and PIN debit was born.

For those who might not know;

• A single-message (SMS) transaction is requested and settled in one go, which is what happens when a customer enters a PIN.

• A dual-message (DMS) transaction authorizes first and clears later, which is what happens when they sign or tap. The domestic networks grew up on the first kind.

Visa and Mastercard grew up on the second. Almost everything that followed traces back to that split.

At the peak, there were more than a hundred of these networks. By the mid-1990s, around fifty. Today you can count the ones that matter on one hand.

The position wasn’t taken. It was sold.

Ownership went first, and the members handed it over willingly.

Concord EFS bought STAR in 2001. First Data agreed to buy Concord in April 2003 for roughly $7 billion in stock. The DOJ, eight states and DC sued that October to block it, and the parties settled on 15 December, the day trial was due to start, with First Data forced to divest its entire 64% stake in NYCE. The deal closed in February 2004.

Discover bought Pulse for $311 million, with the deal closing in January 2005. Pulse’s 4,100-plus member institutions voted it through. This was the network that had spent two decades marketing itself as the one that answered to banks and credit unions rather than Wall Street.

NYCE went to Metavante, then to FIS when it bought Metavante for $2.94 billion in 2009. Fiserv took First Data for about $22 billion in 2019, which put STAR and Accel under one roof.

Nobody was robbed here. Members got liquidity, buyers got scale, and every board could show its shareholders a number. The position was sold, not lost, and that distinction matters, because it means the networks understood what they had and priced it anyway.

Brand went second, and that one was taken.

On the other side, Visa and Mastercard owned the brands that the customer actually saw. The domestic networks became a routing decision made by a merchant, invisible to everyone else. The numbers show how far that went: single-message volume fell 2.2% between 2021 and 2022, the first absolute decline ever recorded, and by 2023 dual-message networks carried 71.4% of US debit transactions.

Then the mandate arrived, and it arrived backwards.

In 2011, Regulation II required that every debit card operate on two unaffiliated networks. Congress handed back a form of distribution that the networks had already given up commercially. They didn’t win it, they didn’t build it, and they can’t defend it. It’s a statute, and statutes move.

How thin the trade actually is

On regulated debit, interchange barely differs by network. In 2023, the covered rate was $0.22 per transaction on dual-message and $0.24 on single-message. That’s it. The cap flattened the thing merchants were supposed to be shopping for.

So the savings from routing don’t come from interchange. They come from network fees. Visa and Mastercard assessments run around 13-14 basis points, plus roughly 1 for licensing. A PINless switch fee runs around 10, often with no license fee at all.

A few basis points. That’s the arbitrage.

Now look at who pays. Total network fees came to $12.95 billion in 2023, and merchants and acquirers covered 64.9% of them, up from 44.3% in 2009. Over the past fifteen years, the networks have quietly shifted most of their fee burden onto the merchant side of the transaction.

The routing behavior follows the arithmetic. Break-even sits around a $15 ticket, because PIN networks charge less as a percentage and more as a fixed fee. Grocery, big-box and fuel route away hard. Coffee shops and drive-throughs stay exactly where they are.

CMSPI puts the total US routing opportunity at roughly $4 billion a year, rising to about $5 billion if every card were PINless-enabled, which they aren’t. As of October 2024, only 51.1% of US debit cards had PINless switched on.

Set $4 billion against $4.66 trillion of US debit value in 2023. That’s under a tenth of a percent of the money moving. It’s real money, and it is astonishingly thin, and it’s the entire commercial case for the routing mandate that keeps these networks alive.

It’s worth being clear about who is even making this choice, because it isn’t the network. The merchant sets the routing policy. The acquirer or gateway executes it. A specialist vendor usually designs the logic, and the system selects the cheapest available option for each transaction, limited to whatever the issuer has provisioned on the card.

So a domestic network competes on price, into a decision it doesn’t participate in, on a spread of a few basis points, for volume its own issuer clients can switch off. And in card-not-present, it competes with token vaults that Visa and Mastercard control, which is a constraint the domestic networks have been debating for years and haven’t resolved.

What Capital One actually bought

Capital One closed the Discover acquisition on 18 May 2025 for $51.8 billion in purchase consideration and told investors to expect $1.2 billion in network synergies in 2027. Migration of its debit portfolio onto Discover and Pulse began that June, card by card.

Here’s the mechanism, because it’s the hinge of everything that’s followed. Reg II caps what a covered issuer, meaning one with $10 billion or more in assets, can earn per swipe. It doesn’t cap what a network charges when the issuer and the network are effectively the same house. A bank that owns the rails its own cards run on can argue its way out of the cap, and that argument is worth a great deal of money.

Once a card moves, it leaves the cap behind.

Unregulated card-present debit at roughly 1.10% plus $0.16, against the covered cap of 0.05% plus $0.22. On a $50 sale, that’s about 71 cents instead of about 24 and a half cents. Nearly three times the economics on the same swipe, the same rails, the same fraud logic.

SRM sized it at around 7 million cards and $60 billion of purchase volume, with the exemption alone worth more than $350 million a year.

Nothing about the switch got better. Capital One didn’t buy a faster network. It bought permission to charge more for the same one.

There’s a bill for that, and Pulse pays it. A network that spent 40 years selling neutrality to banks and credit unions now belongs to a top-six bank that competes with many of those banks and credit unions.

My read is that this is the real cost of the deal and the one nobody has priced yet, because no disclosure yet shows what third-party issuers are doing in response. That gap is genuine, and I’d watch it more closely than the synergy number.

So why are the US Debit Networks interesting again?

On 7 July, the Wall Street Journal reported that Fiserv had held preliminary talks with JPMorgan, Bank of America, Wells Fargo and PNC about selling STAR, possibly with Accel, at a valuation of about $15 billion.

All just speculation, of course.

However, the asset underneath is real enough. STAR runs around three billion transactions a year for more than 115 million cardholders across 2,800-plus financial institutions, including 24 of the top 50 US issuers. KBW put the prize for the ten largest debit issuers at $460 million to $3.5 billion in extra gross revenue. William Blair called a deal unlikely, citing regulatory and merchant pushback and the risk of alienating Fiserv’s community bank and credit union base.

But look at the structural difference between this and the Capital One deal.

Capital One bought one company and controls it outright. It sets its own routing, migrates its own cards, and answers to no one about pricing. A consortium has to agree. Four owners, repeatedly, about how to price a network that thousands of competing institutions route through.

The precedent cuts both ways. Zelle works, and it works beautifully: more than $1.2 trillion in payments and over 150 million enrolled accounts, distributed within banks’ own apps, where the owners control the shelf. Paze is the other outcome. It auto-enrolled 200 million-plus cards and still had to sign distribution deals with Fiserv, Worldpay and Nuvei in 2025 to get accepted anywhere.

A debit network is a merchant-acceptance business. That’s the side bank consortia have historically been worst at, and I don’t see what makes this attempt different.

Then there’s the room the deal would be filed in. The DOJ is currently suing Visa for monopolizing the debit market, alleging that it handles more than 60% of US debit transactions and collects over $7 billion a year in processing fees. Handing the four largest issuers a route around the interchange cap is a difficult transaction to bring into that environment.

One more thing worth flagging plainly. Fiserv doesn’t break out STAR or Accel anywhere. They sit inside Financial Solutions with no separate revenue or margin disclosed. The $15 billion comes from deal chatter, not a filing, and no one outside the room can tie it to any official documents.

The thing nobody buying can control

But let’s get back to North Dakota.

On 6 August 2025, Judge Traynor vacated Regulation II in full, holding that the Fed had gone beyond its statutory authority in deciding which costs the cap could recover. He then stayed his own vacatur to stop interchange becoming, in his words, a completely unregulated market.

The Fed appealed to the Eighth Circuit that October. It filed its opening brief on 30 December, Corner Post responded in February, and the Fed’s reply landed on 23 March. Briefing has closed. As far as I can tell, no oral argument date has been posted yet.

So the fork runs in three ways, and each one changes what a network is worth.

Vacatur affirmed. The cap disappears, and covered interchange rises for everyone. The exemption stops being scarce, which means the thing a buyer paid a premium for becomes the default setting. You’d have bought a privilege on the eve of it becoming ordinary.

Vacatur reversed. The cap survives, and the Fed’s November 2023 proposal to cut the base from $0.21 to $0.144 remains unfinalized and frozen. Finalize it, and the gap between capped and exempt widens. The exemption gets more valuable, not less.

Nothing happens for years. Which, given appellate timelines and a possible return trip to the Supreme Court, is the base case.

There’s a wildcard on top. The Credit Card Competition Act returned on 13 January 2026 as S.3623, with a House companion, and was endorsed by the President the same day. It’s still in committee, and an attempt to attach it to the CLARITY Act at the end of that month failed. If credit routing ever becomes law, every network that isn’t Visa or Mastercard reprices overnight.

And the Visa case has slipped since the July reporting. Under the current schedule, fact discovery now closes on 16 October 2026 with expert discovery running to April 2027, which pushes any trial into 2027 or 2028.

Add it up. Whoever buys a debit network today is underwriting an appellate panel, a frozen rulemaking and a bill in committee. None of that is a business you can operate your way out of.

The case against my own argument

Two things cut the other way, and I want to make sure I highlight them properly.

The first is that PIN debit is genuinely better plumbing. This isn’t sentiment, it’s in the Fed’s own 2023 data. Single-message transactions carried the lowest fraud losses at 6.7 basis points of value against 17.9 for dual-message transactions, and the lowest authorization, clearing, and settlement costs at $0.030 against $0.039. Cheaper and safer, measured rather than claimed. If the networks were purely a compliance artifact, that wouldn’t be true.

They’ve also moved where they were weakest. The 2022 clarification took effect in July 2023 and forced PINless routing into e-commerce. Single-message share of card-not-present volume went from 6.1% in 2021 to 6.6% in 2023, and vendor One Inc reported a 50% rise in debit routed PINless to STAR, NYCE and Pulse between February and September 2023. The direction is right; however, the slope is shallow.

The second is Shazam. Founded in Iowa in 1976, still independent, still member-owned, still not-for-profit, and it marked fifty years in 2026. It never sold, and it still serves the community banks and credit unions that the big consolidators treat as an afterthought.

That’s an existence proof: a service model can hold a position that scale can’t buy.

It’s also about $191 million of revenue, which tells you the size of the ceiling on that model. Both things are true at once, and anyone arguing this industry from one of them is only half right.

What the survivors kept

To learn what can happen, we can look at other regions. The sharpest evidence isn’t in the US at all.

Look at the domestic schemes that thrived. Interac in Canada, founded the same year as STAR. Girocard in Germany, with more than 100 million cards and around 31% of German non-cash payments in the first half of 2025. Cartes Bancaires in France, Bancontact in Belgium.

Every one of them kept three things. Member or bank ownership, so the scheme answered to its users. A mandate or default routing that regulators actively protected. And the brand on the front of the card, so the consumer knew whose rails they were on.

The US networks kept none of the three. They sold ownership, lost the brand to Visa and Mastercard, and then received a routing rule from Congress that they neither built nor control.

Even holding all three isn’t permanent, and it’s worth being honest about that. Nilson’s 2024 Canadian data shows Interac’s share of purchase transactions falling 191 basis points to 37.63%, with Visa at 36.50%, and its 2025 reporting has Visa edging ahead. The strongest domestic scheme in the world is losing ground too. It’s just losing it from a position it actually owns.

Brazil and India went further, skipping the argument entirely. Pix and UPI are rails the regulator built, and adoption followed the mandate rather than the marketing.

Ownership, mandate, brand. Three legs, and the American networks are standing on a borrowed one.

What this means for whoever ends up holding it

When a business earns from what it does, you can make it better. Invest, price, ship, compete.

When it earns from what it’s allowed to do, there’s nothing to improve. You can only wait, and hope the permission holds.

That’s the trade on the table right now. Capital One made it and is collecting on it today, which is a perfectly good reason to have made it. The banks reportedly looking at STAR would be copying a deal whose main input is a rule that a district court has already vacated and an appellate court hasn’t yet ruled on.

Fiserv, for its part, would be selling infrastructure at the exact moment its own market value sits near $29 billion, roughly 70% below where it stood in March 2025. That’s not the behavior of a seller with options.

These networks solved a genuine problem in 1984 and got properly paid for it. Everything since has been an argument about who gets to hold the permission that replaced the advantage.

A loophole isn’t infrastructure. It’s a lease, and somebody else writes the terms.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

By early February 2026, it was worth less than $40 billion, and for a few days it was worth less than eBay, the company that had spun it off six years earlier.

Nothing in the business collapsed.

Volume kept climbing. So did transactions. PayPal moved $486 billion in the second quarter of 2026 alone, $1.88 trillion over the trailing year, and still counts 439 million active accounts.

Venmo is growing in the mid-teens. Braintree is accelerating. Buy now, pay later volume rose 26%.

Then, in mid-July, Stripe and Advent International offered $53 billion to take the whole company private. The board declined their offer, saying it wasn’t good enough.

So this isn’t the story of a company that shrank. It’s the story of a company that kept nearly everything except the one thing that made the rest of it worth owning.

Let me break it down.

What PayPal actually owned in 2015

When eBay let PayPal go in July 2015, the market valued the parent at roughly $35 billion and the spinoff at roughly $49 billion. PayPal arrived with about 179 million active accounts and one genuinely scarce asset.

Not rails. Not a network in the Visa sense. What PayPal owned was consumer intent at the exact moment of payment.

A shopper who saw the PayPal button clicked it because, in 2015, typing a card number on an unfamiliar website was a small act of faith, and PayPal was the alternative to faith.

Merchants paid a premium for that button over raw card processing, and they did so willingly because the button converted browsers into buyers. That premium was the entire franchise. Everything else was infrastructure.

Two other things came in the box. Braintree, bought in 2013 for around $800 million, brought a developer-friendly gateway that already powered Uber and Airbnb. Braintree also brought Venmo, which it had picked up in 2012 for $26.2 million.

And one liability dressed as revenue: the eBay operating agreement, high margin, sitting on a clock that everyone could read.

Three assets, one franchise, and a known expiry date on part of the income. What PayPal is the crux of the matter.

The shopping years

Between 2015 and 2021, PayPal bought adjacencies at pace.

Xoom for remittances in 2015, around $890 million. Paydiant the same year. TIO Networks in 2017, later shut down after a breach. Hyperwallet in 2018 for around $400 million.

Then the big ones: iZettle in 2018 for $2.2 billion, closed that September. Honey in November 2019 for roughly $4 billion, PayPal's largest acquisition to date. Paidy in 2021 for about $2.7 billion. Curv for crypto custody, and Happy Returns, which was sold on to UPS in 2024 for around $465 million.

Most of these had a decent thesis on paper. Two of them show the pattern rather than the price.

Honey cost $4 billion and bought a browser extension that hunts discount codes. The logic was to move upstream, to be present when shopping starts rather than when it ends.

What PayPal got was a coupon tool with a loyal following, an awkward relationship with the creators whose affiliate links it sat atop, and litigation over attribution that ran from 2024 into this year.

It never became the front door. Nobody opens Honey and then decides to buy something.

Zettle is the sharper tell. PayPal closed the $2.2 billion purchase in September 2018 to get a physical presence in shops. Zettle launched in the United States on 30 June 2021. Just under three years to bring an existing, working product to the largest card market on earth.

Square didn’t spend those three years waiting.

Here’s the red lining running through all of it. PayPal bought capabilities and then ran them as separate brands. Braintree, Venmo, Xoom, Honey, Zettle, Hyperwallet: each kept its own product surface, its own login, its own risk logic, its own place in the org chart. The company assembled a portfolio and called it a platform.

Alex Chriss, PayPal’s former CEO, put it more bluntly than any outside analyst did when he arrived in 2023, telling investors he was working out which acquired units were core and which were, in his words, boat anchors. That is a striking thing for a chief executive to have to say about assets his own company chose and paid for.

The asset it owned and never priced

Venmo is the clearest case of the same habit, and the most expensive.

Venmo cost Braintree $26.2 million in 2012. By 2021 it was moving $230 billion a year. Its growth was never the problem. Its growth was extraordinary, viral, and essentially free, the kind of adoption curve companies spend billions trying to buy.

Look at when PayPal actually built a business on top of it. Pay with Venmo, the Venmo debit card, the Venmo credit card, business profiles: those arrived in a cluster around 2020 and 2021. Amazon acceptance came in 2022. Venmo has existed since 2009.

So the monetization machinery showed up roughly a decade after the product found its audience, and only once PayPal’s own growth had started to wobble.

The results now are genuinely good. Venmo TPV hit $93.8 billion in the second quarter, up 14% currency-neutral, the second straight quarter in the mid-teens. Venmo debit card monthly actives grew more than 50%. Pay with Venmo monthly active users grew by around 30%.

That’s a well-run asset. It’s also one that showed up about eight years late, into a market where Zelle, Cash App and Apple Cash had all had time to settle. PayPal owned the best consumer payments brand of the smartphone era and treated it as a feature for most of a decade.

What replaced the eBay revenue

The eBay agreement wound down from 2018 to 2023. PayPal knew the date. The interesting question is what it built to replace that revenue.

The answer, mostly, was volume. Braintree scaled hard into large enterprise processing, winning the kind of accounts that move enormous sums at thin, annually renegotiated margins.

That was a real commercial achievement. It also quietly changed what PayPal is.

Where the growth actually is

This is the part that explains the valuation, and it comes straight out of PayPal’s own second-quarter deck, published on July 28th, 2026.

Branded checkout online, the high-margin business the company was built on, grew 2% currency-neutral. It is now 28% of total volume.

Unbranded payment service provider volume, mostly from Braintree, grew 13%. It is 45% of total volume.

Venmo grew 14%. Branded experiences, including in-store, grew 6%. International volume grew 0% currency-neutral, and international revenue actually fell 3%.

Now put the two headline numbers side by side. Total payment volume grew 10%. Transaction margin dollars grew 1%.

That’s the argument, in two figures. PayPal moved 10% more money and kept 1% more of it.

The take rate shows the same thing from another angle. Total take rate went from 1.87% in the second quarter of 2025 to 1.78% a year later. Transaction take rate went from 1.68% to 1.61%. Transaction margin fell from 46.4% to 44.9%, and from 47.7% in early 2025.

None of that is a demand problem. Demand is fine. It’s a mix problem, and not the good kind of mix. A bad mix is what’s left behind by years of decisions about where to compete and what to defend.

PayPal is growing fastest precisely where it earns the least, and the business it earns most from has stopped growing in any meaningful way. Active accounts have sat at 439 million for five consecutive quarters, roughly four million above where they stood at the end of 2022. That’s four years of essentially flat user growth, over a stretch when global e-commerce was adding buyers by the hundred million.

Who took the button

Bernstein’s estimate is the number to hold on to. PayPal’s share of US digital wallets has fallen from roughly 90% in 2017 to about 50% in 2023, and is now around 40%. Apple Pay sits near 20%. Shop Pay is in the high teens and compounding at about 30% per year, with an estimated $110 billion in volume last year.

Losing half your share of a market you effectively invented takes real competitive pressure. Look at where it came from.

Apple had the device, and the button that lives behind a fingerprint doesn’t need to be trusted; it just needs to be there. Shopify had the merchant relationship, so Shop Pay arrived pre-installed across millions of storefronts. Stripe had the developer and built Link on top of that access. PayPal had the habit.

Habit is the most perishable of those four. It survives only as long as it stays the easiest option, and it stopped being the easiest option somewhere around the time Face ID became standard.

The structure of e-commerce moved too. Amazon, Walmart and Shopify now account for around 55% of US retail e-commerce volume, up from 49% in 2023, and they are growing faster than the market. Not one of them has a reason to hand a third party the checkout experience. The open web, where a trusted third-party button was worth a premium, is the part of e-commerce that stopped growing.

PayPal’s answer was Fastlane, launched in January 2024. It’s a good product. One-click guest checkout, strong conversion data, exactly the right idea.

It also arrived years after Shop Pay and Link had established the behavior, and PayPal chose to distribute it through Fiserv and Adyen, announcing both partnerships in August 2024.

There’s a real argument for that. Reach beats purity, and Fastlane inside somebody else’s stack still touches merchants PayPal would never sign directly.

But look at what it means. PayPal now earns a licensing return on its checkout technology by shipping it through firms that compete for the same checkout, at a moment when its own branded checkout is growing 2%. That is monetizing the technology and conceding the position in the same transaction.

The easier problem

In August 2022, Elliott Investment Management confirmed a stake of more than $2 billion. PayPal responded the same day with cost cuts, a $15 billion buyback authorization, and a new chief financial officer.

Elliott had a case. PayPal had overhired into a pandemic that ended, then spent early 2022 abandoning its own 750 million-account target.

But look at what Elliott diagnosed. Cost, not position.

Cost is the easier of the two. It’s measurable, responds to instruction, and shows progress next quarter.

Losing the button was slower, harder, and invisible on any dashboard. PayPal executed the diagnosis it was given and never looked for the one underneath.

Read the preference straight off this year’s guidance: roughly $6 billion of share repurchases against about $1 billion of capital expenditure. Six to one, at a company whose leadership has publicly concluded it underinvested in its own technology platform.

That ratio is a strategy whether anyone calls it one or not. A buyback says the best available use of a dollar is to have fewer owners.

Fair enough when the shares are cheap. An admission when the alternative was fixing whatever made them cheap.

To be fair, the mechanics work, and technology spend did rise 10% in the quarter. Second-quarter profit fell 12% while per-share profit fell only 3%, because roughly a tenth of the company was retired over the year.

But PayPal has spent around $36 billion on its own stock since 2015, and the company now trades at around $50 billion. Eleven years of returning capital, and the market has priced almost all of it back out.

Three plans in three years

Dan Schulman’s PayPal was going to be a super app, then after Elliott it became a cost story. When Alex Chriss arrived in September 2023, his goal was to turn a payments company into a commerce platform, and he shipped Fastlane, PayPal Ads, the PayPal Everywhere debit push and a set of agentic commerce partnerships. He left on 2 February 2026.

Enrique Lores took over on 1 March 2026. By the end of April, he had split the company into three units: Checkout Solutions & PayPal, Consumer Financial Services & Venmo, and Payment Services and Crypto.

In May came the restructuring, targeting roughly 4,760 roles, or about 20% of the workforce, and aiming to deliver at least $1.5 billion in gross run-rate savings, phased out to 2029. The plan runs in three stages: strengthen the fundamentals through 2027, build momentum through 2028, then accelerate and disrupt thereafter.

Judge the plan on its merits, and it holds up. The diagnosis is honest, the reorganization maps to real customer groups, and about $400 million of savings should land by year-end.

The problem isn’t the plan. The problem is that it’s the third one.

The verdict arrived from outside

With so many changes in such a short time, without any tangible results, others start to take notice and form their own plans. Bloomberg reported Stripe had been looking at PayPal in February.

On 15 July, Stripe and Advent submitted an offer: $60.50 per share, more than $53 billion, roughly $50 billion in committed bank financing, an ownership split evenly, and no stated intention to break the company up. That was about a 28% premium.

On 20 July, the board met and rejected it as inadequate, without closing the door. During their Q2 results call, Lores said he would consider any path that offers shareholders “superior value.”

That sounds great, but what does that mean strategically?

Stripe processed $1.9 trillion last year, up 34%, and was valued at $159 billion in a February employee tender. Stripe does not need volume. It has more than PayPal does, and it’s growing at three times PayPal's rate.

What Stripe cannot build quickly is the consumer side. A brand that hundreds of millions of people already have accounts with, a peer-to-peer network with genuine social gravity, and a wallet consumers will actually load money into. That takes fifteen years or an acquisition.

Which answers the question this piece opened with. The assets are worth considerably more than the company that owns them, and the bid is someone saying so out loud with $50 billion of committed financing behind it.

In other words, PayPal’s board is making a specific bet. It’s betting that a plan running to 2028 and beyond, in the hands of a chief executive who started in March, will pay shareholders better than cash on the table today. That’s a defensible bet. It also has to be re-earned every ninety days.

Here’s what each outcome would prove. None of this is a prediction.

• If a higher offer arrives and the board takes it, the rejection was competent price discovery, and the assets were always worth more than the plan.

• If a higher offer arrives and the board refuses again, the board is claiming it can beat a control premium organically, and every future quarter becomes a referendum on that claim.

• If talks end with no deal, PayPal gets measured against its own numbers with no bid to flatter the share price, and the number that decides it is branded checkout growth, currently 2%.

Waiting costs Stripe and Advent almost nothing. The money is committed, the strategic case doesn’t expire, and every quarter that looks like the last one moves the argument their way.

PayPal has spent eleven years learning that a portfolio and a position are not the same thing. The market cap was only ever the receipt. The click was the business.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

Today it’s worth a fraction of that, down close to 70% in the past year alone. In roughly the same window, it burned through three CEOs and lost its president less than twelve months after she took the job.

The easy read is financial. A bad quarter, a guidance cut, a market that soured on an old-line processor.

That read is wrong. It isn’t even close.

Fiserv didn’t fall apart because the market repriced it. The market repriced it because it had already fallen apart. The stock was the last thing to break, not the first.

What broke first was the strategy. Or the fact that, under all the growth, there wasn’t much of one.

So this isn’t a piece about earnings. It’s about what happens when a company mistakes momentum for direction, and keeps sprinting on the first long after the second is gone.

Let me break it down.

What Fiserv was supposed to be

Start with the ambition, because the ambition made sense.

In 2019, Fiserv bought First Data for about $22 billion, one of the biggest deals payments had ever seen. The logic was clean. Bolt the largest US merchant-acquiring business onto one of the largest bank-technology businesses, and own the whole pipe.

Fiserv already sat inside thousands of banks. Now it would sit inside millions of merchants too. Money moving on one side, the tools to move it on the other, and a fee clipped at every step.

The prize inside First Data was Clover: the point-of-sale and business-management system for small merchants. Picture the tablet and card reader on the counter at your local cafe, plus the software that runs the shop behind it. Clover became the growth story.

So Fiserv wasn’t sold to investors as a sleepy utility. It was sold as a compounder. Boring plumbing on the outside, double-digit growth on the inside, the kind of stock you buy and forget for ten years.

That’s a great pitch. Own the most rails, move the most volume, and somehow grow faster than everyone while you do it.

Here’s the catch, and it’s the whole piece in one line. Owning the most rails is a position. It is not a strategy.

A goal is not a strategy. “Grow double digits” is a goal. “Be the biggest processor” is a goal. A real strategy is the harder thing underneath. An honest read of your actual problem, a clear choice about how you’ll deal with it, and a set of actions that all point the same way.

Fiserv had the goals. It had the scale to chase them. What it slowly lost was the choice about what to actually be.

And once that choice goes missing, something still has to feed the growth number every quarter. That’s where the trouble starts.

The diagnosis Fiserv avoided

The first test of a real strategy is whether you’ve honestly named your problem. Most bad strategy fails right here. It skips the diagnosis, because the diagnosis is a hard pill to swallow.

Fiserv’s hard pill was simple. Strip out a few things that couldn’t last, and the core business grew at low-to-mid single digits, like the infrastructure utility it actually is.

Look at what was holding the headline number up.

Argentina. Fiserv ran a business there that paid merchants early, before their card money cleared, and charged for it. In a country with runaway inflation and sky-high rates, that spread is enormous. It’s a license to print, right up until the economy calms down.

How big was it? By the company’s own account, Argentina alone contributed 10 of the 16 points of organic growth in 2024. Most of the growth the whole company was celebrating came from one inflation-soaked market.

That’s not a franchise. That’s carry, the kind a trading desk books, not the kind a software platform earns. When Argentine inflation cooled, the growth cooled with it, and what was left underneath was ordinary.

Argentina wasn’t the only prop. A few others were quietly doing the work:

• Tech investment deferred for years to protect margins. That flatters today’s profit and starves tomorrow’s product. It’s borrowing from your own future and calling it discipline.

• Pricing pushed above what the market would bear, then pulled back when it wouldn’t hold.

• Merchants moved onto Clover from older Fiserv systems in ways that lifted volume, and that the company is now defending in a shareholder lawsuit. More on that shortly.

The moment of truth came in late October 2025. New leadership opened the books, cut the outlook, and reset organic growth guidance from around 10% toward 3.5–4%. The stock fell about 44% in a single day, its worst ever, erasing roughly $30 billion. Analysts who cover the sector called it, more or less, unrecognizable.

But the crash was the symptom. Here’s the diagnosis Fiserv hadn’t said out loud for years. The growth wasn’t the output of a strategy. It was a substitute for one.

That is textbook bad strategy. Not the absence of a plan, but the refusal to face the real problem, papered over with a number that looked like success.

The carousel at the top

You can usually spot a strategy vacuum before it hits the financials. Look at who’s running the place.

In a little over a year, Fiserv’s top job changed hands three times. The long-serving CEO who built the era left to run a government agency. His successor lasted about thirteen months, then left to run a bank. A third stepped in. Around the same time, a new president, hired with a resume most boards would fight over, walked out less than a year after arriving.

Three CEOs in roughly thirteen months. It’s tempting to make this about the individuals. Don’t. The people aren’t the story. The pattern is.

Good people leave when there’s nothing solid to hold. A real strategy keeps them, because there’s an obvious thing worth staying to finish. Take that away, and every new leader has to start by working out what the company even is.

That re-diagnosis is expensive and invisible. A new CEO spends the first couple of quarters learning which of the old assumptions were real and which were wishful. By the time they set a direction, half their runway is gone. If they leave before it proves out, the next arrival starts the same clock at zero.

So nothing compounds. The company lives in a permanent year one.

The people in the middle learn the obvious lesson: don’t commit to this plan, wait for the next one. That’s how ambition drains out of a company that still, on paper, has everything.

Scale is supposed to protect you from this. Fiserv is enormous. It processes for a huge slice of the US economy, and its rails aren’t going anywhere. But scale only buys time. It doesn’t buy direction. Spend that time arguing with yourself about what you are, and the size stops being a moat and starts being weight.

Shopping the crown jewels

If you want the sharpest picture of a company that’s lost the plot, look at what it was reportedly willing to sell.

Reports this year said Fiserv had quietly shopped one of its debit networks, STAR, to some of the biggest US banks. A debit network isn’t a side asset. STAR routes debit, ATM, and online payments for more than 115 million cardholders across 2,800-plus financial institutions. It’s real infrastructure.

The company hasn’t confirmed the talks, and plenty of analysts think nothing happens. Forget whether it closes. The reason it even came up is the tell.

Ask why a mega-bank would want a debit network. Not because the technology is magic, and not because they couldn’t build one. The pull comes down to a single rule from 2010.

Under the Durbin Amendment, big banks face a cap on the debit interchange they can earn per swipe. If a bank owns the network the payment runs on, it can argue its way around that cap. Analysts at KBW put the prize at somewhere between $460 million and $3.5 billion in extra gross revenue for the largest issuers.

So read what that means. The rails aren’t wanted as rails. They’re wanted as a loophole.

When the only reason anyone wants your core infrastructure is a regulatory workaround, the market is telling you what those rails are worth in your hands. Not much.

And even that thin logic tends to fall apart. Route your debit onto a network you just bought, and the card giants can pull the volume discounts they give you across all your other cards. Retailers push back. Regulators circle. Some banks reportedly ran the numbers and walked.

Here’s the strategic point, and it isn’t about debit at all. A company that knows what it is looks at an asset like STAR and asks what it can build on it. A company adrift looks at the same asset and asks what it can get for it. One treats its infrastructure as a foundation. The other treats it as inventory.

For contrast, a rival in the same seat took the opposite view. Asked about her own debit network, the CEO of FIS said flatly that she has no interest in selling it, and that there are interesting things she wants to do with it. Same kind of asset. Opposite instinct. That instinct is the whole difference between having a strategy and not.

The engine that was really a story

Now to Clover, the engine that was supposed to make the whole thesis true.

For years, Clover’s growth was the proof that Fiserv wasn’t just a commodity processor but a real software platform. And to be fair, Clover is a real product. In 2025, it did about $3.3 billion in revenue, up 23%. Merchants use it and like it.

But look at how some of that growth got made. A chunk of it came from moving merchants onto Clover from Fiserv’s older systems. Those migrations lifted the numbers. They now sit at the center of a shareholder lawsuit that claims the company pushed as many as 200,000 merchants onto the platform to inflate its growth, then stayed quiet as many of them churned back off.

I’m not here to try the case. The strategic point holds either way.

As the one-time boosts faded, Clover’s underlying growth slowed. By management’s own comparison, its US payment volume in one recent quarter grew about 7.5%, landing, in their words, between Visa’s and Mastercard’s growth rates.

Think about that. A platform sold as a share-taker had settled into growing at roughly the speed of the card networks it runs on. That’s the quiet death of a growth story. Not a crash, just a convergence. The engine keeps running. It simply stops pulling ahead of the pack it was meant to leave behind.

The lesson isn’t really about Clover. If your growth engine’s growth was partly manufactured, you didn’t build a franchise. You bought time. And time without a direction runs out quietly, which is the dangerous way, because you don’t hear it go.

Can this still be turned around?

So is Fiserv finished? No. And that’s what makes it a case study rather than an obituary.

The assets are real. Millions of merchants. Thousands of banks. A genuine software platform in Clover. Rails that half the industry’s disruptors would love to own. Fiserv didn’t lose the pieces. It lost the plan for them.

Turning this around isn’t about another acquisition or another leader. It’s about doing the three things Fiserv skipped. Richard Rumelt, author of Good Strategy/Bad Strategy, calls them diagnosis, guiding policy, and coherent action. In plainer terms: say what’s actually wrong, decide what you’re going to be, then make every move serve that decision.

First, an honest diagnosis. Say it out loud. The core is an infrastructure utility that grows in the low-to-mid single digits, and the double-digit years were borrowed. That sounds like surrender. It’s the opposite. You can’t fix a problem you’re still dressing up. To their credit, current leadership has started this, resetting growth guidance toward 3.5–4% and unwinding the pricing tricks. Painful, and correct.

Second, a guiding policy, which means a choice. This is the one Fiserv has dodged for years. Is it a software company or a processing utility? Those are different companies, with different owners, different margins, and different reasons to exist.

• If Clover is the future, then Clover gets the money, the engineers, and the roadmap, and it wins on product depth, not by migrating captive merchants onto it. The test is whether it can take share again on its own merits.

• If the rails are the future, then own that. Be the most reliable, best-priced infrastructure in the market, keep networks like STAR and build services on top of them, and stop apologizing for being a utility. Utilities are excellent businesses when they’re run and priced like utilities.

What Fiserv can’t be is both-and-vague, which is exactly what it’s been.

Third, coherent action, where every move serves the choice.

• Reinvest the tech spend that was starved to flatter margins. The bill for deferral has come due.

• Decide the debit network on strategy, not desperation. Keep it and build, or sell it cleanly to fund the chosen direction. Just don’t shop it around because you’re unsure what it’s for.

• Stabilize the top. A strategy needs someone to hold it for more than a year. The single most valuable thing the board can do now is let one direction run long enough to compound.

• Reset expectations honestly, so the company is measured against a real base instead of a fantasy. A durable mid-single-digit grower that knows what it is beats a fake double-digit grower that doesn’t.

None of that is glamorous. All of it is a decision. And the decision, not the market, is what’s been missing.

What it actually costs

Fiserv’s fall reads like bad luck if you only watch the stock. Watch the decisions, and it reads like something more ordinary, and more avoidable.

A company with every advantage in payments- the scale, the rails, the merchants, a real platform- spent years chasing growth numbers instead of choosing what to be. The numbers came from places that couldn’t last. When they ran out, there was no strategy underneath to catch the company, and the people at the top left rather than build one.

The good news is that the thing Fiserv is missing is also the one thing it can still choose. It can’t buy it, and it can’t reorganize its way into it. It has to decide.

Scale was never the strategy. It was just the part that came easy.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

The same paper reports that the model lost on money-laundering detection. Down 47% against the system it replaced.

Both numbers come from the same team, in the same paper. That honesty is the story, and this edition is about what sits between those two figures, and why it matters to anyone running payments strategy rather than a data science team.

Revolut isn’t alone. Stripe shipped a payments foundation model in May 2025, Visa published one in November, and Mastercard is building one with NVIDIA and Databricks. Nubank runs one in production for credit in Brazil, and Plaid trained one across 12,000 institutions.

This is the same pattern we’ve tracked all arc. Usage-based billing was about owning the meter. Machine-to-machine was about owning identity and settlement. This is the same instinct applied to a different layer: own the intelligence that sits atop the transaction, because that’s where the next margin is.

Two questions run through everything below. Is the value real, or is it data scientists admiring their own work? And can anyone outside the giants build one?

Let’s break it down.

A decade of groundwork, not a six-month launch

While the announcements all happened in the last six months. The work behind them took ten years.

Stripe launched Radar in 2016 and has published its progress in fraud modeling almost every year since. Adyen was catching marketplace fraud with graph-based models and running dozens of model versions in parallel long before anyone used the phrase foundation model.

Every company on this list walked the same road, in the same order: store the data, build the analytics, train one model per task, then compress what those models learned into a single reusable form.

The foundation model is the last step on that road, not the first. This matters strategically, because it tells you who can follow. A competitor doesn’t reach this by reading Revolut’s paper and copying it. They reach it by having spent ten years pouring the floors underneath it: the data pipelines, the labeling, the model infrastructure. If your organization hasn’t started, the paper isn’t a shortcut. It’s a map of how far ahead the leaders already are.

What these models actually are

Here’s the idea in one line: instead of building a separate model for every job, you build one model that understands your customers, and every job draws from it.

Think about how a “modern” payments business works today. It runs a zoo of specialist models, one for fraud, one for credit, one for churn, each built and maintained by its own team, each fed its own hand-picked signals. Six specialists, six projects, and none of them shares what it learned with the others.

A foundation model reads the customer’s full history once and turns it into a single summary that every team can use. That summary is called an embedding: a compact string of numbers that captures how a customer behaves. Fraud, credit, and retries all read from the same summary instead of each building their own.

The technical unlock, for those who want it, is self-supervision, the same method behind ChatGPT. The model learns from raw history without anyone labeling it first, the way a child picks up language by hearing it rather than studying a textbook. That removed the labeling bottleneck that had kept these systems small and single-purpose. If that’s a layer of detail you don’t need, the takeaway is simpler: one model now does the learning that used to take six teams.

To be fair on the history, models that read transaction sequences aren’t new. Call it one-third genuine breakthrough, two-thirds a new name on work that was already underway. The one-third is the part worth your attention.

Where you sit decides what you can learn

Before looking at any company’s results, ask a more useful question than “how good is the model.” Ask what the model can even see. In payments, that depends entirely on where you sit in the flow, and no two seats see the same thing.

Issuers and neobanks see the person. Revolut and Nubank watch a customer’s whole financial life: the salary landing, the rent leaving, the FX trade, the phone plan, the insurance, the savings. Every product they add is another feed into the same picture of that customer. That’s why their strongest results are in credit. Nobody knows the borrower better than the bank that watches their money all month.

Acquirers and processors see the merchant side. Stripe and Adyen watch millions of businesses at once, so their models learn what an attack looks like across the whole economy rather than one shop. The card-testing pattern that hits a gaming platform on Monday hits a ticketing site on Wednesday, and a processor’s model has already seen it twice.

The networks see everyone. Visa and Mastercard sit between all the issuers and all the acquirers, the widest view in payments, and they’re turning that view into AI-powered services they sell back to both sides.

Keep this map in mind for the rest of the piece. What each model is good at follows directly from what its owner can see, and the recent wave of acquisitions, which we’ll come back to, is largely about buying a wider view.

Reading the results without a PhD

Now the numbers. And here’s a shift worth naming, because it’s rare in this industry.

Payments has always been a business of hidden moats. You did not publish how your fraud engine worked; that was the edge. What’s changed is that the AI world Revolut, Visa and the rest now compete in for talent runs on openness. The OpenAI and Anthropic model is to publish, because publishing is how you attract researchers and how others validate that your claims are real. So these companies are putting their methods, and even their failures, on the public record. Revolut’s paper was co-published with NVIDIA. Visa’s, Nubank’s and Featurespace’s are on arXiv too. For a data scientist and strategist like myself, that’s a gift: for once, I can check the work rather than trust the brochure.

The useful way to read a claim is to ask one question. Was this graded by the company selling it, or by an outside process anyone can inspect? Published research you can check. A press-release number you take on faith. Both can be true; they don’t carry the same weight.

But the number that should decide whether you care isn’t the accuracy score. It’s whether the improvement shows up as money, and whether a merchant would ever feel it. So for each builder below, the question isn’t “how clever is the model,” it’s “what did it move, and does it reach the merchant?”

The issuer side. Revolut’s PRAGMA is the strongest public evidence in the category. It trained on over 20 billion events from 25 million customers, and it published the results: +130% on credit scoring, +65% on fraud recall, and that honest −47% on money-laundering detection. The credit number is the one that matters, because for a lender, a better credit model is directly fewer defaults and more approvals. That’s not a benchmark; that’s the P&L.

Nubank reports a 1.2% accuracy gain over its already heavily tuned models. Tiny, until you translate it: Nubank frames it as worth two to three years of normal improvement, and at their scale a fraction of a percent is worth millions in avoided losses. This is the honest shape of these gains. They look small on a slide and land large on the balance sheet.

The acquirer side, and why merchants should care. Stripe is the clearest example of the long road. For years it ground down card testing with specialist models: catch an attack, label it, retrain, redeploy, all inside an incident. That work cut successful card-testing attacks by 80% over two years, on more than $1 trillion of volume.

Then the foundation model arrived, and Stripe describes it as an addition, not a replacement: a large model trained on billions of transactions that catches novel attacks the specialists miss and turns each payment into an embedding that the rest of the stack reuses. On its hardest slice, fraud hidden inside big merchants’ volume, detection jumped from 59% to 97%.

Here’s why that reaches the merchant, which is the whole point. That same embedding now feeds into Stripe’s retries, authentication choices, and acceptance optimization. Better fraud detection means fewer good customers wrongly blocked. Better acceptance means more sales clear. This is not a back-office metric; it’s conversion and revenue that a merchant sees in their own dashboard. One model, many jobs, all of them pointed at the merchant’s top line.

Adyen took a path no one else on this list took: its own data centers across five regions, with the GPU fleet sitting physically next to the data. The engineering discipline shows in the pace. A single model runs 10 to 25 versions at once, half of them changing every week or two, and moving tuning to GPUs made a two- to three-day job 40 times faster. Faster iteration is not a vanity metric here; it’s how quickly a better model reaches a merchant’s checkout.

From Adyen’s published Uplift work: +22% fraud recall, a 46% cut in authorization-rate loss, and a 13% gain on issuer refusals. Every one of those is a merchant-facing number: more approved payments and fewer false declines. And Adyen offered the most instructive admission in the category: it tried replacing its many models with one global model and, in its own words, tried, failed, and will keep trying. A company with hundreds of petabytes and its own hardware telling you the single-model future is hard even with every advantage. Read that as a caution against anyone claiming they’ve already cracked it.

The network side. Visa’s TransactionGPT is published research that beats Visa’s production model by 22.5% on a business metric the paper doesn’t name. Mastercard says its model outperforms advanced machine learning in early testing, though the metric isn’t public yet. Both are building this into the value-added-services layer, the part of a network’s business that grows faster than the rails, and the part they can sell to everyone else.

The result that ties it together. Featurespace trained on 5.1 billion transactions from 180 European banks and reports up to 140% improvement over hand-built signals, tested at banks the model had never seen. That last detail is the important one strategically: it means the learning travels. A model trained on other institutions’ data still worked at a new one, which is exactly what makes a rentable, buy-it-off-the-shelf version of this possible.

The reality check. One piece of independent research keeps everyone grounded. A major 2022 academic study found that conventional models, the tree-based workhorses most teams already run, still lead on medium-sized data. Even NVIDIA’s own demo admits that its embeddings alone lost to a plain conventional model and only won when combined with existing signals.

So the honest verdict, and the one to take into a strategy meeting: a foundation model is worth it when you’re moving enormous volume across many use cases at once. On a single task at ordinary scale, the model you already run probably matches it, costs a rounding error, and can explain its decisions. The foundation model is a freight locomotive. Most companies need a delivery van.

The bill, and the three things that actually gate you

Most people assume these models need a frontier-lab budget. The public numbers say otherwise, and this is where many strategy conversations go wrong.

The training run is cheap. Visa’s runs cost roughly $700 to $1,400 each in cloud time. Nubank’s came in near $1,500. Revolut’s billion-parameter model cost somewhere between $21,000 and $40,000. The single most expensive model in payments costs less to train than a junior analyst’s salary.

So compute is not the gate, contrary to what chip vendors imply. Three other things are, and they stack.

• Data. You need roughly 5 million labeled customer records before this even beats the model you already run, and tens of millions before it’s worth the trouble. That’s not something money buys quickly; it’s years of history. This is the real moat, and it’s why the acquisitions matter.

• Team. Ten to fifteen machine-learning and infrastructure engineers, permanently. The $2,000 training run is nothing next to the $3 million a year it takes to keep the thing alive.

• A decision you own. A point where the model’s output becomes money: the authorization, the credit line, the retry, the checkout. A brilliant score with no decision attached is a weather report for a field you don’t farm.

Networks, large PSPs and big neobanks hold all three. A mid-sized player usually holds none, and that single fact is the whole “can everyone do this” debate.

One apparent exception, since it comes up: Adyen’s data centers look like proof you need to pour concrete to compete. You don’t. Those buildings exist for regulatory control and uptime, and the GPUs moved into infrastructure Adyen already owned for other reasons.

Follow the money behind the story

Most of what’s been written about this category has been co-published by the companies that sell the tools to build it. NVIDIA co-announced Stripe’s model and Mastercard’s, supplies Revolut’s and Adyen’s stacks, and publishes a “build your own transaction model” guide. NVIDIA sells the compute, the clouds rent it, and Databricks preps the data. They profit whichever model owner wins, which makes them the most reliable beneficiaries in the room, and the most motivated storytellers.

So grade the narratives the way you grade the numbers. “You need a data center” sells hardware, and the training costs above show it’s overstated. “Our data moat is unassailable” flatters incumbents, and it’s half true: data is the gate, but 5 million records is a lower bar than the marketing implies.

There are two honest ways out for everyone else. Rent it: Mastercard, Plaid and Featurespace are all building versions you can buy, which doesn’t remove the moat so much as move it to the vendor. Federate it: Swift and Google Cloud have a dozen banks training a shared fraud model without any of them handing over data, still a sandbox, fraud only, but the one genuine opening for a player without a giant dataset of its own.

The regulator already read the paper

There’s a constraint the launch announcements skip, and it decides where this value is actually allowed to land.

The EU AI Act treats credit scoring as high-risk, and it specifically exempts fraud detection. The high-risk obligations begin from 2 August 2026, with some deadlines still being finalized. Europe’s top court got there earlier: its SCHUFA ruling held that a score which effectively decides a loan is itself a decision, with a right to explanation attached. The US demands the same through the reason codes on every decline.

The tension in plain terms: the law requires you to explain the “no,” and an embedding can’t. It may be the best-informed judgment in the building, but it can’t fill in the reason code a declined applicant is legally owed.

Now line that up against the results. The biggest published gains are in credit; credit is exactly where a black box is hardest to use, and fraud, the exempted task, is where every one of these models went live first. That’s not a coincidence. Regulation is quietly directing where the models get deployed, toward fraud and conversion, away from autonomous credit decisions. For anyone planning a build, that routing is the single most important design constraint, and the reason the emerging answer everywhere is a hybrid: the foundation model reads, an explainable model decides, a human signs off on credit.

Two different AI bets, landing at once

It’s worth separating two AI stories that are easy to blur, because they call for different decisions.

One is the foundation model, built on transaction data you own, sitting underneath your risk and credit decisions. The other is the frontier labs- OpenAI, Anthropic, Google- renting out reasoning that sits on top of workflows: investigating a fraud case, drafting a credit memo, handling a dispute. Anthropic powers FIS’s financial-crimes agent with the data staying at FIS. OpenAI and Stripe built a checkout protocol, not a risk model. The labs don’t train on your transactions, and the data owners won’t rent theirs out. Two markets, stacked, not one, and you’ll buy from both.

There’s one place they may collide, and readers of the machine-to-machine edition will see it coming. Agent-initiated purchases don’t behave like human ones, and Stripe already tells businesses to update their risk models to tell good bots from bad. If agent traffic breaks today’s fraud signals, it creates fresh demand for exactly this kind of model. Plausible, not yet proven.

What to actually do about it

Strip away the hype, and the guidance is specific by cohort.

• Networks and top-tier PSPs: build. You hold the data, the team, and the decision points. Aim at fraud and conversion, where the gains convert straight to volume and dodge credit’s regulatory weight.

• Neobanks: build. You’re the archetype this was made for. Every product you add deepens the same customer picture, and credit is where it pays off.

• Banks: rent and buy before you build. Rent reasoning on the FIS model, buy embeddings where you can, and build in-house only where you own a distinctive decision and have tens of millions of records.

• Mid-sized players: don’t build below the bar. Under roughly 5 million labeled records and a ten-person team, the model won’t beat what you run and the team will cost more than the lift. Federate for fraud, rent scores where you can, and keep your existing model, which at your scale matches the giant on any single task and can explain itself.

Where this goes

Look at the deal sheet through this lens and a pattern jumps out.

Global Payments bought Worldpay, which had just bought Ravelin, a machine-learning fraud firm. Adyen bought Orb and Talon.One, adding the meter and the loyalty data on top of the payment. Nuvei and Payoneer changed hands. And Stripe has reportedly put a $53 billion bid on the table for PayPal, which would give a merchant-side processor access to the consumer-side behavior of hundreds of millions of wallets.

Read each of those twice: once as the strategy on the press release, and once as what it feeds a foundation model. More history, more products per customer, more decisions where a score becomes money. Consolidation in payments is increasingly, among other things, a way to buy the one input money can’t create quickly: data with depth and time behind it.

Four signals will tell you where the category settles:

• The first network or PSP to report revenue attributable to the model, a hard authorization or fraud-loss number. That moves this from research to P&L.

• A mid-sized player shipping a competitive model on rented GPUs. That would break the “only giants” story.

• A regulator ruling on foundation models in credit. That decides the biggest prize.

• A rentable model reaching real scale. That means the capability is commoditizing, and the margin moves to whoever owns the decision.

My read: the value is real, it’s narrower than the headlines, and it concentrates where deep data, real teams, and owned decisions already sit.

Which closes the loop across this arc. The meter is worth more than the rail. Identity and settlement beat the transfer. And now the transaction record itself becomes the asset because it is what trains the intelligence.

Everyone can afford the training run. Almost no one can buy the data. That’s the whole game.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

After a decade with this report, I’ve settled into a particular way of reading it. The skill I care about most in payments strategy isn’t reading a number off a chart. Anyone can do that. The work, and the part I enjoy most, is reading what sits underneath the number. A share figure tells you that something is winning. The deeper question, the one worth your time, is how it’s winning, on whose rails, and what that means for everyone else in the chain.

GPR 2026 is full of numbers that reward exactly that kind of reading. Three of them made me stop and rethink something I thought I understood.

[Click here to Read the full Global Payments Report 2026]

Let's break it down.

1. The contactless story most of us are watching is the wrong one

If you follow the Western payments press, the big in-store story this year is Apple. For years, the iPhone’s tap-to-pay NFC chip was reserved for Apple Pay. In mid-2024, Apple agreed to open it to rivals in Europe, and the field moved in fast. The report tracks Klarna rolling out tap-to-pay across 14 European markets in December 2025, PayPal launching in-store in Germany, and Vipps in Norway claiming the first real alternative to Apple Pay on an iPhone (p. 16). In Europe, that genuinely is a landmark.

But the GPR does something useful here. It puts Europe back into a global picture, and in that picture the tap isn’t the headline at all. The scan is.

Look at how people actually pay by phone in a shop. The share reaching for a QR code rather than a tap:

• Argentina: 84%

• China: 80%

• Colombia: 76%

• Philippines: 74% (all p. 17)

These aren’t edge cases. The report shows five markets where payment apps already carry the majority of in-store transaction value, led by China at 89% (pp. 14-15), and it forecasts that payment apps will reach 46% of all in-store transaction value globally by 2030, growing 135% faster than the point of sale growth itself (pp. 12-13).

Source: Global Payments Report 2026, p. 17. [Click here to Download the full report]

Why the rail that won isn’t the more advanced one

Here’s the part I keep coming back to, because it travels far beyond QR codes. NFC, which involves active microchips and radio-frequency transmission and allows for encryption, dynamic data updates, and two-way communication, is widely considered a more sophisticated technology than a QR code, which is simply a static 2D barcode that can be read optically by a camera. It needs a secure chip in every phone and a contactless terminal on every counter, so it spreads only as fast as that hardware gets replaced on both sides. A QR code needs a printout and a camera. One scales at the speed of a hardware refresh cycle. The other scales at the speed of a sheet of paper.

So in markets that already had deep, well-functioning card terminals, there was little urgency to change, and the tap simply carried the existing card rails onto the phone. In markets that didn’t have that base, QR didn’t wait for the hardware. It routed around the gap. The rail that wins is rarely the most elegant one. It’s the one that fits the infrastructure and the cost structure already on the ground. That’s the lesson, and it holds across almost every payments shift I’ve watched.

Why this matters more than “contactless is growing”

This is where the shorthand gets risky, because it hides the only question that really matters: which contactless. A tap usually rides card rails, which means interchange, chargebacks, and the card networks earning on every transaction. A QR scan in Brazil, India, or much of Southeast Asia usually rides an account-to-account rail, money moving directly from bank to bank, at close to zero cost, with different settlement timing and a different fraud model, and the card networks sitting outside the flow.

So the QR story isn’t really about technology. It’s about who controls the in-store rail and who earns on it. Europe’s NFC opening is, at heart, card networks and their challengers competing over the tap. The QR boom across APAC and Latin America is driven by national systems and fintechs building rails that bypass the tap-and-card economics altogether.

The report makes that tangible. It tracks how these QR-based systems are now pushing past their own borders, with Brazil’s Pix and India’s UPI gaining acceptance abroad, and Project Nexus linking the instant-payment rails of India, Malaysia, the Philippines, Singapore and Thailand (pp. 18-21). The in-store rail question is quietly becoming a cross-border one. If your markets sit outside the card-led West, those guides will show you which rail you’re actually building on.

2. “Digital wallets are winning” is true, and it’s only half the story

This is the number that will travel furthest from GPR 2026: digital wallets are now the leading way people pay worldwide, at 56% of online transaction value and 33% of in-store transaction value in 2025 (p. 38). It’s true, and it matters. It’s also the doorway, not the room.

Because a digital wallet isn’t a payment method. It’s a container. Apple Pay, Google Pay, PayPal, Alipay, these are wrappers, and the question I always want answered is what sits inside the wrapper when someone pays.

• Is the tap funded by a credit card?

• By a direct bank transfer?

• By money preloaded into the app?

The same gesture at checkout can sit on completely different rails, with completely different economics. The single most useful addition to GPR 2026 is that its market guides now estimate the funding mix inside the wallet, market by market (p. 23). It was the first thing I turned to.

Source: Global Payments Report 2026, pp. 38-39. [Click here to Download the full report]

The same wallet is two different products

Put Japan and India side by side, and the point makes itself.

In Japan, a digital wallet is essentially a credit card living in a phone. The report’s funding-mix estimate shows cards, especially credit cards, leading what’s inside Japanese wallets (pp. 56-57), which fits the market with the highest direct credit-card share online in the entire report, at 58% of e-commerce value. In India, a wallet with the same generic name is an entirely different machine. Its funding mix is led by account-to-account transfers through UPI, ahead of stored cards (pp. 52-53), because UPI moves money straight from one bank account to another for free.

Two markets. The same rising “wallet share.” The same small icon at checkout. Underneath, two different rails, two different cost structures, and two different sets of companies getting paid.

It’s the difference between two people who both tell you they drink from a cup. One holds espresso, the other holds water. The container was never the story.

Source: Global Payments Report 2026, India pp. 52-53 and Japan pp. 56-57. [Click here to Download the full report]

It also corrects the lazy reading of “cards are losing”

There’s a second, quieter point here, and the report is refreshingly direct about it. The easy reading of “wallets up” is “cards down.” The GPR pushes back: total card volumes are still rising globally, both directly and through their use inside wallets (p. 38). Cards were still 48% of in-store and 31% of online value in 2025, over $16 trillion in direct spending, and where their direct share slips, much of that volume hasn’t left the card rails at all. It has simply moved inside a wallet. In a card-led market, “the wallet won” and “the card won” can be the same sentence. The wallet just became the new front door to the card.

Why funding mix is the number I’d track

This is why “we support digital wallets” tells you almost nothing on its own. It describes the front door and says nothing about the plumbing. The funding mix is what determines:

• Cost of acceptance: interchange on a card-funded wallet, versus close to zero on an A2A-funded one

• Fraud and dispute exposure: chargebacks behave very differently across the two

• Who your real counterparty is: the card networks and issuers, or a domestic bank rail

The report even breaks out the networks inside each market. Japan’s domestic JCB still holds 29% of card-scheme share; India’s RuPay, 13% (pp. 53, 57), detail the headline number erases. And it shows where this ends up: in a handful of markets, wallets have become superapps, Alipay and WeChat Pay in China, and increasingly Mercado Pago and the neobanks across Latin America (pp. 26-27). When the wallet becomes the place where the customer lives, the funding-mix question becomes the oldest question in this industry: who owns the relationship?

[Click here to read the full Global Payments Report 2026]

3. Cash isn’t dying. It’s settling.

We’ve all heard the line for fifteen years. Cash is on its way out, heading inevitably to zero. GPR 2026 quietly tells a more interesting story.

Globally, cash at the point of sale eases only from 14% of value in 2025 to 12% in 2030 (p. 38). That’s a two-point drift across five years, which isn’t the profile of something dying. And in several of the wealthiest economies on earth, it’s barely moving at all:

• Germany: 32% of in-store value (2025)

• Spain: 34% (both p. 79)

• Japan: 38% (p. 43)

These aren’t developing markets waiting for card infrastructure to arrive. They’re rich, fully banked societies that have simply chosen not to give cash up on cue.

The clearest signal is in Japan, and the report says it plainly

The Japan guide is unusually candid. It expects Japanese cash to stay “effectively flat in absolute value with only modest declines in share” (pp. 56-57), drifting from 38% to 31% by 2030, and it puts that down not to a lack of infrastructure but to trust, convenience, reliability, and long-ingrained habit. That’s a striking thing for a payments report to highlight: that the move from analog to digital isn’t always linear, and that in some markets the floor is structural, not temporary.

India is the control experiment

Hold Japan up against India and the whole pattern explains itself. India’s cash didn’t gently decline. It fell off a cliff, from 71% of in-store value in 2019 to 14% in 2025 (pp. 52-53). Same six years, opposite outcome. The difference wasn’t wealth, or smartphones, or willingness to go digital. It was one piece of infrastructure. India built UPI, a free, instant, government-backed bank-to-bank rail, and gave cash somewhere better to go. Japan never built an equivalent pull, and its consumers had no reason to leave a system they trust.

So the lesson isn’t “cash is resilient,” and it isn’t “cash is dying.” It’s sharper than either. Cash leaves precisely when a genuinely better rail arrives to replace it, and it stays, indefinitely, where one hasn’t. The decline isn’t a function of time. It’s a function of whether something has been built to displace it, and whether habit will yield.

Why this is a planning problem

A lot of strategy quietly assumes cash trends to zero on a predictable curve. Acceptance roadmaps, store-format decisions, the business case for retiring cash handling, many of them assume the cash line takes care of itself if you wait long enough. The report says that in Germany, Spain and Japan, it won’t, certainly not inside any horizon you’re planning around. The floor is real, and in some major markets it sits three to four times higher than a “cashless by 203X” assumption would budget for.

And notice how this loops back. The markets where cash actually collapsed, India and Brazil, are the same markets where a domestic account-to-account rail displaced it. Cash didn’t lose to “digital” in the abstract. It lost to a specific rail. Which is the same thing the wallet-funding data and the QR data were telling us. Every market in the report carries its own cash floor and its own line to 2030. The useful question was never “when does cash die.” It’s “what, if anything, is positioned to replace it here,” and you answer that market by market in the guides.

What the three add up to

Three numbers, three stories that run deeper than the headline. Read them together, and they stop being separate observations:

• Wallets are winning, but the funding inside the wallet decides the economics and the owner.

• Contactless is booming, but the rail under the tap decides who earns and who’s bypassed.

• Cash is declining, except where nothing has been built to replace it, in which case it holds.

Put plainly: payments in 2026 is no longer a story about digital replacing analog. That contest is largely settled. It’s a contest about which digital rail wins in each market, card rails, account-to-account rails, or stored-value and superapp rails, and about who controls and earns on that rail. The headline shares compress all of that into a single percentage. The real value of this report is that it lets you decompress it again: to see the funding mix inside the wallet, the rail under the contactless figure, and the force, or the absence of one, behind every cash line.

That, for me, is the whole point of payments strategy. Anyone can read a share number off a chart. The value is in reading the rail and the economics underneath it, and that reading now sits in the GPR’s 42 market guides, each carrying the wallet funding mix, the card-scheme breakdown, and the forecast to 2030.

I’ve only pulled on three threads here, and deliberately the most quoted ones. I haven’t touched the report’s read on:

• BNPL quietly reinventing itself: the supposed card-killer now fuelling card-backed installments while its pioneers turn into banks (pp. 28-31)

• Crypto finding real consumer relevance: not through the revolution it promised, but through ordinary card rails, with crypto-linked card spending up sharply in a single year (pp. 32-35)

• “Glocalization”: the slow knitting-together of domestic rails into cross-border networks that could reshape who even needs the global card schemes (pp. 18-21)

Each of those rewards reading at the same depth.

So here’s where I’d normally tell you to go read the report. Instead, let me tell you what to read it for: not the share numbers you’ll see quoted everywhere this month, but the composition underneath them, and specifically the market that matters most to you, which is almost certainly not one of the three I used here.

[To Download the Global Payments Report 2026, click here]

Thanks for reading, and feel free to share this newsletter with anyone you think could benefit from my breakdown, or the Global Payments Report itself.

Dwayne Gefferie

We swiped and signed. Then we dipped a chip, and the card’s secret moved into the silicon. Then we typed the number into a browser. Then we stopped typing it at all.

Apple Pay, Google Pay, a tap, done. Visa has provisioned more than 12.6 billion network tokens. Most people who use them never learn that the real card number stays with the bank.

Each step abstracted the human from the credential. A person still decided, and a person still paid.

Then came agentic commerce, the story of the past year. An AI agent acts on a person’s behalf, sometimes only suggesting a purchase, sometimes preparing it and waiting for a click, and increasingly completing the buy on its own inside limits the person set once. (If you need a refresher on which protocols exist, check out “Payments Protocols”).

Agentic commerce still points one direction, though. An agent buying from a merchant, for a human, mostly on the card rails.

There’s one step past that. The machine stops shopping for a person and starts buying for itself.

That’s machine-to-machine payments. The payer is software, the payee is software, and the thing changing hands is usually a service the machine needs to do its own job. Compute, a model call, data, an API.

This isn’t theoretical. In the first half of 2026, Coinbase, Stripe, Google, Visa, Mastercard and AWS all shipped agent-payment products. In May, AWS wired agent payments into Bedrock’s AgentCore in preview, built with Coinbase and Stripe, letting software discover a service, agree a price, pay, and move on inside a single loop, settling in a fraction of a second with no human approval.

Two of those launches are the clearest lens on where this goes. Coinbase built a rail for machine payments, called x402. And in June 2026, Mastercard built a product for them, called Agent Pay for Machines. Most of this piece breaks down those two.

[Image 1: The evolution of payments, from swipe to machine-to-machine]

What follows is what machine-to-machine payments are, why they work now when they never did before, and the question that matters most. Is this a frontier where new payment companies get built? Or is it early crypto again, interesting to specialists and no one else for a few more years?

Let’s dive in.

What machine-to-machine actually means

Two terms get used as if they’re one. They aren’t.

Agentic commerce is an agent transacting on behalf of a person. It runs across three modes. Discovery, where the agent finds and recommends. Human in the loop, where it prepares the purchase and waits for approval. Fully autonomous, where it completes the buy within the guardrails the person sets in advance. All three are agentic commerce, and all end with a merchant buying something for a human, mostly on cards.

Machine-to-machine moves the buyer. Here the software is the customer. It pays for what it needs to finish its own work, and it pays other software, not a store. Compute, an API call, a dataset, or a task handed to another agent.

Autonomy isn’t the dividing line, which is where the terms get muddled. A fully autonomous agent buying someone a pair of headphones is still agentic commerce. The same agent buying the compute it runs on is machine-to-machine. The question is who the purchase serves and what’s being bought.

There are four kinds of machine payment, and only one is really live.

[Image 2: The four types of machine payments]

Agent-to-service is an agent paying for an API, a model call, or a slice of compute. That’s the one transacting today. Agent-to-agent, one agent hiring and paying another, is real but mostly experiments. Agent-to-merchant is agentic commerce again, running on cards. Device-to-device, for example a car paying for a charger or a sensor selling its data, is still mostly a promise in 2026.

So “machine-to-machine payments are exploding” means one narrower thing. Agents are paying for software services, in small amounts, very often.

The honest number is worth stating plainly. On rails like x402, agents have run well over 100 million transactions and tens of millions of dollars in volume since 2025, at an average ticket near 20 cents.

That figure needs a caveat. A large share of it is test traffic and meme-coin activity, not commerce anyone would recognize. The plumbing is real and shipping fast. The genuine demand is still small.

Both are true at once. A transaction count quoted without that caveat is a headline, not an analysis.

So why would we care?

Small today doesn’t mean small forever. Tokenization started small too, now more than half of all transactions are tokenized either via a wallet or network tokens.

Why the card can’t follow the money down here

The mechanical reason a new rail exists at all comes down to a fixed cost.

A card transaction incurs roughly 30 cents in fixed costs before any percentage is added. That’s invisible on a 90-dollar pair of shoes. It’s fatal on a 20-cent API call.

A fixed 30-cent fee on a 20-cent purchase isn’t a fee. It’s a barrier. Most machine payments live below it.

When an agent calls a model a thousand times a minute, each call costing a fraction of a cent, card economics don’t survive the contact.

[Image 3: Why the card can’t reach the bottom of the stack]

Cost is only half of it. The rest of the card stack assumes a person is present.

No one taps a phone, approves a push, clears a 3DS check, or squints at a CAPTCHA. Those steps exist to prove a human authorized the payment. Remove the human, and they lose their meaning.

A machine can’t prove it isn’t a robot. It is one.

The theme is familiar from these breakdowns. Payments got cheaper; complexity did not. The card rails are built around one assumption: a person at checkout decides to buy, and machine payments break that assumption at its root.

That break is the opening. It’s why there’s room, for the first time in years, to build something new underneath.

x402, the rail built on a part of the web nobody used

x402 comes with a good backstory.

When the web was designed in the 1990s, its authors reserved a status code, number 402, labeled “Payment Required,” and marked it for future use. The web might one day need to charge for something directly. It sat unused for nearly thirty years, because the web found other ways to make money and no machine needed to pay another one mid-request.

Coinbase built on it in 2025. That’s x402.

The mechanic is simple. An agent asks a server for something. The server replies that it costs, say, a quarter of a cent, and asks for payment first. The agent pays in stablecoin in the same exchange, asks again with proof, and gets what it wanted. No account, no signup, no human, in about the time it takes to load a page.

It spread for three plain reasons. It’s nearly free to add, and a developer can wire it up in an afternoon. It’s cheap to run because a stablecoin transfer costs a fraction of a cent rather than 30 cents. And Coinbase gave it away.

In April 2026, x402 moved to the Linux Foundation, the neutral home behind much of the internet’s shared infrastructure. Visa, Mastercard, Google, AWS and Circle are all involved. That shift turned x402 from one company’s product into something closer to a shared standard, which is how a payment rail reaches critical mass.

One point carries the rest of this piece. x402 is a single thin layer. It’s the moment of payment and nothing else.

It doesn’t know who the agent is. It doesn’t know whether the agent was allowed to spend. It keeps no invoice, runs no dispute, files no report. It moves value and steps aside.

That’s by design, and it’s useful. It also means x402 is a component, not a payments company.

Agent Pay for Machines

Mastercard took a different route, and a revealing one.

Rather than treat a stablecoin rail as a threat, Mastercard built a machine-payments product that puts stablecoins to work inside it. Agent Pay for Machines, announced in June 2026, treats machines as a genuine class of payer. In Mastercard’s own framing, businesses will build services for AI agents to buy, and those agents will transact with each other continuously, at machine speed, down to microtransactions.

It runs on four steps. Authorize, Discover, Execute, Settle.

Mastercard’s own example is a coffee shop, and it’s worth walking through because it makes the model concrete.

The shop wants a website and gives an AI agent one instruction: build and launch our online presence. In Authorize, the owner sets a spending limit and picks a funding method, a card, a stablecoin wallet, or a credit line, and that permission is recorded on-chain, so there’s a tamper-proof record of what the agent may do.

In Discover, the agent finds trusted providers for the images, the copy, the domain and the hosting, using network-issued credentials so each side knows the other is legitimate. In Execute, it buys all of it, firing off a mix of small and larger payments at machine speed, using lightweight vouchers rather than a full card transaction each time.

In Settle, the providers batch what they’re owed and get paid in the currency they choose, fiat or stablecoin, with Mastercard standing behind the settlement.

Two details matter more than they first appear.

The batching step answers a hard problem. A million sub-cent payments can’t run through the card system one at a time; the economics above would kill it. So the agent transacts against cheap vouchers in the moment, and the network aggregates them into proper settlements afterward.

And Mastercard isn’t resisting stablecoins here. It’s building them in. The funding can be a stablecoin wallet, the payout can be a stablecoin, and the network adds what a raw transfer lacks: identity, trust between strangers, and a settlement guarantee.

A network committed real product to this while the volume was still small. That’s a considered read on where commerce is heading, not a reaction to where it is.

The stack a machine payment needs

Two anchors, then. A free crypto-native rail in x402, and a network wrapping it in a guarantee in Agent Pay for Machines. Set against the same picture, they show the whole shape of a machine payment.

[Image 4: What a machine payment is actually made of]

A machine payment isn’t one thing. It’s a short stack of layers, each a separate job.

• Identity and authorization: who is this agent, and was it allowed to spend this.

• Protocol and negotiation: how the two sides agree on what’s bought and for how much.

• Rails and settlement: the actual movement of money, where x402 lives and where Mastercard settles.

• Metering and billing: counting what was used, pricing it, keeping the record. Readers of the usage-based billing edition already know this layer. It’s the meter.

Plot the anchors on it. x402 sits at the rails layer, and only there. Agent Pay for Machines reaches wider, across authorize, discover and settle, because a network can.

Look at the top and bottom of the stack. Identity and the meter. Not the rail.

The lesson repeats from earlier editions. The rail is the cheapest layer in the stack. Moving the money is nearly free and getting freer, and the value sits above it, in knowing which agent to trust and in metering what it actually used.

For machines that’s truer still. Here the rail isn’t just cheap. It’s free, and run by a foundation.

So is this a real opportunity?

Now the real question. Not what x402 is, but whether a serious payment company gets built here, one that never looks at a human transaction. And if so, what does it build?

The answer is yes, and the money was never in the rail.

There’s no opportunity inside x402 itself. Coinbase gave it away, the protocol fee is zero, and settlement is heading to the floor, the way it did once stablecoins made moving a dollar almost free. The transfer is free, so there isn’t much of a business in charging for it.

The opportunity is the work around the rail. That’s where a machine-native payment company gets built, and the pieces are clear once the stack is in view.

It has to handle identity: proving an agent can be trusted before anyone transacts with it. Call it Know Your Customer for a customer that can’t hold a passport. It needs guardrails: hard limits, allow-lists, and a way to cut off an agent’s spending the moment it misbehaves.

It needs metering: counting and pricing what the agent used, the same monetization layer software companies already pay well for. It needs settlement and reconciliation: turning a flood of tiny payments into books a finance team and an auditor can close. And it needs compliance for a payer that fits none of the existing boxes.

These companies already exist in early form. Skyfire is building agent identity and a wallet, anchored on a “Know Your Agent” check. Crossmint is building agent wallets that work across rails. Nevermined is building metering and settlement made for agents.

None of them is trying to own the transfer. All of them are building the workaround for it. And none of them cares about a human checkout.

That’s the shape of the answer. x402 is a foundation to build on, not the building. The serious companies treat it as one component, the way they treat the internet’s transport layer, and put identity, metering, guardrails, and reconciliation on top.

Some lean on the network path instead, letting Mastercard carry settlement and the guarantee while they own identity or the meter. Most do both because, early on, you support every rail and let the market decide.

Two settlement models are worth separating. The pure stablecoin path, x402 style, is instant and final, with no middleman and no guarantee. Fast and cheap, but if an agent pays the wrong party, there’s no one to call. It sits outside chargebacks and outside Regulation E, with no dispute path and no one clearly on the hook.

The network path, Mastercard-style, is a beat slower but comes with identity, a guarantee, and a reconcilable settlement. One is built for speed and cost, the other for trust. The interesting companies bridge the two.

Reaching the machines is its own question, and you don’t sell to a machine. You integrate where they already operate: the AI platforms through the tool connections agents already use; the protocols through x402 and its peers; and the network programs through partner rosters like Mastercard’s. Plug in where the agents are, and their spending runs through you.

Should PSPs and acquirers care yet?

The harder question is the one that comes up most in strategy rooms. For a PSP or an acquirer, is this move now, or watch-and-wait?

The honest answer has two halves.

As a revenue line, wait. The genuine volume is small, most of it isn’t real commerce yet, and no one makes their year on 20-cent agent payments in 2026. Reorganize around this today, and you’re early in the way that costs money.

As a control point, don’t wait.

The pattern is recent enough to remember. Stablecoins looked like a niche for years while specialists built the infrastructure, and then the incumbents bought their way in. Stripe bought Bridge. Mastercard bought BVNK.

The companies that owned the infrastructure when the market moved got paid. The ones who waited paid up.

Machine payments are following a similar pattern, with one difference worth noting. The timeline is shorter. In the stablecoin cycle, the incumbents had years to react. Here, Mastercard shipped a machine product while the volume was still a rounding error.

So the choice, stated plainly. Machine payments won’t pay for themselves soon. But identity and settlement are control points, and control points are cheap to hold early and expensive to reclaim late.

If agents become a real class of payer, and the trend points that way, no one wants to buy their position back later from whoever took it while they waited.

The best time to take that position was probably last year. The second best is now.

The takeaway

For thirty years, every advance in payments moved the human another step from the machinery. The line ends here. The payer is now software, buying from software.

The rail for that world will be free. Coinbase proved it by giving away x402, and Mastercard made the same point by wrapping free rails around the one thing still worth charging for: trust.

The companies that matter won’t be the ones moving the money. They’ll be the ones that worked out what to charge for instead: the identity of the agent, the limits on its spending, and the meter that prices its work.

Machine-to-machine payments are small today, and easy to wave off. So was every other step on that line, right up until it wasn’t.

The human has left the loop. The only question left is who owns it now.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally, through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

Then it bought two in seven weeks.

In April, Adyen paid €750 million in cash for Talon.One, a Berlin loyalty platform, made its first acquisition in its history. Eight weeks later, on 11 June, it signed a second: $335 million for Orb, a usage-based billing startup out of San Francisco. The deal is set to close on 1 July.

Look at when this happened. The timing tells us a lot.

In February, Adyen reported its second-half 2025 numbers, and the stock dropped about 20% in a day. Net revenue grew 17%. The problem wasn’t the result.

It was the outlook. Adyen guided to 20%-22% growth for 2026, a shade under what analysts wanted, from a company the market had priced for more. Berenberg cut its target from €1,550 to €1,000 and still called the platform best-in-class.

That’s the bind Adyen is in. It can meet expectations and still watch the stock fall. It happened before, in August 2023, when a single soft quarter knocked 39% off the share price in a single day.

So picture a company whose growth is settling into the low twenties, trading well below its highs, under pressure to find margin somewhere new. In that window, it breaks its founding rule twice and rolls out an agentic-commerce suite at the same time. At a retail show this spring, it demoed a setup where autonomous agents shop on a merchant’s behalf.

For a firm that spent two decades insisting it would build whatever it needed, that’s a real change of posture. Every one of those moves points the same way: toward owning more of what happens around the payment, not just the payment itself.

This isn’t Adyen buying more processing. It’s Adyen buying a different layer.

And it’s not the only one.

One move, four buyers

In about six months, four of the best engineering shops in software and payments made the same move.

Stripe bought Metronome. Adyen bought Orb. Airwallex bought OpenPay, then Leapfin. And Salesforce, which isn’t a payments company at all, bought m3ter.

Different firms, different continents, one target: the software that meters what a customer uses, prices it, bills it, and books the revenue.

Usage-based billing.

Two editions ago, I put the whole M&A wave on a single grid. One column kept lighting up for the firms whose payment rails were already finished. They weren’t buying rails. They were buying the cart, the layer that decides on the payment before it happens.

This is that column, up close.

Here’s my read, and the rest is the argument for it. A processing fee gets cheaper as it scales, because any rival can undercut it by a few basis points. A billing system gets harder to leave as it scales because it is wired into how a company sets prices, books revenue, and closes the quarter.

The rail commoditizes. The meter compounds. And AI is accelerating the meter compound.

What the layer actually is

Quick definition, because “billing” is one of those words everyone nods at and nobody pins down.

The monetization layer sits on top of the payment stack. It does five jobs:

• Metering: captures raw usage, the tokens, API calls, and GPU-seconds a customer burns through.

• Rating: turns those events into amounts, under whatever pricing logic you’ve set.

• Billing: produces the invoices, credits, and caps.

• Revenue recognition: books it all under ASC 606 and IFRS 15.

• Reconciliation: closes the books.

Picture the stack (see diagram below). Networks and rails at the bottom, priced in basis points. Gateway, processor, and acquirer above them, also basis points, also compressing. Orchestration in the middle, thin and neutral.

Then a line. Below it is “the payment.” Above it is the money: metering, rating, billing, revenue recognition, and reconciliation.

Orb’s CEO, Alvaro Morales, makes the case plainly. Standalone billing systems, he says, are “fundamentally limited because they operate blind to transaction execution.”

The acquirers are betting the opposite. Fuse the two layers, and billing and payments are worth more together than apart.

One last point about that layer, and it’s the whole reason it matters. It’s where the price gets set. Everything below it just moves money. This is where you decide how much money there is to move.

Of those five jobs, metering is the hard one to build at scale. It’s the piece these deals were really chasing.

Why now, and not three years ago

For thirty years, software was sold by the seat. That model rested on one assumption: each extra user cost the vendor close to nothing.

AI killed the assumption.

Every model request costs real money. Tokens, GPU time, inference, all of it variable, all of it landing on the vendor’s own bill the moment it happens. When a customer fires 50 requests per minute, someone is paying for 50 requests per minute.

Run the per-seat math against that. If one AI agent does the work of ten people, you’ve sold one seat where you used to sell ten. Per-seat pricing undercharges for the value and leaves the vendor exposed to costs with no ceiling. It’s the worst of both: capped upside, uncapped cost.

Take a fifty-seat support team replaced by one AI agent that resolves tickets on its own. You’ve lost forty-nine seats of revenue and kept all of the costs. Intercom charges about a dollar per resolved ticket for exactly this reason: the value is in the tickets closed, not the seats filled.

The only way to line price up with cost is to meter what gets used.

You don’t need a survey to see where this is going. Watch the incumbents reprice their own products.

Salesforce charges for its Agentforce agents by the action, about ten cents each, through what it calls Flex Credits. GitHub moved every Copilot plan onto usage credits. Microsoft sells Copilot Studio in credit packs. Adobe meters its AI image generation the same way.

The firms with the most to lose from killing per-seat pricing are the ones killing it first. And the buyers are asking for it: enterprise procurement increasingly wants to pay for usage and outcomes, not seats nobody logs into.

Metronome’s surveys point in the same direction. Metronome’s 2025 study of usage-based pricing found that 77% of the largest software firms have already built consumption pricing into their products. Metronome sells billing software, so weigh it accordingly, but it matches what the incumbents are doing in plain sight.

The buyers say the timing out loud, too. Stripe calls metered pricing “the native business model for the AI era.” Adyen ties its move to the way AI is reshaping the way software is priced and consumed. Three buyers, one reason, no coordination.

Five deals, one target

Line the deals up, and the pattern is hard to miss.

• Stripe bought Metronome, the metering engine behind OpenAI and Anthropic, for a reported $1 billion.

• Adyen bought Orb, the billing platform for the AI-native crowd, for $335 million.

• Salesforce bought m3ter and wired it straight into its CRM.

• Airwallex bought OpenPay, then Leapfin, to cover billing all the way through to the closed books.

Five deals, four buyers, one year. That’s not a coincidence. It’s a market repricing a layer.

Notice who’s missing, too. Not one of these is a processor buying another processor. That game is being played elsewhere. This is scope, not scale.

Are they all buying the same thing?

From a distance, yes. Up close, no, and the differences are where it gets interesting.

Metronome handles the heaviest volumes. It ingests more than 100,000 events a second on streaming infrastructure, and it ran the metering behind OpenAI, Anthropic, Nvidia, and Databricks before Stripe bought it. If your product bills by the token at frontier-AI scale, this is the one built for you.

Orb sits a notch in from there. It stores every raw usage event, then works out the bill by querying that history when the invoice runs. Because the raw events sit there untouched, you can change next quarter’s pricing and re-rate last quarter’s usage against it. Its customer list reads like the AI-native roster: Vercel, Replit, Supabase, Pinecone, Perplexity, and Glean.

Orb rebranded in January as a “revenue design platform.” That’s a dressed-up way of saying pricing should change as fast as the product does.

m3ter is the enterprise plumbing. Near-real-time metering and rating that feeds a company’s CRM, ERP, and quote-to-cash systems. Its customers include Paddle, Onfido, and Sift, and now it lives inside Salesforce.

Set those against Stripe Billing, which is subscription-first and, according to independent accounts, tops out at around 100 requests per second on its events API. That ceiling is the gap. AI products need a thousand times the headroom, which is roughly why Stripe wrote a reported billion-dollar cheque instead of shipping a patch.

Airwallex went furthest down the chain. It bought OpenPay for billing, then Leapfin for revenue recognition and reconciliation, the unglamorous work of turning a year of transactions into numbers an auditor will sign. Stitch those together, and you get one stack that runs from accepting the payment to closing the books. For a company that already moves money across borders for 250,000-plus businesses, that’s a pitch aimed at the finance team, not just the developer.

Salesforce is the tell that this isn’t only a payments story. It bought m3ter to give its customers usage billing inside the CRM. It also needs the meter to charge for its own AI: Agentforce runs on per-action credits and has already crossed a billion dollars in recurring revenue. When the buyer has to meter its own product to get paid, the layer isn’t optional.

So the buyers chose for different reasons. Stripe bought raw throughput and the Frontier-AI book. Adyen bought the scale-up book and a data loop. Salesforce bought a meter that speaks fluent CRM and can point at itself. Airwallex bought the whole lifecycle.

Same shelf. Different prizes.

The best builders in payments just became buyers

These aren’t companies that struggle to build software. So when all of them choose to buy instead of build, that tells you something.

Stripe built its own billing product back in 2018. It extended it for usage models. It even bought a small billing startup for its team in 2024. Then it paid a reported billion dollars for Metronome anyway.

That sequence is the evidence. If real-time, high-volume metering were a simple feature, Stripe would have shipped it. It bought instead, which tells you how hard the problem is.

Hard in a specific way. Metering at AI scale means swallowing a firehose of events without dropping any, storing them so you can re-price later, and rating them under pricing that changes every few weeks. Get one event wrong, and you’ve misinvoiced a customer who pays you millions a month. The people who can build this reliably are not in short supply.

Look at where the founders came from. Metronome’s are ex-Dropbox. Orb’s spent five years running engineering at Asana. m3ter’s spent three years inside AWS, watching how a hyperscaler bills usage at scale.

That’s a small, expensive talent pool, and the buyers paid for the teams as much as the code.

One detail here is hard to ignore.

The people who built Adyen’s own billing engine, the one that ran at over €970 billion in volume, left in 2022 to start Solvimon. They raised about €9 million from Northzone, they’re processor-agnostic, and they’re scaling. Adyen still chose to buy Orb, a US company, rather than rebuild on the talent it used to employ.

That’s not a knock on Adyen. It’s a measure of how specific the prize was.

Adyen wasn’t shopping for billing know-how. It was shopping for Orb’s AI-native customer base and for the speed of owning it now instead of in three years. You can build a billing engine. You can’t build someone else’s customer list.

What owning the meter actually gets you

Four things, and they stack.

First, margin. Processing is a thin spread on a regulated cost. Billing software runs at software margins, roughly 80% gross, and the buyer can charge for it on top of the payment.

Stripe takes about 0.7% of billing volume on top of its processing fee, and its revenue-automation suite is on track for a billion-dollar run rate in 2026. The meter is a richer floor than the rail.

Run it on a merchant doing $100 million a year. As its processor, you might keep 20 to 40 basis points after interchange and scheme fees, so $200,000 to $400,000. Run its billing too, at 0.7% of the same volume, and that one line is $700,000, at software margins, on revenue far harder to walk away from.

Second, lock-in, and this is the real one. Swapping a payment processor is a weekend’s work by comparison, which is the entire reason orchestration layers exist. Ripping out a billing engine means re-wiring how you meter, invoice, book revenue, and pass an audit.

Once your processor also runs your billing, leaving stops being a project and becomes a year you don’t have. Ask anyone who has re-platformed billing. It’s measured in quarters, not weeks, and every quarter carries the risk of billing a customer incorrectly.

Third, the data loop. This is Adyen’s own framing, and it’s the sharpest of the four. Adyen described the Orb deal as closing the loop between what a merchant charges and how those charges perform.

In practice, that means seeing both the price a merchant sets and whether the resulting payment goes through, then using one to sharpen the other. A billing vendor that never touches the transaction can’t see the second half.

Done well, that could flag when a merchant’s heaviest users sit on its cheapest plan, or lift the success rate on recurring usage charges by retrying them more cleverly. Both are worth real money.

There’s a catch, though. No acquirer has yet to publish a hard number showing that the loop lifts authorization rates or reduces churn. The logic is strong. The proof is pending.

Fourth, speed. Whoever owns the meter lets a customer change pricing without booking an engineering sprint. In AI, where prices move every time a new model lands, that isn’t a luxury.

Mistral runs 32,000 invoices a month on per-token billing through Lago. Orb says Replit grew its revenue 40-fold after it moved onto usage billing. Those numbers break anything built for monthly subscriptions.

And one more the buyers rarely say out loud. Whoever meters OpenAI’s tokens gets a front-row view of how AI pricing behaves, across the whole market, before anyone else does.

One doubt worth holding

None of this is settled.

Stripe launched its own Billing product in 2018 partly to undercut the independents and fold billing into payments. The same move could turn billing into a near-free add-on and collapse the margin everyone is now paying for.

If that happens, the value climbs again, to whoever owns the next scarce layer: revenue intelligence, or the settlement of agent transactions. The meter is the prize today. It won’t stay the prize on its own.

What this means for you

Depends where you sit.

If you run a PSP or an acquirer, read this one twice. The per-transaction business is your floor now, not your ceiling. The question on the table is whether you can credibly offer the meter.

If you can’t, you become a commodity, rail-fed into someone else’s revenue system. The undifferentiated mid-market processor is exactly the spot everyone else is paying billions to leave.

If you’re a merchant, turn the question around. The headline reads “your processor wants to run your revenue,” and yes, that carries lock-in. The more useful question is whether your own pricing is keeping up.

Your customers increasingly expect usage and hybrid models. If your stack can’t meter and bill that, the processor offering to do it for you is solving a real problem, not inventing one.

It bites hardest in the mid-market, the merchants too big for self-serve and too small for white-glove. That’s the band where one accept-to-bill stack is most tempting, and where lock-in costs the most.

If you do hand it over, keep your options open. Hold your own usage data, keep more than one way to move money, and settle the exit terms before you’re in deep. The upside is speed. The risk is waking up unable to leave.

If you’re building in this space, the window for selling at the top just narrowed. The two best AI-native independents, Orb and Metronome, are gone. What’s left splits two ways.

There’s the neutral camp, betting that merchants will pay to keep billing independent of their processor. Lago, an open-source, multi-PSP platform, has raised over $22 million and runs billing for Mistral, PayPal, and Synthesia. Solvimon, the ex-Adyen team, took about €9 million from Northzone and stays processor-agnostic by design. Sequence raised around $19 million from Andreessen Horowitz on the same bet.

Then there’s the finance-first camp. Chargebee and Maxio are strong on subscriptions and revenue recognition for companies that care more about the close than the metering. Maxio came out of the SaaSOptics and Chargify merger; Chargebee keeps buying its way wider. Both are clean bolt-ons for an acquirer that wants billing without having to build it.

Any of them could be next. Checkout.com, PayPal, GoCardless, Fiserv, or Global Payments could all use a meter, and there are only so many good ones left.

Where this goes next

Watch four things over the next year or two.

A fifth big buyer is moving on a billing target. Billing revenue is growing faster than processing revenue. Any acquirer finally putting a hard number on that data loop. And the opposite signal, billing bundled toward free to win the payment, which would tell you the layer is commoditizing in turn.

By 2027, expect PSP pricing to shift. Less pure per-transaction basis points, more of a layered price: a platform fee to be there, a consumption fee on the meter above, and a thin processing spread underneath. The processor, still priced only in bps, is competing in the cheap seats.

There’s a thread back to the story everyone’s chasing, too. When AI agents start transacting on their own, paying per call and per outcome, the meter stops being a billing tool. It becomes the thing that settles the transaction.

The usage-based billing layer and the agentic-commerce layer are converging. The firms buying billing now are buying an early seat at that table.

One question for your next strategy session. Look at your stack and find the meter. Then ask who owns it.

If it’s the same company that moves your money, you’ve handed them the one box you can’t take back.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally, through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

Read the headline, and it looks like one more processor is bulking up. It isn’t. Nuvei’s own release says the quiet part out loud: the combined company will let businesses “accept, hold, and move money, including stablecoin transactions, across 190+ countries.”

That’s not a scale play. It’s a company buying the parts of the money lifecycle it didn’t already own.

I’ve been watching this build for a while, and Nuvei and Payoneer are the cleanest examples yet. Here’s the breakdown: what’s driving the wave, why it’s the mirror image of the last great consolidation, the four shapes it’s taking, and where it goes next. I’m also going to give the pattern a name, because it’s earned one.

The one-line version, and the rest of this is the argument for it:

The 2018-2021 mega-mergers bought scale. This wave buys scope.

We’ve been here before, and scale didn’t deliver

I’ve spent the past few years making one argument on stage and in this newsletter: the payment service provider is being disintermediated. The mega mergers of the last decade are where that story starts, so let’s set them down clearly.

That wave was enormous, and it was horizontal. Processors bolting themselves to other processors:

• Fiserv and First Data, January 2019, $22 billion in equity (closer to $46 billion with debt)

• FIS and Worldpay, March 2019, $35 billion (roughly $43 billion with debt)

• Global Payments and TSYS, May 2019, $21.5 billion

• Worldline and Ingenico, February 2020, €7.8 billion

• Nexi, SIA, and Nets, 2020, roughly €12 billion combined

The logic was no secret. FIS chairman Gary Norcross summed up the era in three words: “Scale matters.” Payments is a volume business; the cost per transaction falls as you grow, so combine and win on cost. Simple.

It didn’t pan out that way. Within a few years, FIS and Fiserv were both under pressure from their own investors over these deals, and the savings and cross-selling the decks promised proved harder to get than expected.

FIS separated Worldpay, which then passed to private equity and, as we’ll see, straight into this current wave. Global Payments and TSYS got re-sorted in the same set of moves. Two of the five mega mergers came apart and were rebuilt inside the next one.

None of these are weak businesses. That’s not the point. Buying volume in a commoditizing layer didn’t stop the commoditizing; it just built bigger companies stuck in the same squeezed middle.

Same pressure, a new response

Here’s the thesis I have developed, which I believe lies beneath all of it. Value in payments is fleeing the middle of the stack toward the two ends. Up toward the customer, the software, the integrated platform. Down toward the rails, the infrastructure, the new settlement layer.

The plain PSP and acquirer in the middle get squeezed. Not much to innovate on, not much to differentiate on, no good reason to raise prices. Commoditization sets in, and the company in the middle gets disintermediated.

This wave is the same pressure, now playing out through M&A instead of in-house build. And it needs a name.

I’ve spent years describing one half of this: the disintermediation of the payment service provider. This is the other half. Call it the Great Re-Bundling.

“The Great Re-Bundling of 2026” - by Dwayne Gefferie

For more than a decade, fintech pulled the payment stack apart into thin, specialized layers. A gateway here, an orchestrator there, processing, acquiring, and infrastructure, each sold as its own product. That pulling-apart is what commoditized the middle and hollowed out the classic PSP.

The Great Re-Bundling is the answer. The survivors are putting the layers back together, this time around, whoever owns the customer or the rail, so the margin stops leaking out at every handoff.

And it’s surgical. Nobody buys the whole stack; each buyer takes the one missing control point that closes an economic loop, whether that’s payouts, bank rails, stablecoin settlement, or a hard-won license.

It runs through acquisition rather than build for one reason: speed. If you can’t ship a capability in about 18 months, you buy it. Even Adyen, which has built in-house for twenty years, has started buying, and that tells you how the calculus has shifted.

Then, two firms in the same squeezed layer merged to survive on cost. Now, firms reach into the next layer over to grab margin that used to go to a partner, and onto new rails before those rails get priced like a commodity.

Scale was about cost. Scope is about margin, lock-in, and data.

The money is not small either. The three biggest deals of this cycle alone, Capital One and Discover, Global Payments and Worldpay, and FIS and Issuer Solutions, clear $70 billion between them. The full set pushes toward $100 billion.

Here’s the side-by-side.

The clearest tell sits inside the Global Payments and Worldpay deal. Global Payments acquired Worldpay for $24.25 billion. In the same transaction, FIS took Global Payments’ Issuer Solutions business, the old TSYS, for $13.5 billion.

Two of the original mega mergers fell apart and were re-sorted into a single three-way deal. The scale-era bundles are being broken up and rebuilt along cleaner lines. That doesn’t happen if the first wave worked.

The four shapes of the Re-Bundling

It’s tempting to file all of this under one word: consolidation. That word hides more than it shows. Look closely, and the wave splits into four shapes, and the strategy only makes sense once you pull them apart.

1. Horizontal bundling (accept + hold + move)

Buy the money-movement layer sitting next to acceptance.

• Nuvei and Payoneer is the textbook case. Nuvei could already accept payments; Payoneer brought cross-border payouts, multi-currency accounts, a banking network and settlement in 150 markets. Accept, hold and move, in one company.

• Mollie and GoCardless, roughly €1.1 billion, December 2025, is the European version. A card-first PSP buys a bank-to-bank network across 38 countries. The timing isn’t an accident: PSD3, the Open Banking Regulation and instant rails like EPI’s Wero are turning account-to-account from a sideshow into core plumbing, and Mollie bought its way onto that rail instead of building it.

2. Vertical integration (software and payments fold into each other)

Think Shopify, Square or Toast: software that pulled payments inside the product. This wave does the same thing through acquisition, and it runs both ways.

• Xero and Melio, $2.5 billion, June 2025. This one isn’t accept-hold-move. It’s an accounting platform that already owns the small-business relationship, buying the bill-pay layer so it can earn on the payment itself.

The general ledger stops being a record and starts being an operator. Whoever owns the workflow takes the payment margin. That’s the disintermediation of the PSP, in cash.

• Shift4 and Global Blue, $2.5 billion, is the same move from the other direction. Here, a payments company reaches up into commerce services: tax-free shopping, currency conversion, and a way into Alipay+ and WeChat Pay through Ant and Tencent. Software pulling in payments, or payments pulling in software, and either way, one owner ends up holding the whole column.

3. The new-rail land grab (stablecoins)

Buy the next settlement rail before it commoditizes the current one. This is where most of the deal activity is clustering.

• Stripe didn’t just buy Bridge for $1.1 billion in 2024. It bought Privy for embedded wallets, the “hold” layer, picked up the Valora team, and launched the Tempo blockchain with Paradigm. Stripe now has the whole stablecoin stack: acceptance, issuance, wallets, and settlement.

• Mastercard and BVNK, up to $1.8 billion, March 2026, is the largest stablecoin infrastructure deal on record. Watch the framing: Mastercard treats BVNK as a multi-rail layer across fiat, stablecoins, and tokenized money. The sharpest incumbents aren’t betting on one rail winning; they’re buying the tollbooth between rails.

This sits on years of non-card buying, from the Nets account-to-account business to Recorded Future. Visa has done the same with Pismo and Featurespace.

• Ripple spent roughly $2.7 billion in 2025 building the institutional version of the same stack: Rail for stablecoin payments, Hidden Road for prime brokerage, GTreasury for corporate treasury, and Palisade for custody, all wired to its RLUSD stablecoin.

The volume backs it up. S&P Global counted at least 14 stablecoin deals in 2025 alone and concluded that incumbents would rather buy this than build it. The buy-versus-build question is settled.

4. Specialisation and unbundling (the counter-move)

The mirror image of bundling, and the shape that’s easiest to miss.

• Global Payments and FIS is the clearest case. Instead of staying sub-scale generalists across the middle, each giant picked one end and went deep. Global Payments becomes a pure merchant acquirer; FIS becomes a pure issuer and banking-tech provider.

They specialized by swapping halves, and Global Payments put a number on the focus: about $600 million in cost synergies and more than $200 million in revenue synergies, the most explicit bridge in the set.

• On the issuing side, the biggest deal of the whole cycle is a vertical play. Capital One and Discover, $35.3 billion, closed in May 2025. An issuer and bank buys a card network to run a closed loop and stop paying the Visa and Mastercard toll.

People are calling it a “triopoly of vertically integrated networks,” and the two are expected to reach roughly $2.7 billion in pre-tax synergies by 2027. Own accept-hold-move on the acquiring side, or own the issuer, the network and the deposits on the other. Same logic, opposite end.

One rule ties all four together: stop being a sub-scale generalist in the squeezed middle. Own the full column for a segment, or own one layer at world scale. The undifferentiated middle is the death zone, and that’s where the scale wave left its biggest players.

The whole wave on one grid

Pull every deal onto a single grid, and the pattern is hard to miss. Six control points run across the top: acceptance, embedded workflow, payouts, account-to-account rails, stablecoin settlement, and licenses. Almost every deal is a move to own one or two of them.

Read the columns, not the rows. Workflow lights up for the software and the rails-complete buyers; payouts, stablecoin settlement and licences cluster on the firms filling a rail. Nobody buys everything, just the one box that closes the loop.

The two questions every buyer is answering

Strip the deals down, and each buyer is answering two questions. What’s missing from my stack, and is it too important to rent?

The first question cleanly splits the field. Firms with gaps in their rails buy rails: Nuvei buys payouts, Mollie buys bank rails, and Mastercard buys stablecoin settlement. Firms whose rails are already complete buy workflow instead, the billing, pricing, and promotion logic that decides a payment before it happens.

Watch Adyen for the proof. It built everything in-house for twenty years, then broke that rule twice in one quarter: Talon.One for promotions, Orb for usage-based billing. Stripe did the same with Metronome.

The two best builders in the industry are buyers now. They’re buying the cart, not the rails, because their rails are done, and AI-era usage pricing turned billing into the next contested layer.

The second question is sharper than the old build-versus-buy debate: own versus access. Stripe bought Bridge, Mastercard bought BVNK, and Visa just partnered with Bridge for stablecoin-linked cards.

Three routes to the same rail. The only real question is which capabilities are too strategic to leave in someone else’s hands.

And often what you’re buying isn’t code, it’s time. When the scarce resource is a license or an approval that takes years to win, building loses to buying on the maths alone. Mastercard said it plainly about BVNK: the licenses across multiple markets were the point, and building the same thing would have taken too long.

What comes next

If the last cycle re-bundled the lifecycle we already understood, the next one is a fight over the parts that barely exist yet. Three are worth watching. The first is the big one.

The agentic authorization layer

When an AI agent does the buying, something has to decide the purchase, authorize it, hold the mandate, and carry the trust between shopper, agent, and merchant. Whoever owns that layer owns the front door to agent commerce. That’s why everyone is piling in.

The card networks are bolting it onto tokenization. Visa shipped Intelligent Commerce, then Intelligent Commerce Connect, and tied up with OpenAI in June. Mastercard shipped Agent Pay and its Agentic Tokens.

Google has the Universal Commerce Protocol and AP2; Stripe has the Agentic Commerce Protocol with OpenAI and the Machine Payments Protocol with Tempo. Four-plus competing standards in barely a year, which tells you nobody has won yet.

Here’s the tension that makes it worth your time. The networks don’t want to be disintermediated by the AI platform sitting in front of them, and OpenAI already charges 4% on Instant Checkout, so the platform plainly intends to take a cut.

The companies solving agent identity, mandates, and orchestration are the ones to watch. They’re the quick movers, and they’re acquisition bait the moment a standard starts to win. This is the first new layer of the stack in years, and it will be bought, not built.

Treasury, reconciliation, and stablecoin settlement

Ripple buying GTreasury is the tell. Once a company owns accept, hold, and move, it finds the real lock-in living in the back office, in the treasury, liquidity, and reconciliation layer nobody enjoys, and everybody depends on. Boring and sticky are good things to own.

This is also where stablecoins go mainstream without the fanfare. Mastercard just opened card settlement to stablecoins across eight chains, Circle is pushing its payments network, and the banking-as-a-service crowd, BitGo among them, now sells stablecoin settlement as a product.

Watch the crypto firms climb up into corporate treasury, and watch the treasury and reconciliation vendors get rolled up by whoever wants that lock-in. Ripple moved first. It won’t be the last.

The issuer-side stablecoin stack

Capital One and Discover proved that issuers want closed loops. The next closed loop is on-chain, and almost no large issuer owns that rail yet.

The banks know it. JPMorgan launched its JPMD deposit token on Base, Wells Fargo filed for WFUSD, and SoFi launched SoFiUSD on BitGo’s stablecoin-as-a-service. JPMorgan, Citi, Bank of America, and Wells Fargo are also building a shared tokenized deposit network through The Clearing House, designed to prevent deposits from leaking to Circle and Tether.

Most of them will buy the rail rather than build it, just as the acquiring side did. Issuer processors and banking-as-a-service providers, Marqeta, Galileo, Lithic, Fiserv, and FIS, will need stablecoin issuance and settlement inside the loop, and the license is the scarce part, so stablecoin-as-a-service shops like BitGo and Paxos are the obvious targets.

The acquiring side moved first. The issuing side is next, and it sits on a bigger pool of money.

When the next deal lands, and it’ll land soon, don’t ask how big it is. Ask which control point just got claimed, and which one is still sitting in the open.

The middle is for rent. The ends are for keeps.

Thank you for reading.

P.S. If you’re reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally, through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

Six months ago, most of us signed off on a strategy. Our companies, our teams, our boards. We placed a set of bets about where payments were heading: what to build, what to defend, where the margin would come from, which threats were real and which were noise.

Good moment to see how those bets are holding up.

The first half of 2026 didn’t drift. It moved.

Four shifts did real work, the kind that changes who gets paid for what, not the kind that fills a slide and is forgotten by lunch.

I spent half a year in the rooms where this got argued out. Boardrooms, merchant calls, conference floors across Europe and beyond. What follows is my read, not a news roundup: what happened, why, and what I think it changes for you.

One pattern ties all four together.

Value is moving away from processing the payment and toward what sits around it: the relationship, the rails, the economics. Processing is becoming the cheap part, and the money is heading for the edges.

Read them that way, and they stop being four news stories. They’re one story told four times. Let’s get into them.

Shift 1: Agentic Commerce Met Reality

For over a year, “AI agents will do your shopping” was the loudest promise in payments. This half, it met the rails, and the collision taught us more than the hype ever did.

What happened. In March, OpenAI quietly killed ChatGPT’s Instant Checkout, the in-chat buying feature it had shipped only months before. According to the reporting at the time, barely a dozen of Shopify’s millions of merchants had ever switched it on. OpenAI never built the boring parts either: sales tax handling and inventory sync at scale.

The behavior underneath was the real verdict. People are happy to let AI help them shop, then freeze when it’s time to pay.

The surveys back that up:

• ~70% of consumers are open to AI agents for shopping and travel, but far fewer will hand over payments.

• 54% are fine letting AI plan a whole trip, yet only ~12% are fine with it booking on its own.

High interest in the assistant. Low trust in the wallet.

And under all that hesitation, the rails kept getting built. At Money20/20 Europe in early June, Worldline, ING, and Mastercard ran what they billed as Europe’s first end-to-end agentic payment in live production. A real ING cardholder, a real Dutch merchant, the transaction flagged to the issuer as agent-initiated.

Read that wording closely, because it’s carefully chosen. Santander and Mastercard already claimed Europe’s first agent-executed payment back in March, but that one ran in a controlled environment. So the June “first” is really “first in live production,” and we’ve had a steady drip of these, each claim a little narrower than the last.

That drip is the actual signal. The tech clearly works. What nobody will say out loud yet is that it works at scale, in the open, with money that matters.

PayPal and the UK search platform Hey Savi launched the UK’s first agentic shopping app over the same window, with Debenhams as the launch retailer. And a standards war went public:

• Google Universal Commerce Protocol (shown at NRF in January, with Shopify, Walmart, Target, and Visa behind it)

• Visa Intelligent Commerce Connect

• Mastercard “Know Your Agent” credentialing

• Coinbase x402 micropayment rail

Why is it happening? AI crossed from helping you find things to making the purchase, and the networks saw the threat at once. If an agent picks, checks out, and pays, the shopper relationship, plus the data and money attached to it, is suddenly up for grabs.

So everyone’s racing to set the standard before someone else sets it. Don’t read this as a technical debate.

It’s a land grab with “interoperability” on the label.

What it means. The lesson isn’t “agents will shop.” It’s about where control sits, and you can see exactly where the fight is by watching where the protocols disagree.

Google’s UCP keeps the retailer as the merchant of record, so the retailer keeps the checkout, the transaction, and the customer. Other models don’t.

That one design choice decides who owns the customer and who keeps the margin when an agent buys.

Here’s the part most strategy decks miss. The agentic activity that’s actually live at scale today isn’t consumer shopping at all. It’s machines paying machines. On rails like Coinbase’s x402, something like 69,000 agents have run tens of millions of dollars in transactions, at an average ticket of around 20 cents. AWS wired agent-to-agent stablecoin payments straight into Bedrock in May, in preview, on the same protocol.

Read that with one eye open, though. A big chunk of x402 volume is test traffic and meme-coin games, not real commerce. So my honest read is this: the plumbing for machine-to-machine payments is real and shipping fast, the genuine demand is still small, and almost none of it is covered. An agent paying over a stablecoin rail sits outside chargebacks and Regulation E. No dispute path, nobody clearly on the hook.

The rails showed up this half. The rulebook didn’t.

If you parked agentic commerce under “2027 problem,” here’s your correction: the infrastructure is a 2026 reality, the trust isn’t, and the two are moving at completely different speeds.

Shift 2: Stablecoins Became Plumbing

This is the one that jumped from panel chatter to the balance sheet, faster than most of us budgeted for in January.

What happened. Fiat-backed stablecoin supply surpassed $300 billion in early 2026, many times what it was at the start of the decade. On-chain volume in 2025 ran into the tens of trillions. Plenty of that is crypto-native, but a growing slice is real payments.

The regulatory unlock mattered most. The GENIUS Act became law, so US issuers finally had rules to build against.

Then the moves piled up:

• Visa settlement now runs across nine blockchains at roughly $7 billion a year.

• Visa and Stripe, through Stripe’s Bridge deal, started pushing stablecoin-linked cards into 100+ countries.

• Mastercard picked up a New York BitLicense.

• Tether launched its US-regulated USAT.

Watch the structural move underneath the headlines. Circle, Tether, and Stripe, with their Tempo chain, are each building their own settlement infrastructure to keep economics that used to leak out to public networks.

Why is it happening? Two things lined up. Regulatory clarity removed the excuse to stay away, and the cost and friction of cross-border settlement gave everyone a reason to want in.

Now look at what the incumbents chose to do.

They didn’t fight stablecoins; they swallowed them. Visa and Mastercard are pushing harder on services and settlement precisely because their own rails are under fee pressure from all sides, and absorbing the new rail is cheaper than defending the old one.

What it means. The point is where the pressure lands. Stablecoins aren’t coming for the checkout first.

They’re coming for settlement and treasury, the unglamorous, very profitable back office.

That makes this a B2B, cross-border, and treasury story before it’s anything else, which is easy to underrate if your strategy is built around front-end acceptance. If you sit in the mid-stack, this is the shift I’d model first, not the flashy agent one.

So here’s the uncomfortable question I’d put to you. If settlement gets near-instant, programmable, and cheap, how much of your margin is paid for moving money, and how much is paid for the delay and complexity that stablecoins are built to remove?

Shift 3: Consolidation Split Into Two Playbooks

Want to know what the industry believes about its own future? Watch what changes hands. A lot did in H1 2026, and the deals split cleanly into two playbooks.

What happened. Global Payments was the headline. In January, it closed its $24.3 billion buy of Worldpay and, in the same deal, sold its Issuer Solutions business to FIS for $13.5 billion. FIS confirmed the close on 15 January.

What’s left is a focused, pure-play acquirer running about $3.7 trillion a year. It simplified itself by dropping issuing and concentrating on merchant acquiring.

Then the one nobody saw coming. Adyen built its whole identity on building in-house and buying almost nothing. In April, it signed its first acquisition ever: €750 million in cash for Talon.One, a Berlin loyalty and incentives platform with more than 300 merchants.

When a company breaks a fifteen-year principle, pay attention to what it breaks it for. Adyen broke it to own the loyalty and retention layer that sits on top of the transaction. That’s where it now thinks the next margin lives.

Worldline went the other way. Rather than buy into new ground, it sharpened its focus on Europe. It agreed to sell its India business to BillDesk for about €60 million, moved its PaymentIQ orchestration platform to Incore Invest, and raised €500 million to fund the refocus on its European core.

Same pressure, opposite response. One company buys to expand what it owns, the other concentrates on what it does best.

And the market hasn’t stopped moving. In June, Nuvei, taken private by Advent back in 2024, was reported to be in advanced talks to buy cross-border SMB specialist Payoneer for around $2.7 billion. Treat that as direction, not a done deal. It’s reporting, nothing’s signed, and the stock jumped about 25% on the news, but it points the same way as everything else: cross-border reach is the prize.

Why is it happening? One pressure sits under all of it. Margin in plain processing is shrinking, and scale alone no longer buys growth.

That single fact explains every move:

• Buy focus and scale: Global Payments

• Buy the retention layer: Adyen

• Concentrate on the core: Worldline

• Chase cross-border reach: Nuvei, if Payoneer is real

And Visa and Mastercard keep leaning on value-added services, because moving a transaction isn’t the earner it once was.

What it means. For anyone in acquiring, these deals are a map with one spot clearly marked “do not stand here.” The undifferentiated mid-market processor whose pitch is “we’ll process your transactions reliably” is the squeezed middle everyone else is paying to escape.

So the question I’d put to you is blunt. In a market moving this fast, what do you own that a $3.7-trillion rival can’t copy and a loyalty-rich platform can’t absorb?

If the honest answer is “nothing structural,” these deals just told you where you sit.

Shift 4: Banks Built Roads Around the Cards

The least-covered of the four, and probably the one with the longest tail. This half, banks stopped piloting card alternatives and started rolling them out.

What happened. In the UK, the UK Payments Initiative went live. It’s an FCA-regulated account-to-account scheme backed by the biggest lenders: Barclays, HSBC, Lloyds and NatWest, plus Santander, Monzo, Revolut, Starling, with rails from TrueLayer and Token.io.

It goes after recurring and variable payments first, starting with utilities, government, and charities, and it’s pitched straight at cards and direct debit. There are commercial rules attached, not just open-banking goodwill.

In the EU, Wero, the European Payments Initiative’s wallet on SEPA Instant rails, pushed past peer-to-peer toward pan-European e-commerce. Worldline and Airwallex are enabling merchant access, and banks in Germany, France, Belgium, and the Netherlands are onboarding. Even Mastercard hedged, teaming up with PaidBy for cross-border A2A.

Why is it happening? Three forces stack up:

• Regulation is turning instant payments from an option into the default across Europe.

• Sovereignty is now a stated political goal, the wish to stop routing Europe’s everyday payments through two US networks said out loud rather than muttered.

• Economics: banks see a chance to win back the margin they handed the schemes years ago.

When the regulator, the politics, and the P&L all point one way, things that sat still for a decade suddenly move.

What it means. Take these seriously because of where they start. They’re not chasing the high-value, dispute-heavy, rewards-driven purchase.

They’re going after recurring, predictable, low-risk volume: subscriptions, bills, top-ups. That’s the quiet, reliable base acquirers and issuers lean on.

A2A won’t replace cards at the checkout this year, and it doesn’t need to. It only needs to peel off the steady volume at the bottom of the stack, and that’s now exactly what it’s built to do.

If your plan treats recurring card volume as locked in, this is the shift quietly working against it while you watch the louder three.

The Reflection: Does Your January Strategy Still Hold?

Line the four up, and the thread is hard to miss. Agentic commerce is a fight over who owns the relationship when a machine makes a purchase. Stablecoins are a fight over settlement economics. Consolidation is firms repositioning past commodity processing, some by buying, some by focusing. The bank-led A2A schemes are a fight over the rails.

Four fronts, one direction: value is leaving the middle of the payment and heading for the edges.

The point of seeing this clearly isn’t to call the back half of the year. Predictions are cheap and usually wrong. The point is sharper: you wrote a strategy in January, and a strategy is only as good as the bets under it.

So go back through your plan, and be honest:

• Did it allow for agentic commerce being won or lost on the merchant-of-record question, not the AI itself?

• Did it assume your settlement and treasury margin was safe, just as the cheapest, fastest rails in a generation get built to squeeze it?

• Did it leave you in the squeezed middle, while the strong spend billions to escape it, and others struggle hard to get out of it?

• Did it treat your recurring volume as locked in, just as a bank-backed, regulator-blessed alternative shows up aimed straight at it?

If you saw all four coming, you wrote a sharper January plan than most of us. If you didn’t, the midpoint isn’t for feeling behind. It’s for fixing things while you’ve still got half a year.

The plain truth is this. The strategy you’re running was written for a payments industry that’s already changing under it. Not collapsing, changing. The value is moving.

The only question that counts at the half-year mark is whether you’ve moved with it.

Thank you for reading.

P.S. If you're reading this and are looking for the insights to help you improve your own strategy, that’s exactly what I help payments companies figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally, through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these breakdowns and deep dives, buy me a coffee.

Then one day, the account changes hands. Maybe through a romance scam. Maybe a fake job offer. Maybe outright threats. The person who opened the account is no longer the person controlling it.

And here’s the problem: the bank’s compliance system already trusts it.

Every checkpoint was cleared at onboarding. There is no second verification. There is no trigger. The account is clean on paper and dirty in practice. According to new research from Incognia, only 16% of financial institutions, from traditional banks and credit unions to neobanks, fintechs, and money service businesses, catch this before any suspicious transactions occur. The other 83% are cleaning up damage after the fact.

That number alone should keep fraud and risk leaders awake. It kept me up.

Over the past few weeks, I’ve been writing about the shifting fraud and risk landscape. What’s changing, why the current defences aren’t keeping pace, and where the real exposure sits. But Incognia’s The State of Mule Account Handovers in 2026 stopped me cold. It’s based on a survey of over 500 fraud, risk, and AML professionals across the US and Europe, and it doesn’t just confirm what practitioners have been feeling. It puts hard numbers behind it.

This edition isn’t a summary. It’s a breakdown. I’m pulling out the data that matters most, explaining why it matters, and giving you something you can take back to your team on this week.

Read the full report here →

1. The Structural Blind Spot

Mule account handovers exploit a gap that the financial services industry designed into its own compliance framework. To understand why, look at how mule accounts have evolved.

Twenty years ago, mule accounts were purpose-built. Someone opened an account with the intent to move dirty money. KYC was your first and best shot at catching it. A decade later, the model shifted to recruitment. People were lured in through fake job offers or money transfer schemes, used their accounts for a period, and sometimes got them back. The compliance system could still cope because there were signals at or near onboarding.

Now we’re in a third generation. The account was opened legitimately. The customer used it normally for weeks, months, sometimes years. Then, through coercion or manipulation, they handed control to someone else.

At that point, every compliance checkpoint has already been cleared. The account is verified. It’s clean. And the system trusts it indefinitely.

That’s not a detection failure. That’s a design flaw.

The report confirms this at scale. 81% of respondents across banks, neobanks, fintechs, credit unions, MSBs, and payment processors report an increase in mule activity involving account handover over the past twelve months. Not a marginal uptick. An industry-wide trend.

But the more revealing number is what happens after the handover. Only 16% catch it before suspicious transactions occur. The breakdown of the other 83%:

• 46% detect it after early suspicious behaviour, but before significant funds move

• 26% catch it after funds have already started moving

• 7% find it after the money has exited the institution entirely

• 4% only learn about it through customer complaints or external alerts

Each stage represents real losses, real compliance exposure, and real customers caught in the crossfire.

So what are the 16% doing differently? They’re not treating individual signals as standalone triggers. They’re watching for compounding risk over time. A device change alone is noise. A location shift alone is noise. But a device change followed by a location shift followed by a behavioural anomaly in quick succession? That’s a pattern. That’s what separates proactive from reactive.

Strategy: Shift from point-in-time identity verification to continuous account-level monitoring that extends well beyond onboarding.

Action:

1. Map your current detection triggers against the full account lifecycle, not just the transaction layer.

2. Ask your team one question: “At what point after onboarding do we re-verify who is behind this account?”

3. If the answer is “when something suspicious happens at the transaction level,” you’re in the 83%.

2. The Cross-Border Dimension Most Teams Are Missing

Mule account handovers aren’t just a domestic problem. And this is where banks, neobanks, and MSBs are likely underestimating their exposure.

64% of respondents report confirmed or suspected cross-border mule activity. 34% strongly suspect it based on behavioural, device, or transaction signals. 30% have confirmed cases where account control shifted to individuals outside their primary operating country.

The numbers are striking on their own. The correlation underneath them is what matters.

Financial institutions reporting significant increases in mule activity are nearly twice as likely to have confirmed cross-border account handover. If you’re seeing a domestic spike, there’s a real chance you have a cross-border problem you haven’t uncovered yet.

Larger institutions are especially exposed. Respondents at banks and financial institutions with over 50 million accounts are 40% more likely to confirm cross-border handover than the overall respondent base.

And the benchmarks vary sharply by institution type. MSBs report the highest rates of significant mule account handover increases at 44%, nearly half again as much as the industry average. Neobanks are the most likely to confirm account control shifting outside their primary country at 38%. If you’re in either category, industry-wide benchmarks understate your risk.

Strategy: Extend mule detection beyond domestic patterns and build cross-border visibility into your geographic risk model.

Action:

1. Layer geographic behaviour analysis into your mule detection workflow.

2. Ask: “Can our systems detect when account control shifts to a device or location in another country?”

3. If you don’t know the answer, you probably don’t have the visibility. That’s the gap.

3. Coercion Changes Your Entire Response Model

This is where the report gets uncomfortable for most fraud and risk teams.

33% of respondents say most cases of mule account handover at their institution involve coerced or manipulated customers. Another 30% say it’s roughly a 50/50 split between coercion and willing participation. Only 20% say most cases involve willing mules.

Credit unions stand out: they’re 3x more likely than all respondents to report that nearly all cases involve coercion.

This isn’t academic. It changes how you should treat a flagged account.

And here’s what the industry is actually doing: 51% of financial institutions respond to suspected handovers by temporarily restricting account activity or payment capabilities. The default playbook at most banks, neobanks, and fintechs is restricting the accounts of people who are already being exploited.

There’s logic to restrictions. Slowing things down protects funds, deters fraudsters, and forces a conversation with the customer. But that logic assumes the customer will engage honestly when you reach out.

Coerced customers often don’t. Someone caught in a romance scam may not believe they’re a victim. Someone under threat may lie to you because they’re scared. Someone who’s been socially engineered may actively work against your investigation because they’re still under the fraudster’s influence.

A bifurcated response isn’t just about being gentler with one group. It’s about designing a process that anticipates resistance from the very people you’re trying to protect.

Strategy: Build a bifurcated response model that separates willing mule handling from coerced victim intervention, and design for customers who may resist your help.

Action:

1. Review your current response playbook. Does it distinguish between suspected willing mules and suspected coercion?

2. For willing mules: restrict and investigate.

3. For suspected coercion: restrict if needed, but pair it with customer outreach that assumes the customer may not cooperate. Design the conversation accordingly.

4. The goal isn’t just containment. It’s intervention.

4. Your Detection Signals Are Probably Creating More Problems Than They Solve

This is the section that matters most if you run a fraud operations team.

Device identity is the most frequent cause of false positive alerts for mule account handovers, cited by 28% of respondents. Location verification is second at 24%. IP address flags sit at 23%. Behavioural biometrics at 22%.

Here’s the twist. Device identity and location intelligence are also the two most widely used detection signals across banks, fintechs, and payment processors. The tools these institutions rely on most are the same ones generating the most noise.

The report draws a sharp line. The problem isn’t the signal category. It’s the resolution.

Most implementations can’t tell the difference between a legitimate device upgrade and a handover. Or between a user who’s travelling and an account that’s changed hands. A device ID that resets every time the app is reinstalled or the phone is wiped creates a cascade of false alerts for what is, in most cases, normal customer behaviour. Location checks built on city-level IP data will flag a business traveller as suspicious. One built on behavioural patterns at higher precision can tell the difference between a customer who moved house and an account that’s been compromised.

53% of respondents say there’s a greater risk of false positives when detecting mule account handovers compared to other fraud types. That tracks. When your signals can’t separate real threats from routine behaviour, everything looks suspicious. Your investigators end up chasing ghosts instead of catching handovers.

Financial institutions don’t need different signals. They need more precise ones.

Strategy: Stop evaluating detection tools by signal category and start evaluating them by signal precision. The right instruments at the wrong resolution create more problems than they solve.

Action:

1. Separate signal category from signal quality in your next vendor review.

2. The question isn’t “do we use device intelligence?” It’s “can our device intelligence tell the difference between a factory reset and a new person controlling the account?”

3. Run a precision audit on your top three detection signals. How many alerts lead to confirmed mule activity versus false positives? That ratio tells you more about your stack than any vendor deck.

5. Where the Industry Is Putting Its Money (And Where It Might Be Wasting It)

78% of respondents say improving mule account handover detection is a high or top priority for the next twelve months. 52% call it a high priority with planned initiatives underway. 26% say it’s a top-tier priority with active executive sponsorship. Only 3% say it’s lower priority or not a strategic focus.

The money is coming. Whether it’ll be spent well is a different question.

Top investment priorities across banks, fintechs, neobanks, and MSBs:

• AI and machine learning: 51%

• Data integration across fraud, AML, and payment systems: 46%

• Network analysis: 37%

• Device integrity checks: 34%

• Location behaviour analysis: 30%

Each targets a different symptom. AI and ML can spot behavioural pattern changes post-onboarding. Data integration connects signals that live in silos. Network analysis identifies shared devices and locations across accounts. Device integrity catches tampering, emulators, and app cloners. Location behaviour reveals the geographic mismatch that signals a handover.

Together, they all try to answer one question: who is actually behind this account after onboarding?

But here’s where the industry needs to be careful. AI and ML topping the list at 51% sounds strategic. Sounds modern. But if your device ID can’t survive a reinstall, and your location data comes from IP addresses that anyone with a VPN can spoof, you’re building models on top of unreliable inputs. That’s not intelligence. That’s sophisticated pattern matching on noisy data. Garbage in, garbage out. No amount of ML fixes bad inputs.

Fix the signal quality first. Then model on top of it.

The institutions in that proactive 16% aren’t there because they threw AI at the problem. They’re there because they use behavioural analytics post-onboarding (41%), real-time monitoring (35%), and location intelligence (35%) together, as a single system. The lesson isn’t to buy all three. It’s to make whatever you already have work as an integrated whole.

Strategy: Fix the plumbing before you add AI. Clean, connected, precise data is the prerequisite for effective modelling, not the other way around.

Action:

1. Before adding another tool, audit whether your current tools actually talk to each other.

2. 46% plan to invest in data integration. That tells you the industry already knows its signals are siloed.

3. The first investment shouldn’t be AI. It should be plumbing. Connect the signals you have. Make your onboarding data available to post-onboarding monitoring. Make device intelligence talk to location intelligence.

4. Once the data is clean and connected, then layer on the modelling.

The full report goes deeper than I’ve covered here. Institution-specific benchmarks, technology breakdowns by organisation type, the US versus Europe split on regulatory concerns. If you’re building or reviewing your mule detection strategy, it’s worth the read.

Read the full report here →

This edition of the Payments Strategy Breakdown was created in partnership with Incognia.

Your compliance system verified the account holder’s identity once. The question is whether it’s still the same person.

The identity check verifies who is transacting. The merchant’s fraud model scores the session against hundreds of signals. The authentication challenge tests the claim.

Each layer does real work with real data. And each layer trusts that the credential being presented wasn’t compromised three months ago in a breach that the issuer hasn’t yet detected. When that trust is misplaced, every correct decision above it produces the wrong outcome. The identity check passes because the credential is genuine. The merchant’s fraud model sees a familiar card range on a returning customer’s device. The authentication challenge succeeds if it fires at all. The transaction clears. The merchant ships. The chargeback arrives sixty days later.

The Identity Theft Resource Center recorded 3,158 US data compromises in 2024, with roughly 1.7 billion victim notices driven by six mega-breaches, including Ticketmaster, Change Healthcare, and AT&T. Stolen credentials were the leading attack vector in cyberattacks on publicly traded companies that year. Four of the six mega-breaches were credential-driven.

Tens to hundreds of millions of stolen card records circulate on dark web marketplaces at any given time, priced at roughly $5 to $15 per card. According to Recorded Future’s lifecycle analysis, a compromised card is typically used for fraud within three to four days of purchase, after a checker-service validation step that costs 20 cents. The credential layer is the foundation of every trust decision in the stack. It’s also the layer the industry treated as solved while it was quietly becoming the weakest link.

The Architecture Built for Plastic

The Primary Account Number, the long card number printed on the front, is a permanent, reusable, non-revocable identifier. It was designed for a world of plastic cards and carbon-copy imprinters. It was never designed to be stored in merchant databases, embedded in recurring-billing engines, or typed into browsers. But that’s exactly what happened. The PAN became the default key for card-on-file storage, subscription billing, and every card-not-present transaction the e-commerce economy runs on. A static identifier, permanently bound to an account, is replicated across every merchant that has ever stored it.

The Card Verification Value is a three-digit static secret printed on the back of the card. PCI DSS prohibits storing it after authorization. Breaches continue to surface CVV alongside PAN anyway, because compliance is a posture, not a guarantee.

The CVV has two structural weaknesses that I want to name directly. The first is the Newcastle University Distributed Guessing Attack, published in 2016. Researchers demonstrated that with a valid PAN, an attacker can recover the expiry date in roughly 60 guesses and the CVV in under 1,000 guesses by distributing attempts across hundreds of e-commerce sites. Each site enforces its own lockout threshold. None of them shares velocity data with the others. The attack recovered a full set of card credentials in under six seconds. The original research found Visa’s network vulnerable. Mastercard’s centralized detection caught it within roughly ten attempts, even when distributed. No public scheme statement claims this structural weakness has been closed. The global web security community (OWASP) still lists card cracking as an ongoing automated threat in 2026.

The second weakness is simpler. The CVV isn’t required for all online card transactions. Recurring payments, subscriptions, merchant-initiated charges, and some tokenized flows don’t require it. The three-digit code on the back of your card doesn’t even cover the full range of online card usage.

The credential architecture was built for a world where a human holds a card, physically presents it, and the static values on it are secret. Two of those three assumptions are already gone. The third is under structural pressure from a direction the original architects never anticipated.

The card industry’s security standard, PCI DSS 4.0, had its latest requirements take effect on March 31, 2025. It tightened controls around how card data is stored, transmitted, and accessed. No public compliance rate has been published since the deadline passed. The standard itself is rigorous. But the enforcement model relies on self-assessment for most merchants and third-party auditing only for the largest. PCI measures whether organizations have the right controls in place. It does not measure whether those controls actually prevent card data from being stolen. The distinction matters. Every major breach that surfaced card credentials in the ITRC dataset occurred at an organization that was, at the time of the breach, either PCI-compliant or in the process of attesting to compliance.

The Rotation Spectrum

The industry’s answer to the static-credential problem comes in two forms. Both are attempts to stop trusting static values. They sit at different points on the same architectural spectrum, are deployed by different actors, and operate under different economic models. Neither was designed to complement the other.

Network tokenization is the dominant response. Instead of sending a card’s real number to the merchant, the network replaces it with a token that only works for that specific merchant or channel. If the token gets stolen, it’s useless elsewhere. Visa reported 13.7 billion tokens issued cumulatively as of late 2025, with roughly 50% of Visa's global e-commerce transactions now using tokens instead of real card numbers. Mastercard reported that roughly 35% of all transactions would be tokenized by mid-2025.

Both networks frame tokenization as a Value-Added Services revenue line and disclose growth rates that outpace those of core processing revenue. Visa reports a 28-40% reduction in fraud for tokenized versus non-tokenized online transactions, depending on the analysis window. Mastercard reports similar directional improvement. Both figures are network-published and should be read accordingly.

Tokenization has made the credential layer meaningfully better. But it hasn’t made it honest. The real card number still exists in the network’s vault. The token is a stand-in, not a fix for the underlying architecture. And it creates a downstream effect the industry hasn’t fully reckoned with. When a token is present, the bank’s fraud-scoring system treats the transaction as more trustworthy. The merchant’s fraud model relaxes too. The token looks clean, so the authorization looks clean, so the approval comes through. That’s the right outcome when the underlying customer is legitimate. When they aren’t, when the account was opened with a fake identity or has been taken over, tokenization actually makes the fraud harder to catch because the token itself is genuine. The same mechanism that helps legitimate customers can help a bad actor who controls a valid card.

A Visa token can’t be used on the Mastercard rail. Cross-network tokenization remains fragmented. India’s Reserve Bank mandated tokenization in October 2022, banning merchants from storing raw card numbers and pushing adoption of tokenization on consumer devices and bank-side vaults. It’s the only major market where a regulator forced tokenization as a baseline rather than leaving it to commercial incentives. Everywhere else, tokenization adoption is voluntary, commercially driven, and accelerating inside each network while interoperability across networks remains unresolved.

Dynamic credential rotation is the other architecture. It replaces the static printed CVV with a dynamic value. Two deployment models exist. Card-side display puts an e-paper screen on the back of the card, refreshing the code every one to four hours. IDEMIA’s MOTION CODE has been deployed across issuer programs in France, Spain, and Malaysia, as well as embedded-finance platforms, with over 600,000 cards issued by 2018. App-side generation produces a biometric-protected dynamic code through the issuer’s mobile banking app.

SafeCypher’s deployment with An Post Money in Ireland is the most publicly documented example of this model, with An Post and SafeCypher jointly reporting no CNP fraud through the An Post Money App over an 18-month period.

IDEMIA has published a similar zero-CNP-fraud claim across its MOTION CODE deployments. Both claims are vendor-issued. The transaction volume denominators are not publicly disclosed. The structural argument doesn’t rest on zero-fraud claims. A rotating secret tied to a separate channel collapses the three- to four-day breach-to-fraud reuse window that makes stolen static credentials monetizable. That argument is independently defensible.

The gap between these two architectures isn’t technical. It’s commercial. Tokenization sits inside the networks’ Value-Added Services revenue lines. Visa and Mastercard report token growth to investors. Dynamic CVV is an issuer-side cost with no network revenue share. No US top-25 issuer by card volume has publicly disclosed at-scale dynamic CVV deployment. The technology is mature. The barriers are commercial, not technical. The credential layer’s two responses to the same structural problem are deployed by different actors with different economics, and the one with the network flywheel is winning.

Both are points on the same spectrum. A network token is a long-lived, rotating identifier. A dynamic CVV is a short-lived rotating secret. Neither was designed to complement the other. Neither addresses what happens when the entity presenting the credential isn’t human. And neither is measured against a common outcome metric. The credential layer has two partial answers, two different owners, and no shared measurement of whether either one is working at the system level.

When the Presenter Isn’t Human

The credential model was designed for a human typing a card number into a browser or tapping a card against a terminal. That assumption is under structural pressure from AI agents, software that can browse, compare, and buy on behalf of a human, which are beginning to initiate transactions without the cardholder being present at all.

Visa launched its Trusted Agent Protocol in October 2025 with Cloudflare and twelve launch partners, including Adyen, Stripe, Worldpay, and Shopify. TAP is open-spec, published on GitHub, and uses HTTP Message Signatures with a Visa-operated public-key directory to carry Agent Intent, Consumer Recognition, and Payment Information. No money moves over TAP itself. It’s an identity-of-agent layer that composes with the existing payment rails.

Mastercard introduced Agent Pay in April 2025, extending its tokenization infrastructure with Agentic Tokens. The September 2025 expansion added Insight Tokens for permissioned data-sharing and Verifiable Intent for transaction-level authorization. Google’s Agent Payments Protocol (AP2) has sixty-plus coalition members, including Mastercard, Visa, PayPal, Stripe, and American Express, and is live with select merchants. EMVCo acknowledged in November 2025 that its specifications for 3DS, Payment Tokenization, and SRC need to be developed and enhanced for agentic payments. That acknowledgment is itself the evidence. The layers were not designed to compose.

These are real efforts at a real problem. TAP, Agent Pay, AP2, and EMVCo’s emerging workstream are all attempts to retrofit the missing system layer onto a stack that was never designed to compose. But the list of what hasn’t been solved is longer than the list of what has.

Who pays when an agent makes a purchase that the customer didn’t authorize?

That remains formally unresolved at the scheme level. The payment message format that banks use to talk to each other, a standard designed decades ago, has no field for “this transaction was initiated by an AI agent acting on behalf of this person.” How fraud-detection models should interpret software behavior rather than human behavior remains unresolved. No regulator in India, Singapore, Australia, Japan, or Hong Kong has published a public position on agent-initiated payments specifically.

The credential-rotation spectrum directly surfaces this tension. A dynamic CVV on a physical card requires a human to read a code. That’s fundamentally incompatible with an AI agent making a purchase. An app-generated dynamic CVV with biometric approval could, in theory, work with delegated agent flows, but no one’s tested it in production. Network tokens, because they’re already digital and machine-readable, naturally compose with agent-initiated payments. Agentic Tokens are explicitly built on that foundation. The credential architecture that’s winning on commercial adoption is also the one that works with the agentic future. The architecture with the stronger argument against breach-driven fraud may not survive the transition to machine-initiated payments.

The regulatory picture is equally unresolved. Europe’s upcoming PSD3 regulation includes expanded liability for payment providers when customers are defrauded through impersonation, but doesn’t specifically address agent-initiated payments. The US has no equivalent federal mandate. The gap between where regulators are and where the technology is heading is measured in years.

The volume reality is honest about where this stands. As of late 2025, Visa reported “hundreds” of completed agent-initiated transactions across all production partners. The forecast is millions by the 2026 holiday season. That’s a forecast, not a fact. The structural break is real. The commercial volume is still measured in pilots.

The Layer Nobody Measures

Every other layer in the risk stack has a measurement regime, however imperfect. Authentication has fraud-rate thresholds set by European regulators, with specific ceilings that determine whether a bank can skip the extra verification step. The merchant layer has chargeback monitoring programs like Visa’s VAMP and Mastercard’s EFM, with defined thresholds, escalation schedules, and financial penalties. The identity layer has at least conceptual assurance levels, even though they don’t travel in the payment message.

The credential layer has PCI DSS attestation. It’s binary. It’s self-assessed at most levels. And it measures whether you have the right security controls in place, not whether those controls are actually preventing card data from being stolen.

There is no public metric for how well the credential layer is actually performing. No scheme program penalizes or rewards credential-layer security the way VAMP penalizes merchants for chargebacks. No regulator publishes credential-compromise rates the way the European Banking Authority publishes fraud rates by authentication method. The credential layer is the only layer in the risk stack with no outcome measurement whatsoever.

The industry measures what it penalizes. It penalizes fraud at the merchant layer, friction at the authentication layer, and identity failure at the onboarding layer. It does not penalize credential compromise at any layer. And it does not measure the gaps between them at all.

What Five Layers Taught Us

The risk stack was not designed. It was accumulated.

Identity verification was solved by KYC providers and credit bureaus. But there’s no standard way for the result of that check, “this is a real person, we’re confident,” to travel inside the payment message when a transaction happens months or years later. The identity layer has no owner, no standard, and no way to measure whether its decisions still hold at the moment of purchase.

Merchant fraud decisioning was addressed by data-enrichment vendors and merchant data science teams. The merchant knows more about the customer than anyone else in the chain, evaluating hundreds to thousands of data points at checkout. But the payment message that reaches the bank carries only a fraction of that context. The richest signal in the chain dies at the handoff.

Authentication was solved by 3DS, the protocol that produces those “verify your purchase” prompts. The protocol is identical for every bank. But every bank is choosing which cost to absorb: fraud, conversion, or compliance. The same protocol produces radically different outcomes depending on geography and strategy.

Credential security was addressed through network-level tokenization and issuer-level dynamic credential rotation. Tokenization has a network revenue line. Dynamic CVV has an issuer cost line. There is no commercial alignment between them and no shared way to measure whether either is working.

Each layer was solved by the entity with the strongest commercial incentive to do so. None of them was built to work with the others.

That’s the pattern this series has documented across five issues. The Nilson Report trajectory that opened this series, $33.41 billion in global card fraud in 2024, heading toward $41.06 billion by 2030, is not evidence that any single layer is failing at its own job. Each layer is doing exactly what it was designed to do. The problem is that nobody designed the connections between them.

The false decline problem that opened Issue 1, $50.7 billion in refused legitimate revenue across four markets, has the same root cause. The risk stack penalizes fraud approval and ignores trust refusal. It optimizes each layer individually, resulting in an overall outcome worse than the sum of its parts. A compromised credential passes every upstream check. A legitimate customer on a flagged card range fails them. Both outcomes are produced by a stack that’s working exactly as designed. The layers aren’t bad. The system they’re supposed to be part of was never built.

Why This Matters Beyond Risk

If like me, you’ve spent your career on the optimization side of payments, on authorization rates, processing speed, interchange management, and routing efficiency, this series might have felt like a detour into someone else’s problem. It isn’t.

The payments industry has spent two decades getting extraordinarily good at processing transactions. Faster rails, smarter routing, better data, higher approval rates. That work is real, and it matters. But it was done while the risk infrastructure underneath it was accumulating without design, each layer solving its own problem, each owned by a different commercial actor, none of them measured as a system.

The professionals who only see the optimization side are missing half the picture. They’re building on a foundation they haven’t inspected. The ones who only see the risk side are missing the other half. They’re hardening layers without asking whether those layers connect to the business outcomes that pay for everything.

The companies that will navigate the next decade well are the ones that understand it’s about balance. Authorization optimization and risk architecture are not separate disciplines. There are two views of the same transaction. An approval rate means nothing if the fraud that clears it costs more than the revenue it generates. A fraud-prevention system means nothing if it refuses more legitimate customers than fraudsters. Both problems have been accelerating simultaneously, and this series has traced the structural reasons why.

EMVCo’s Digital Identity and Payments Task Force, announced in early 2026, is the most credible attempt to address the composition problem at the standards level. Visa’s Intelligent Commerce, Mastercard’s Agent Pay, and Google’s AP2 are each approaching it from their own position. These are real efforts. But they’re also, structurally, the next layers being added to a stack that has never been designed as a system.

The question this series leaves you with isn’t technical. It’s organizational. Who in your company owns the full picture, from identity to credentials, from fraud prevention to customer trust? If the answer is “different teams, different budgets, different KPIs,” you now know what that produces. You’ve seen it across five layers and five issues.

The risk stack doesn’t need better layers. It needs someone to own the system.

Thank you for reading.

This edition was made in partnership with SafeCypher, which provided its expertise, insights, and reports to help me better understand the challenges with credential security.

P.S. If reading this series made you wonder where your own authorization chain is leaking value, that’s exactly what I help institutions figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally, through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these deep dives, buy me a coffee.

Nothing in the protocol explains that. EMV 3DS 2.x is identical for every participant. The message structure is defined the same way in Sydney and São Paulo and Singapore. The certification test plan is the same. What changes at that border is not the technology. It’s which equilibrium the issuer is positioned for.

That’s the layer this issue is about. Authentication is where the issuer tests the claim the merchant has sent. The protocol is standardized. The economics are not. And the gap between those two facts is where most of the layer’s behavior actually lives.

Where the merchant’s signal goes to be tested

The merchant holds the richest fraud-relevant context in the chain. Almost none of it travels. The thin slice that does arrives at the issuer’s authentication layer as a 3DS authentication request, and that request is where the issuer’s stack is supposed to test what the merchant has claimed.

The quality of that test is not a property of the protocol. It’s a property of the issuer’s strategic position, expressed operationally. Two issuers can receive the same request, on the same network, for the same merchant, from the same cardholder, and answer it in completely different ways. That’s not a bug. The protocol leaves the answer to the issuer.

The issue is that most issuers haven’t thought about it as an answer they’re choosing.

Authentication is not a protocol the issuer runs

EMV 3DS 2.x defines one identical message set, one SDK schema, one publicly maintained registry of approved products. Every certified ACS, every certified 3DS Server, every accredited Directory Server passes the same EMVCo test plan. The protocol leaves nothing technical to interpretation. AReq, ARes, CReq, CRes, RReq, RRes. The same on every rail.

And yet authentication outcomes vary across issuers running this identical protocol by orders of magnitude. Frictionless approval rates that range from below half to above ninety-five percent. Challenge abandonment that varies from a few percent to a third of all challenges. False challenge rates that no one can quantify because nobody measures them consistently, but that practitioners working across PSPs will tell you span ranges that no protocol could produce on its own.

The cause is choice. Each issuer is choosing how aggressively to apply exemptions, which ACS to run, which fraud model behind it, which challenge type to deploy, which fraud-versus-conversion tradeoff to live with. None of those choices are dictated by the protocol. The protocol is the same. The economics are not.

The structural argument the rest of this issue lands on is this. Authentication is where the issuer chooses which kind of cost to absorb. There are three options. There is no neutral position.

Three costs, no neutral position

The first cost is fraud absorption. The issuer accepts that some authenticated transactions will turn out to be fraudulent, treats the losses as an acquisition or retention expense, and runs a permissive authentication posture in exchange for higher conversion. Fintechs and BaaS-issued cards skew strongly toward this position. The frame is conversion preservation with a fraud constraint.

The second cost is conversion absorption. The issuer accepts that some legitimate cardholders will abandon at the challenge or be rejected by a conservative ACS, treats the revenue leakage as the price of low fraud rates, and runs a more cautious posture. Tier-one incumbent banks skew toward this position, particularly in markets where the issuer’s fraud rate is the public KPI. The frame is fraud minimization with a conversion constraint.

The third cost is compliance absorption. The issuer accepts that the regulator’s mandated factor selection or fraud-rate ceiling dictates the operational posture, and runs whatever protocol path keeps the firm out of supervisory trouble. Indian issuers post-RBI 2025, Singapore banks post-MAS digital-token mandate, Hong Kong banks post-HKMA “Authenticate in-App,” Japan post-METI all sit here. The frame is regulatory survival with everything else as a constraint.

The European Banking Authority’s reference fraud rate ladder is the cleanest single illustration that the architecture itself defines these cost lines. To apply the Transaction Risk Analysis exemption on a remote card transaction, the PSP claiming the exemption must keep its reportable fraud rate below 0.13% on payments up to €100, below 0.06% up to €250, and below 0.01% up to €500. Above €500, no TRA exemption is available at all. Those are not technical thresholds. They are cost ceilings. Whichever side of the message wants to run frictionless on more volume has to keep its fraud rate under those numbers. Whichever side runs over loses the lever.

The teams that read those numbers and the teams that bear the consequences of breaching them are usually not the same teams. The fraud team’s KPI is fraud rate. The cards team’s KPI is authorization rate and interchange. The compliance team’s KPI is no enforcement action. None of those three sees the conversion-cost-of-false-challenge that the merchant ultimately bears, because the merchant isn’t a stakeholder inside the issuer.

The ACS is a strategic asset most issuers buy as a checkbox

The Access Control Server is the issuer’s component in the 3DS flow. It receives the authentication request, decides whether to challenge, runs the challenge if needed, and returns the cryptogram. ACS quality determines latency, challenge logic, fraud detection at the authentication moment, and the user experience of the challenge itself. It is a strategic asset.

Almost no issuer builds one. Almost every issuer buys one. The vendor market splits along a legacy-versus-modern axis that maps onto the issuer-type split with surprising consistency.

On the legacy side sit Broadcom’s Arcot stack, Outseer (the carve-out of RSA’s payments authentication business), parts of the Modirum-derived deployments now consolidating under Entersekt after its December 2023 acquisition, GPayments, and the heritage on-premise deployments inside tier-one banks that were sized for 3DS 1.0 and migrated. CardinalCommerce, acquired by Visa for $300 million in December 2016, sits in this neighborhood, dominant by volume on the merchant and acquirer side rather than the issuer ACS side.

On the modern side sit G+D Netcetera, first to certify globally on EMV 3DS 2.3.1 in 2023 and deployed across more than a thousand issuers; Worldline on its modernization track; Entersekt-Modirum on its modernization track; and the BaaS-native multi-tenant entrants like Apata, Verestro and M2P, bundled into program-manager stacks for fintech and neobank issuance.

Mastercard and Visa each operate their own network ACS stand-ins for the moments when the issuer’s ACS is unavailable. Stand-in is a third category. The network answers on behalf of the issuer, and the issuer typically still bears the liability.

The ACS contract is one of the highest-leverage decisions an issuer makes about authentication. It is rarely owned by the team that owns the authorization rate. The choice locks in a fraud-versus-conversion position the rest of the issuer’s payments organization has to live with, and the team that signs the contract is usually procurement working from a vendor RFP, not the cards-P&L owner reading the operational consequences.

Exemption strategy is where the position becomes economically visible

TRA is predominantly an acquirer-side lever. The acquirer, working with the merchant, decides at the moment the authentication request is constructed whether to flag the transaction as TRA-eligible. The acquirer is the party trusting its own fraud signal enough to claim the exemption, and the acquirer is the party whose reportable fraud rate degrades if the call is wrong. The issuer holds the parallel lever on its own side of the message. Through the ACS risk-based decision, the issuer chooses frictionless approve, challenge, or decline. Both levers are forms of the same strategic choice. They sit on different sides of the authentication message.

Visa Europe has estimated publicly that 40 to 50 percent of European e-commerce by volume could qualify for TRA exemption if criteria are met. The actual frictionless rate that the layer produces sits well below that ceiling. The gap between what the regulation permits and what either side actually claims is wide.

The architecture itself adds a cruel mechanic. Under EBA Q&A 2019_4702, the issuer’s reportable fraud rate includes unauthorised transactions regardless of which side applied the exemption. If an aggressive acquirer applies TRA broadly across its merchant book and chargebacks come back, the issuer’s rate degrades on transactions the issuer never approved an exemption on. The regulator itself acknowledged the perverse incentive in the Q&A response. The issue persists. One side’s appetite for the exemption can quietly cost the other side its future ability to claim it.

The lever everyone can technically pull is one many won’t pull, because pulling it badly costs them their future ability to pull it well. Conservative posture is the safe equilibrium. The cost shows up downstream as conversion nobody traces back to authentication.

Same protocol, four different equilibria

The cleanest proof that authentication is a position, not a protocol outcome, is the geographic comparison. Four major markets, the same EMV 3DS 2.x specification, four structurally different equilibria.

The United States absorbs CNP fraud at the issuer level. There is no Strong Customer Authentication mandate. 3DS adoption on US-issued cards is dramatically lower than on European-issued cards. Federal Reserve Kansas City research, in its most recent February 2026 update on 2023 Federal Reserve Board data, shows the US card-not-present fraud rate continuing its upward trend across both the dual-message networks (Visa, Mastercard, Discover) and the single-message networks (Pulse, Star, NYCE). The Pulse debit issuer benchmark CNP rate climbed from 26.1 basis points in 2019 to 41.6 basis points in 2023, and the trajectory has not reversed. Industry estimates put CNP fraud at roughly 71% of total US card fraud loss, in the neighborhood of $10 billion in 2024. The US equilibrium is issuer fraud absorption, with the merchant’s choice of whether to invoke 3DS as the only authentication lever in the system. Stripe’s analysis of US 3DS performance suggests some US issuers may treat a 3DS request itself as a fraud signal, declining 3DS-flagged transactions at higher rates than non-3DS flow because they read the request as confession of merchant suspicion. Same protocol, opposite signal interpretation.

Europe absorbs conversion at the issuer and acquirer. PSD2 SCA, the EBA Regulatory Technical Standards, the TRA exemption ladder. The 2025 EBA-ECB Joint Report on Payment Fraud documents €1.329 billion in card fraud losses across EU/EEA issuers in 2024, up 29% year over year on a base where SCA has measurably suppressed unauthorised remote fraud. The same report finds the fraud share roughly ten times higher when the counterparty sits outside the EEA, where SCA does not legally apply. The European equilibrium is conversion absorption inside the perimeter and a hard fall-off at the perimeter edge.

Latin America is converging on regulator-managed device authentication. Brazil’s central bank has issued Pix-scheme rules, Normative 491 capping unregistered-device transactions at R$200 per transaction and R$1,000 per day, and a cybersecurity package effective March 2026. Mexico’s CNBV fraud-prevention regulation took effect June 2024. Peru’s SBS Regulation 2286-2024 mandates two-factor for all card transactions. The pattern across the region is a regulator that reaches into device-binding and identity-validation rather than relying on a CNP-card SCA mandate.

Asia-Pacific is uniformly retreating from SMS OTP. India’s RBI Authentication Directions, issued September 2025, mandate two-factor for all domestic digital payments by April 2026 and risk-based controls for cross-border CNP by October 2026. Singapore’s MAS announced in July 2024 a phase-out of SMS OTP for digital-token-enabled customers. Hong Kong’s HKMA has made in-app authentication the default for online card transactions since late 2024. Japan’s METI and Japan Credit Association mandated EMV 3DS for all e-commerce as of April 2025, on a base where 92.5% of Japanese credit card fraud was CNP. Korea’s residual ARS-and-financial-certificate stack persists as the canonical example of how legacy authentication infrastructure produces structurally different friction for foreign cards.

Four regions. Same protocol. Four different answers to the question of which cost to absorb. And then there is the cross-border picture, where these equilibria collide.

When a US-issued card transacts on a European-acquired merchant, the US issuer is not bound by SCA. The merchant applies 3DS as fraud control. The issuer’s home regime governs the response. When an Indian-issued card transacts on a European merchant after October 2026, RBI’s cross-border AFA rule may require the Indian issuer to validate an Additional Factor that has no current European analogue. When an Australian cardholder transacts on an overseas merchant, the same cardholder who barely registers any CNP fraud domestically registers an order of magnitude more on the cross-border slice, because domestic and overseas authentication postures are positioned for different equilibria. The seams between equilibria are where authentication-layer breakage actually shows up.

The challenge moment is itself a strategic choice

The cardholder doesn’t see any of the protocol mechanics. The cardholder sees the moment the issuer asks them to prove they are who they say they are. Sometimes that’s a text message with a code. Sometimes it’s a fingerprint or face scan inside the bank’s app. Sometimes it’s a tap to approve. Each of those choices produces a different abandonment rate, a different fraud rate, and a different cost.

Outside the United States, regulators have stopped letting issuers make this choice freely. India’s central bank has effectively pushed banks off SMS codes. Singapore’s regulator has done the same for digital-token customers. Hong Kong has made the bank’s own app the default. Brazil and the rest of Latin America are following. The regulatory direction is uniform: SMS is not safe enough anymore, and the bank’s own app is where authentication is supposed to live.

The United States is the outlier. Most US issuers still rely heavily on SMS or no challenge at all. That’s a strategic position, even when nobody at the issuer has framed it as one.

The networks profit from heterogeneity

Visa and Mastercard each operate their own ACS stand-ins and sell network-private value-added services that wrap around the authentication moment. Mastercard’s Identity Check is the brand wrapper around its 3DS, and Smart Authentication Stand-In handles the cases where the issuer’s ACS is unavailable. Visa Authentication Suite and Visa Risk Manager play the same role on the other rail. Mastercard’s value-added services and solutions reached $3.4 billion in net revenue in the third quarter of fiscal 2025, up roughly 25% year over year, with a meaningful slice tied directly to authentication and fraud scoring. Visa’s equivalent line has grown in the high teens. Each layer of fragmentation creates a value-added service the network can sell. Mandate coherence and you commoditize the revenue.

The networks have built infrastructure around 3DS, not over it. They earn from the seams between issuers, not from closing those seams. Coherence at the authentication layer would commoditize the very services they sell to cope with the lack of coherence. The fragmentation is the business.

The measurement vacuum at the authentication layer

There’s a number nobody measures. The legitimate cardholder, challenged unnecessarily, who then abandons or fails or calls customer service or simply walks away. Call it the false challenge rate. It’s the authentication-layer cousin of the false decline, and like the false decline, it doesn’t appear on any standard report because no party in the chain has both the data and the incentive to publish it.

The issuer can sometimes infer false challenges from chargeback absence over time. The merchant sees abandonment without knowing whether the abandoner was a genuine cardholder or a fraudster who balked. The PSP sees the outcome but not the inside of the issuer’s decision. The cardholder, the only party who actually knows, is silent. There is no industry-standard metric for “challenge issued where the cardholder was the genuine cardholder, who would have completed without it.” There is no scheme report on it. There is no regulatory disclosure on it.

Which metrics an issuer chooses to optimize is itself the strategic position. An issuer watching only the chargeback rate optimizes toward conservative challenge. Every challenge issued that gets approved looks like a successful intervention even when the cardholder was genuine and would have completed without it. An issuer watching conversion-net-of-fraud optimizes toward aggressive frictionless approval. Every false challenge avoided shows up as conversion preserved. Most issuers optimize the first KPI and ignore the second, because the team that owns the second sits in a different building with a different P&L.

You can’t penalize what you can’t measure. The false-challenge tax exists because nobody with the data has the incentive to publish it.

The position is mostly the residue of teams not talking

At most issuers, the strategic authentication position is not something anyone deliberately chose. It’s the operational residue of teams that don’t share P&L language.

ACS configuration is owned by the fraud team or e-banking risk. Exemption thresholds are nominally owned by fraud, but the consequence of those thresholds shows up in the cards-P&L owner’s authorization rate. EBA fraud rate reporting is owned by regulatory compliance, with data assembled from the systems the fraud team runs. The cards-P&L owner reads quarterly auth rates without seeing the ACS configuration that produced them. The fraud team reports its rate to compliance without seeing the conversion impact.

Cross-team coordination is the structural failure point. The teams collectively make the strategic position by accident. The fraud team’s appetite for false positives, the procurement team’s choice of ACS vendor three years ago, the compliance team’s interpretation of EBA Q&As, the risk committee’s tolerance for chargebacks. All of those individual decisions roll up into a strategic posture that nobody in the room ever described as a strategic posture.

This is the structural problem worth naming. Each layer inside the issuer optimizes inside its own data, its own KPI, its own reporting line. The aggregate outcome is worse than the sum of those parts. At the authentication layer, the result is a position that nobody is accountable for. The consequences look like accidents and behave like positions.

The choice is being made whether the issuer notices or not

Here’s what’s actually happening at the authentication layer. Every issuer in the world is running the same protocol. Every issuer is producing different outcomes. Those outcomes are not random and they are not technical. They are the cumulative result of choices the issuer has made about exemption appetite, ACS vendor, challenge type, fraud-rate target, and the implicit weight given to conversion against fraud against regulator. The issuer running aggressive frictionless approval, low abandonment, and a fraud rate at the EBA ceiling has chosen one position. The issuer running conservative challenge logic, higher abandonment, and a fraud rate well under the ceiling has chosen another. The issuer that defaults to scheme stand-in because its ACS is unreliable has chosen a third by accident.

The cost of the choice is paid by three different parties on three different timelines. The cardholder pays in friction. The merchant pays in conversion. The issuer pays in fraud and in regulatory exposure when the rate drifts. None of those three costs sit on the same P&L. None of them surface in the same quarterly review. None of them get reconciled in any forum where the choice could be revisited deliberately.

That is the structural challenge of the authentication layer. The decision is being made. It just isn’t being made by anyone who can see what it costs.

The next layer down is the credentials themselves. Issue 5 picks up there.

Thank you for reading.

P.S. If reading this series made you wonder where your own authorization chain is leaking value, that’s exactly what I help institutions figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally, through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these deep dives, buy me a coffee.

The merchant carries the consequence of that decision in both directions.

If the approval was wrong, the merchant absorbs the chargeback, the cost of goods, the shipping, the interchange already paid, and exposure under Visa’s Acquirer Monitoring Program. If the decline was wrong, the customer leaves, often without ever telling the merchant why, and never comes back. Both outcomes cost the merchant money. Only one of them appears on the dashboard.

That asymmetry shapes how the decision is actually made. The fraud model doesn’t sit in a vacuum. It sits inside an operation that pays for being wrong twice and gets credit for being right once. Everything downstream of that follows from that single structural fact: the model architecture, the threshold settings, the rule layer, the choice to outsource, the chargeback guarantee contract, and the vertical-specific calibration.

This issue takes that decisioning operation apart from the inside.

Let’s break it down.

What An Enterprise Merchant Actually Sees In A Transaction

The fraud-relevant context available at checkout is not comparable to anything else in the authorization chain. The merchant sees the customer in a way the acquirer doesn’t, the network can’t, and the issuer never will.

The signals fall into six categories:

• Identity. Name, email, phone, billing address, shipping address, account reference, and account tenure with the merchant. None of this is novel on its own. What is novel is the merchant’s ability to cross-reference these against its own customer history to a depth no one else has. A returning customer who has placed 43 prior orders over 6 years on the same device is a different proposition than a first-time email subscriber who signed up an hour before checkout. Both arrive at the issuer as the same fifteen-digit account number.

• Behavioral. This is where the merchant pulls ahead. Adyen’s risk field reference is the only first-party PSP documentation that explicitly names interaction patterns with the payment form as a model input. That phrase covers mouse cadence, typing rhythm, copy-paste detection, time spent on each page, navigation path before checkout, and whether the cart was assembled in one session or three. None of this travels in any standardized payment message.

• Device. IP address, anonymous IP, and proxy flags, full browser fingerprint, persistent device identifiers across sessions. The merchant gets these natively at checkout. The issuer gets a partial subset only when the merchant invokes 3DS.

• Network. IP geolocation, IP-to-card consistency, ASN reputation, and the relationship between the device’s network and the customer’s stated address. Run against the merchant’s own database of prior good and bad transactions, plus the cross-merchant network data that its fraud vendor brings.

• Transactional. Amount, currency, MCC, item-level data, payment method, card BIN, descriptor. These mostly travel in the authorization message. They are also the least informative on their own.

• Historical. Velocity counters keyed to the card, email, IP, device, customer reference, and shipping address. Stripe’s Radar exposes more than 40 velocity attributes in its rule editor, such as dispute_count_on_card_number_yearly and total_customers_with_prior_fraud_activity_for_card_yearly. Stripe states that roughly ninety percent of cards on its network have been seen before. That depth of historical context exists only at the merchant or its vendor; it is impossible to reconstruct it downstream.

Add it up, and a serious enterprise fraud model evaluates several hundred to a couple of thousand features per transaction. A typical card-not-present authorization message that reaches the issuer populates between twenty-five and sixty fields out of the 128 or 192 defined in the underlying ISO 8583 standard. The compression is not subtle.

The issue isn’t about what does or doesn’t survive the handoff. The issue is about what the merchant does with that signal richness once it has it. And the answer is that the merchant runs the most informationally rich fraud check in the entire payments chain, and then has to decide how aggressive to be with it under a cost structure that punishes both directions.

How The Decision Actually Gets Made

A fraud model is, mechanically, a probability classifier. The feature vector goes in. A number comes out, almost always between zero and one, almost always thresholded into bands.

The output is fundamentally the same across the industry: a probability, a band, and an action.

• Stripe Radar. Numeric risk score from zero to ninety-nine, mapped to bands of normal, elevated, and highest. Default policy thresholds sit at sixty-five and seventy-five.

• Adyen RevenueProtect. Result type of green, amber, or red, plus a risk level from very low to very high, plus a numeric score. Legacy rule engines block when the accumulated rule score reaches 100 or higher.

The interesting question is who picks the threshold.

Stripe’s Radar Technical Guide is unusually direct about this. The data scientist’s job is to push the model’s accuracy curve outward by building better features. The business’s job is to pick a policy point on whatever curve the data scientist hands over. Two different problems. Two different owners.

The point on the curve is chosen based s acceptable cos’ acceptable cost of being wrong in each direction. A merchant willing to absorb more chargebacks in exchange for higher approval rates raises the threshold. A merchant under VAMP scrutiny, or with thin margins that make every fraud loss a meaningful drag on contribution, moves it down. Neither setting is right nor wrong. They reflect different business positions.

That choice is then complicated by the rule layer. Almost no enterprise merchant runs the model output as the sole decisioning input. Rules sit on top:

• BIN allow and block lists

• MCC carve-outs and geographic restrictions on shipping addresses

• Mandatory 3DS step-up over a certain amount

• Manual review queues between specific score thresholds

• Vertical-specific overlays for high-risk SKUs

• Seasonal calibration during peak demand windows

The rule layer typically belongs to the fraud operations team, separate from the data science team that owns the model. The two teams talk to each other. They are not the same team.

The split is fairly standard at scale. Data science owns the model, retraining cadence, feature engineering, and threshold experimentation. Fraud operations owns the rule layer, the manual review queue, chargeback representment, list curation, and threshold tuning by SKU, region, and season. At the largest merchants, there’s also a payments team owning authorization rate optimization through network tokens, account updaters, and retry strategies.

That last team is functionally separate from fraud, and the distinction matters. Fraud detection and authorization rate are distinct KPIs, owned by different teams and optimizing different curves. A tool that improves one does not necessarily improve the other.

Inside this architecture, the merchant-layer false-positive rate is what most enterprise merchants are quietly carrying. Merchant Risk Council’s 2024 global fraud and payments report puts the typical merchant rate between two and ten percent of total e-commerce orders, with one in five merchants reporting rates above ten percent. That is the rate of legitimate transactions that the merchant’s own fraud check is refusing, separate from anything the issuer subsequently does.

Whether that rate is two percent or twelve percent depends entirely on where the threshold is set, which depends on the cost structure the merchant is operating under.

Build, Buy, Or Outsource: The Decision

The merchant’s first strategic question, before any threshold gets picked, is who builds the model. Four answers exist in practice:

• In-house at the largest merchants. Uber has published a series on its fraud detection systems, including Project RADAR and the Mastermind rule engine; its Michelangelo machine learning platform lists payment fraud detection and chargeback prevention as core production use cases. Booking.com’s data science team has published peer-reviewed work on tabular transformers for fraud detection. Amazon, the large airlines, and the largest marketplaces all run proprietary stacks. The economics work because the volume justifies the team and infrastructure cost, and because their data network effect, the depth of customer history they can train on, is something a third party can’t replicate at the same depth.

• Orchestration with merchant-built rules sits on top of platforms like Stripe Radar, Adyen RevenueProtect, Worldpay’s FraudSight, Cybersource Decision Manager, and Braintree. The model is the platform’s. The rule layer, the threshold, and the manual review policy are the merchant’s. The merchant retains the dial. The platform handles the underlying engine.

• Standalone third-party fraud platforms with proprietary models sit alongside or instead of the PSP-bundled tools. Sift, Kount, Accertify, Sardine, Feedzai, and ThreatMetrix all play here. The merchant integrates the platform, sees the score, builds rules around it, and makes the call.

• Decision-plus-financial-guarantee is the fourth category, and it is the one with the most consequential structural difference. A vendor reviews each transaction, approves or declines it, and reimburses the merchant if an approved transaction is subsequently charged back as fraud. Riskified, Signifyd, Forter, ClearSale, and several smaller players sell into this segment. The merchant pays a premium per transaction, typically between 50 and 150 basis points of approved order value, depending on the vertical, average order value, and contract structure.

Below the in-house volume threshold, the economics get harder. Building a fraud model that performs as well as a vendor’s, when the vendor is training on cross-merchant network data spanning hundreds of billions of dollars in annual transaction volume, is a structural disadvantage that gets steeper every year as the network effect compounds.

Most enterprise merchants run hybrids. PSP-bundled fraud tooling for the long tail of low-risk transactions, a third-party platform or guarantee vendor on top for higher-risk verticals or higher-value SKUs, internal rule overlays everywhere.

The underlying point is that the merchant has effectively transferred ownership of its most consequential customer interaction. Most have done it for cost reasons. The vendor’s network effect, the breadth of cross-merchant signal, is what the merchant is buying. What the merchant is giving up, especially in the guarantee model, is direct control of the threshold.

The Two-Sided Cost

Every fraud decision the merchant makes carries a cost on both sides. The asymmetry is not in whether each side has a cost. It’s whether the cost is visible.

The cost of approving a fraudulent transaction is documented down to the line item:

• Chargeback fee, between fifteen and one hundred dollars, depending on the processor and vertical

• Cost of goods, gone

• Shipping, gone

• Processing fee and interchange, paid and not refunded

• Transaction value itself if representment fails, and across all chargebacks, the net recovery from representment runs around twelve percent

• Exposure under VAMP, where, after April 1, 2026, a fraud-coded chargeback counts twice in the merchant’s monitoring ratio, once as a TC40 fraud report and once as a TC15 dispute, with eight dollars per event for merchants in the Excessive tier

LexisNexis Risk Solutions’ True Cost of Fraud study, fifteenth edition, published in April 2025, puts the total multiplier at $4.61 per dollar of fraud loss in the United States. Up from three dollars the prior year. That number includes all the line items above, plus the operational cost of investigation and chargeback handling, plus the cost of the fraud tools themselves.

The cost of declining a legitimate transaction is harder to see:

• Lost sale, obviously

• Lost lifetime value, less obviously

• Customer attrition, depending on the study, between roughly one quarter and just under half of falsely-declined customers don’t return

• Wasted customer acquisition cost, because the merchant paid to bring that customer to checkout and then declined them at the door

• Reputation damage, particularly when the falsely-declined customer publicly complains

The fraud team is paid on the first cost. The chargeback ratio is the team’s KPI. Fines, MATCH listing, acquirer reserves, all flow from that number. The lost-LTV cost shows up somewhere else entirely, usually in marketing dashboards, sometimes in customer service complaint logs, occasionally in an attribution report nobody reads.

Industry research shows that only about sixty-four percent of merchants even track their false-decline rate. The other thirty-six percent are operating with a one-sided cost picture by default, optimizing against a metric that captures one of the two ways they can be wrong.

This is the first place where the structure of the decisioning operation starts to matter as much as the model behind it. A team paid on chargeback ratio and not paid on false-decline rate will, rationally, set thresholds that minimize chargebacks at the expense of approval rate. The model isn’t broken. The incentive structure produces the behavior.

When The Vendor Owns The Decision

The chargeback guarantee model takes this structural problem one step further. It removes the chargeback cost entirely from the merchant’s books. It also removes something else.

The mechanic is straightforward. The vendor reviews the transaction, approves or declines, and reimburses the merchant if an approved transaction subsequently has chargebacks under a covered fraud reason code. Coverage tiers break down as follows:

• Universally covered. Visa 10.4 Other Fraud, Card-Absent Environment, and Mastercard 4837 No Cardholder Authorization

• Usually covered. Mastercard 4863 Cardholder Does Not Recognize

• Generally not covered by base fraud guarantees. Items-not-received and not-as-described disputes, although some vendors have layered products that extend coverage

• The gray zone. Friendly fraud and first-party misuse, and that gray zone is now the top-reported merchant attack category in the industry

The merchant pays a per-transaction premium for this coverage. The vendor takes the loss when the coverage triggers. From the merchant’s perspective, fraud loss certainty replaces fraud loss volatility.

Here’s where the structural shift happens. The vendor’s loss ratio is a function of the vendor’s own profit and loss. When chargeback losses exceed the premium the vendor collects, the vendor’s gross margin erodes. The vendor’s response, mechanically, is to tighten approval rates. Decline more transactions. Push the loss ratio back below the premium.

The merchant doesn’t set that dial anymore. The merchant has bought certainty and given up control.

This isn’t theoretical. Riskified, the only public pure-play in the chargeback guarantee category, said it explicitly on its Q4 2025 earnings call. Quoting the transcript: “In 2026, we are focused on driving gross profit growth versus optimizing primarily for revenue growth.” Translated, that means tighter approval rates in lower-margin categories. Riskified’s annual filings have, for years, characterized tickets and travel as higher-risk, lower-margin, and fashion and luxury as lower-risk, higher-margin. The vendor’s gross margin profile determines which categories are approved and at what rate.

The merchant in the higher-risk lower-margin category isn’t getting a worse decision because the model is worse. It’s making a worse decision because the vendor is rationally optimizing its own business. The merchant’s conversion rate is downstream of that optimization, not upstream.

This is the most consequential and least discussed structural fact about the chargeback guarantee market. Risk transfer doesn’t make the decision better. It moves who’s optimizing and what they’re optimizing for.

Why The Vertical Matters More Than The Tooling

The same fraud platform, deployed by the same merchant operations team, makes structurally different decisions across verticals because the cost-of-error profile varies. This is not a vendor-quality issue. It’s an arithmetic issue.

• Digital goods is the cleanest illustration. nSure.ai’s published, vendor-flagged data shows average decline rates of around 15% for digital gift cards and 20% for downloadable content. Of those declines, roughly nineteen percent are actual fraud. The other eighty percent of declined customers are legitimate. Two-thirds of the customer base is on the wrong side of an aggressive threshold in a category where a single fraudulent transaction generates near-instant fungible loss because the goods are a code that gets resold at a discount on a secondary market within minutes. The merchant in this vertical isn’t being stupid. The cost of being wrong on the approval side is so high that the rational threshold sits where it does.

• Travel. High average order value, longer fulfillment window, complex, multi-party supply chain, loyalty points that are themselves a target of fraud. Forter’s Fraud Attack Index, vendor-flagged, has historically shown attack rates for airlines and hotels well above retail baselines. Industry estimates put payment fraud at roughly 1.2 percent of annual airline revenue. False-decline rates for new customers have been documented at five to seven times higher than for returning customers in legacy systems.

• Subscription. Sift data, vendor-flagged, shows business-to-consumer software-as-a-service chargebacks up eighty-three percent year over year in late 2025. Industry surveys show that twenty-two percent of consumers admit to filing chargebacks for unwanted subscriptions. Forgot-I-subscribed is now a measurable business problem, and first-party fraud has risen to roughly thirty-six percent of all reported merchant fraud globally.

• Ticketing runs the most extreme profile measured publicly. Sift Q1 2025 data, vendor-flagged, puts the ticketing payment fraud attack rate at 7.4 percent, up eighty-five percent year over year, among the highest of any vertical. Peak-demand windows force automation; manual review at scale during a tour of sale isn’t possible.

• iGaming illustrates the trap of single-source measurement. One vendor’s data shows online gambling payment fraud running far below network averages, because operators’ KYC and pre-screen aggressively at deposit. A different vendor measuring identity fraud at applicant onboarding shows iGaming applicant fraud rates of five to eight and a half percent, up sixty-four percent in two years. Both true simultaneously. Different measurements at different points in the funnel.

The strategic question for the enterprise merchant is not which fraud vendor performs best in some abstract benchmark. It’s what loss profile is acceptable in this merchant’s specific vertical, given this merchant’s specific margin structure, and how the decisioning architecture reflects that. The same vendor, set to the same threshold, across two different verticals, produces materially different conversion outcomes because the underlying cost structures differ materially.

This is why fraud benchmarking across verticals is mostly noise. The right benchmark is internal: what is this merchant’s loss profile, what is its margin structure, what threshold matches both, and what did the threshold produce last quarter?

What The Enterprise Merchant Is Actually Optimizing For

Step back from the mechanics for a second.

The merchant’s decisioning operation is not optimizing for the right answer on each transaction. It’s optimizing for the cost of being wrong, weighted by which kind of wrong is visible, audited, and tied to a KPI someone is measured against.

• Visible costs. Chargebacks incur a fee, trigger a representment workflow, generate a monthly report to the acquirer, and expose the account to VAMP and the Mastercard fraud and chargeback monitoring programs. The chargeback ratio is the most heavily instrumented number in the merchant’s payment operations.

• Invisible costs. Cart abandonment due to over-aggressive fraud rules appears as a slight drag on conversion in a marketing dashboard. As a customer service ticket from someone who knew their card worked elsewhere. As a slow erosion of repeat customer rates that nobody attributes to fraud rules because the attribution data isn’t structured to make the connection.

McKinsey’s 2025 work on fraud prevention is pointed about where this lands. Quoting directly: “Many institutions run false positive rates in the high 90s, whereas best-in-class players are in the 60s.” That’s at the issuer, not the merchant, but the structural logic is the same on both sides. When the cost of one kind of error is visible, and the cost of the other isn’t, the system optimizes against the visible cost and lets the invisible one accumulate.

This is also why bringing in a better model rarely solves the problem in production. A model that produces a sharper accuracy curve hands the operations team a wider menu of policy choices. Whether the team picks a better policy point depends on whether it is rewarded for doing so.

A merchant whose fraud team is paid on the chargeback ratio, with no countervailing metric for the false-decline rate, will keep setting thresholds that minimize chargebacks, even if the model would support more aggressive approval. The model improvement translates into lower fraud loss at the same approval rate, rather than higher approval at the same fraud rate. Both are valid outcomes. They aren’t the same outcome.

The merchant that wants different outcomes from its fraud operation has to change the incentive architecture before changing the model:

• Track false-decline rate explicitly

• Build it into the fraud team’s KPI set

• Surface it to the executive layer with the same cadence as chargeback ratio

Most merchants haven’t done this. The thirty-six percent of merchants who don’t track false-decline rate at all aren’t going to.

The model isn’t producing the wrong answer. It’s producing the right answer to the wrong question.

Where This Leaves Us

The merchant runs the most informationally rich fraud check in the entire payments chain. It evaluates several hundred to a couple of thousand features per transaction, against historical data nobody else in the chain can replicate. It produces a probability, bands it, overlays rules, and sends back a decision that costs the merchant money whether it’s right or wrong.

The decisioning architecture sitting on top of that model is what produces the outcomes the merchant lives with. The threshold setting. The KPI it’s tuned against. The choice to outsource to a vendor whose loss ratio drives approval rates. The vertical-specific cost-of-error profile that determines what threshold even makes sense.

Every component is operating correctly. The model is fitting the data. The rules are doing what they were written to do. The vendor is optimizing its own profit and loss. The fraud team is hitting its chargeback ratio target. The CFO is approving the budget. None of this is broken.

What’s broken is that the question being optimized against, which minimizes the visible cost of being wrong, produces a different outcome than the question that would actually serve the merchant’s long-term economics, which is to minimize the total cost of being wrong, weighted equally on both sides. The first question has a clear owner. The second question does not.

That gap is what the merchant pays for every day: customers turned away who would have paid. The chargeback appears on the dashboard. The lost customer doesn’t.

In the next issue, the same transaction reaches the issuer’s authorization platform, where a different decisioning architecture runs against a smaller data set under a different cost structure and reaches its own conclusion about whether to honor the merchant's approval. That’s where Issue 4 starts.

Thank you for reading.

P.S. If reading this series made you wonder where your own authorization chain is leaking value, that’s exactly what I help institutions figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally, through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these deep dives, buy me a coffee.

Fraud scoring engines evaluate 500+ risk attributes in milliseconds. Authentication has SCA, 3DS 2.x, network tokenization, behavioral biometrics, and device intelligence. 99% of financial institutions are running ML somewhere in the fraud chain. The tooling is real, and it’s expensive.

And fraud is still growing. Nilson put 2024 global card fraud losses at $33.41 billion and projects the number reaching $41.06 billion by 2030.

Those two facts create a question every fraud leader I speak to is already asking themselves in private: if every layer is working harder than ever, where is the fraud actually getting through?

The answer to this issue is: it isn’t getting through. It’s arriving pre-approved.

By the time a fraudulent transaction reaches the authorization decision, the fraud has often already succeeded at an upstream layer. The layer where someone decided who they were presenting themselves as in the first place. The transaction-level stack can’t see it because it was never designed to.

That layer is the identity layer. It’s the first decision in the chain. And it’s the layer nobody owns on either side of the authorization message.

Let’s break it down.

The Decision You Make Before Every Other Decision

Everyone in this industry knows the individual components. KYC at onboarding. Log in and step up at authentication. Device, behavioral, and location signals at checkout. Network services like 3DS and tokenization in between. Merchant underwriting on the acquirer side.

Here’s the reframe worth sitting with: every one of those controls is answering the same question. Is this the party we think it is? They’re just answering it at different moments, with different data, owned by different teams, on different sides of the transaction.

The identity layer runs on both sides of the authorization message.

On the acquirer side, there are two identity decisions. The first is merchant underwriting.

• Who is this business?

• Is it what it says it is?

• Are the beneficial owners real?

• Does the MCC actually match the activity?

The second is cardholder identity at checkout.

• Device fingerprint

• Session behavior

• IP

• Billing address

• CVV match

• 3DS context

• Returning-customer signals

• Account tenure at the merchant

On the issuer side, there are also two identity decisions. The first is cardholder onboarding.

• KYC at account opening

• Identity documents

• Velocity at application

• Bureau data

The second is cardholder identity at authorization.

• The local risk score the issuer builds from account history

• Recent activity

• Whatever arrives in the authorization message

Four identity decisions per transaction on a good day. Each one is real. Each one uses data the others don’t have. And here’s the part nobody states plainly: none of them checks whether the others still hold, and almost none of them travels across the authorization message.

Let’s walk through the compression.

At the merchant, the checkout session is scored against dozens of signals.

• Device fingerprint (OS, browser, canvas rendering, TCP/TLS fingerprint)

• Session behavior (typing cadence, mouse movement, time-to-purchase)

• Network signals (IP, ASN, proxy detection, geolocation mismatch with billing address)

• Account context (tenure with this merchant, prior order history, cart behavior)

The merchant’s fraud model produces a verdict and routes the approved transactions to the acquirer.

At the acquirer, more scoring happens.

• Velocity checks across the acquirer’s merchant portfolio

• Cross-merchant risk patterns

• BIN-level behavior

• Tokenization status

The acquirer applies its own fraud model, often with a proprietary score layered on top of the merchant’s signal. Then the acquirer constructs the authorization message.

This is where the compression lands.

ISO 8583, the standard established in 1987, defines which fields can be transmitted.

• The PAN

• The amount

• The entry mode

• The EMV cryptogram, if there is one

• The 3DS CAVV and ECI if authentication happened

• The merchant name (40 characters) and MCC (4 digits)

• Network-private fields that can carry a risk score, but only if the acquirer chose to populate them and only in formats the issuer chose to parse

The device fingerprint doesn’t have a field. Session behavior doesn’t have a field. Account tenure with the merchant doesn’t have a field. None of the identity contexts the merchant spent milliseconds computing has a standardized destination in the message carrying the transaction to the issuer.

The message arrives at the network. VAA and Decision Intelligence score it against hundreds of attributes using cross-network data that the acquirer couldn’t see, and attach the score to the message. What those models see that the acquirer couldn’t: this cardholder’s behavior across every merchant, the merchant’s fraud profile across every bank, and cross-border patterns. What those models don’t see: everything the acquirer just compressed out. A different kind of intelligence, not a richer version of the same intelligence.

The issuer receives the message and makes the final decision using its own data — onboarding tier, account history, velocity, the network score, whatever rules the fraud team has accumulated. The issuer has the cardholder’s full history, but almost none of the merchant-side context. The merchant has the session but not the cardholder history. The network has both sides in aggregate, but doesn’t push the derived signals back upstream.

Three parties, four identity assessments, and a message format that lets almost none of the identity work travel.

Every upstream identity decision is a local side effect that the authorization layer can’t see, can’t verify, and can’t weigh. That’s the first seam. And it’s where the fraud has learned to live.

The Account That Was Always Fine

Let’s take one failure mode and walk through it properly.

Mule accounts. The UK PSR October 2024 reimbursement rule has put them on every fraud desk on the issuer side, and the mechanics apply far beyond APP fraud.

APP fraud is up. Authorized push payments, social engineering, and the victim moves the money themselves. Every fraud team in Europe has been dealing with this for three years.

Here’s the question nobody is asking cleanly: once the victim has sent the money, where does it go?

It goes to a real bank account. That account is the mule. And the mule account almost always passed KYC.

How does someone become a mule? Three paths, each of which the onboarding engine is blind to.

1. Path one is recruitment. A university student sees a post on Telegram or Instagram offering £300 to “help process international payments.” They open an account at a challenger bank in their own name with their own ID, then hand over the credentials. From the bank’s perspective, every signal at onboarding is legitimate. Real name, real address, real ID document, real selfie, real device. Nothing is fake.

2. Path two is dormant account acquisition. Fraud markets sell access to existing accounts opened years ago by people who have moved on. The mule operator buys credentials in bulk and routes illicit funds through the active ones. From the bank’s perspective, this looks like a dormant customer suddenly becoming active.

3. Path three is a takeover through phishing or SIM swap. A real customer’s credentials are compromised, and the account gets used as a pass-through, often for 48-72 hours before detection.

In every path, the onboarding engine had nothing to catch. The account was opened by a real person with real documents. The mule behavior, rapid inflows from unrelated parties, new payee patterns, and velocity that doesn’t match the stated account purpose only become observable after the account is live.

The UK’s Financial Conduct Authority looked at 25 UK firms in 2023. Between January 2022 and September 2023, those firms offboarded 194,084 accounts for mule activity. Every one of those accounts opened normally. The mule status emerged later. The firms found them.

Here’s what makes it a structural failure rather than a technology one.

Of those 194,084 accounts, only 37% were reported to the shared fraud database, Cifas. Reporting rates across the 25 firms ranged from 6% to 66%. And of those 25 firms, exactly one was running real-time checks against that shared database to catch mule accounts that had already been flagged at another institution.

The FCA’s own words: “Some firms conducted relatively few checks during customer onboarding, relying on post-onboarding monitoring.”

Translate that into practitioner terms. The industry has the data to catch this. Each firm is finding its own mules. The shared database exists. The connective tissue — report what you find, check what others have found — is optional, under-invested, and not enforced.

The mechanism: the fraud passes through the identity layer because the identity layer isn’t really a layer. It’s 25 separate layers running in parallel, not talking to each other. The same account that got offboarded at Firm A opens at Firm B the following week using the same recruit, sometimes with a slightly adjusted phone number and a fresh selfie.

The PSR October 2024 rule is the first real structural force changing this. Sending and receiving PSPs now split APP reimbursement 50/50 up to £85,000. For the first time, the bank that hosted the mule has a direct financial stake in catching it. In the first year of the rule: 269,000 claims reported, 88% of in-scope losses reimbursed, £173 million paid out. Compare to the 65% reimbursement under the prior voluntary code. That’s what accountability does when you build it in.

And it’s only in the UK. Nowhere else has this rule.

The Same Pattern on the Other Side

The mule account is the issuer-side identity failure. Acquirers have their own version. It’s called transaction laundering, and it works the same way.

A merchant onboards clean. Real business name, real address, registered entity, verified beneficial owners, bank account in the business name, MCC that matches the stated business. The acquirer’s underwriting checks:

• The sanctions and PEP screens

• Bureau data on the beneficial owners

• The website content

• A test transaction

• Estimated volume versus the MCC norm

All pass.

Post-approval, the monitoring regime changes character. The acquirer’s systems are looking for volume anomalies, chargeback spikes, and MATCH listings. They’re not re-inspecting the underlying business every month. A merchant approved to sell home goods continues to be treated as a home goods merchant.

Now the drift starts.

• Sometimes the merchant knowingly fronts for another business that couldn’t pass underwriting — adult content, gambling, supplements, cryptocurrency, counterfeit goods.

• Sometimes the merchant sells access to their MPA on a gray market for a percentage of processed volume.

• Sometimes an acquired customer list contains a company whose real business is different from what it said.

The cards being processed don’t change, but the actual nature of the commerce does. The MCC is wrong. The descriptor is wrong. The risk profile approved by the underwriter is wrong.

The detection gap is structural, just like the mule problem. A merchant offboarded for content violation at Acquirer A routinely reappears at Acquirer B within weeks, with a slightly different business name, a fresh set of documents, and sometimes a different beneficial owner on paper. Mastercard BRAM and Visa’s Global Brand Protection Program exist precisely because acquirers don’t coordinate. The networks send violation notifications and impose fines, but those travel merchant-by-merchant rather than as a cross-acquirer intelligence feed. What one acquirer just offboarded for, another acquirer is about to approve.

Same mechanism as the mule. Clean onboarding, non-clean behavior, and no real-time coordination between institutions on what they’ve already caught.

Two sides of the authorization message. Two versions of the same failure. Neither side has a real-time cross-institutional view of what the other side has already seen.

Same Shape, Different Faces

The mule pattern isn’t unique. It’s the shape of every identity-layer failure. Three more, at depth.

Synthetic identity, on the issuer side. A fabricated identity is assembled from fragments of real data. The most common construction: a real Social Security Number belonging to a child (often harvested from a data breach of pediatric health records), paired with an invented name and invented date of birth. Sometimes the SSN belongs to a deceased person, sometimes to an elderly person whose credit isn’t being actively monitored. The name and DOB are fiction. The SSN is real.

Here’s the mechanism that makes this work. Under the Fair Credit Reporting Act, a credit bureau is required to open a file on any credit inquiry. The act of applying for credit with a new SSN-name-DOB combination creates a file for that combination. The first few applications are rejected, but the file now exists. Future applicants with the same combination appear to have a credit history.

The Federal Reserve stated this exactly: “The existence of a credit file can seem to imply ‘proof of life’ and validation of the identity, but it does not.”

Now the aging cycle. The synthetic operator secures one line of credit — a subprime card, a buy-now-pay-later account, or an authorized-user add to an existing account. Months of on-time payments follow. The credit score rises. More accounts open. Limits increase. The file appears increasingly legitimate as it accumulates legitimate-looking behavior. After 12 to 24 months, the synthetic has several accounts with a combined line of $50,000 to $100,000.

Then the bust-out. Every line is drawn down simultaneously. The synthetic disappears. The losses appear across multiple issuers, each of whom books the loss as a credit charge-off — not as fraud. The identity that caused all of it is never labeled as synthetic in any institution’s ledger.

This is why the FinCEN synthetic-identity SAR count for 2021 was roughly 3,000 filings totaling $182 million, while industry consensus estimates actual synthetic-identity losses at $20 to $30 billion annually. Two orders of magnitude between what the federal government counts and what the industry thinks is happening. The fraud detection stack never saw it as fraud. The institution booked it as a customer who stopped paying.

Merchant-side account takeover, on the acquirer side. A fraudster obtains credentials through credential stuffing, running leaked username-and-password combinations from unrelated data breaches against major marketplaces. The hit rate is low, but the supply is effectively unlimited. The Identity Theft Resource Center recorded 3,158 US data compromises in 2024, with roughly 1.367 billion victim notices. Credentials are not a scarce resource.

Once the fraudster is in the account, the device is new. The marketplace’s fraud model notices. This is where the tiered trust pattern kicks in. Most marketplaces use tiered trust, a new device requires verification on first access, typically a text message to a phone number on file. If the credentials came with the phone number (SIM swap, or number porting) or if the account recovery email was also compromised, the verification passes. After the first successful verification, the device moves into the trusted tier. From that point, the marketplace’s fraud model sees the account as a returning customer on a recognized device.

The fraudster ships goods to a drop address or immediately sells digital goods. The authorization message that hits the issuer looks clean. The cardholder recognizes the merchant. The device is known to the merchant. The acquirer sees the marketplace’s “low risk” verdict and forwards the message. The issuer approves.

The chargeback arrives 30 days later, when the legitimate cardholder sees the charge. The issuer’s fraud team codes it as CNP fraud at the merchant. The acquirer sees it as a dispute. The merchant absorbs the loss. Nobody in the chain has enough visibility to see that the actual failure was at the marketplace’s identity trust model, which the issuer can’t see and the acquirer treats as a black box.

The full chain is failing upstream. The 2023-2026 BaaS enforcement wave is telling this story in institutional form. Evolve Bank & Trust, Blue Ridge, Piermont, Cross River, Sutton, Lineage, Thread, Hatch. Every federal consent order in the set names the same failure. The sponsor bank held the charter and the BSA obligation. The fintech or middleware provider held the customer relationship, the onboarding decision, and the identity telemetry. Neither was accountable for the seam between them.

Here’s what that means in practice. A fintech runs onboarding through its own stack, identity document capture, liveness check, device attestation, and bureau match. The sponsor bank has contractual access to the results, but not always to the raw telemetry. When the fintech approves a user, it sends an account-opening event to the sponsor bank. The bank’s BSA obligations include verifying that the fintech has run the required checks, but the bank’s visibility is often limited to the output of those checks rather than the underlying signal quality.

Piermont’s CEO, Wendy Cai-Lee, said it directly to American Banker: “Every bank that touches BaaS is getting an enforcement action.” The reason is structural. When the fintech is optimized for conversion, and the sponsor bank is liable for AML, the incentive to accept a marginal identity signal sits with the fintech, and the consequence sits with the bank. Synapse’s 2024 bankruptcy made the opposite version of the same seam visible: when the middleware holds the ledger and the sponsor bank holds the deposits, the ~$65-96 million gap between the two can persist for years because neither side can see the whole.

Now follow that downstream to the acquirer. When one of those fintechs onboards a fraudulent user and that user transacts with a merchant, the acquirer is processing against a card issued by a sponsor bank that authorized the card on the basis of identity work the bank itself didn’t directly perform. The acquirer can’t see any of that. The merchant certainly can’t. The loss shows up on the merchant’s chargeback ledger, coded as CNP fraud at the transaction. What actually happened is that the identity decision at fintech onboarding was wrong, and the chain of accountability never surfaced the error to the side of the authorization message that ultimately absorbed the cost.

Three different fraud vectors. Same structural mechanism. The fraud succeeds because the identity layer isn’t reconciling its own decisions, and because those decisions were made by parties who don’t share their conclusions across the authorization message.

You Can’t Fix What Nobody Counts

In the last issue I made a case that false declines go unmeasured because no one with the data has the incentive to report, no one with the incentive has the data, and the networks sell the visibility rather than mandate it.

The identity layer has the same shape. Three times over.

• Synthetic identity losses. The gap between FinCEN’s $182 million SAR count and the $20–30 billion industry consensus is two orders of magnitude because synthetic losses get booked as credit charge-offs, not fraud. The federal government counts the ones flagged as identity fraud at the point of SAR filing. Banks book the ones that bust out at the point of unpaid balance. Nobody counts the same thing. Even the Fed admits the number in their own publications is sourced from a vendor rather than federal data.

• Mule-enabled and transaction-laundering losses. I don't believe any regulator publishes a global number for either. Europol counts arrested mules (10,759 in 2023 operations). UK Finance counts APP reimbursements (£450.7M in 2024). Cifas counts database filings (34,000 mule flags in 2024, with the drop attributed to regulator caution about criminalizing under-30s, not a real decline). FinCEN counts SARs. On the acquirer side, BRAM and GBPP outcomes are opaque to anyone outside the networks. Each program measures a slice. None of them aggregates.

• Cross-side identity coherence. There is no metric anywhere that assesses whether the acquirer’s and the issuer’s identities align on a given transaction. No disclosure regime would ever produce one. The question isn’t being asked because no one owns the answer, and because the networks — the only parties that could see both sides — have chosen to treat cross-side coherence as a product surface rather than a protocol obligation.

This is why the Nilson trajectory looks the way it does.

The fraud isn’t growing because the risk stack is getting weaker. It’s getting stronger every year. The fraud is growing because it keeps shifting to parts of the stack that nobody measures. And nobody measures it because measuring it would require both sides of the authorization message to agree that the question matters.

Two Product Lines, One Supplier, Zero Coherence

In the last issue, the growth of value-added services at Visa and Mastercard was the clearest single signal that the networks prefer to sell the solution rather than build it into the protocol. VAS is growing 22 to 28% annually, while core payment-network revenue grows at less than half that rate.

Same pattern in identity, and the picture gets more interesting when you put both product lines side by side.

Visa FY2025: VAS revenue up 25% in constant currency. Mastercard FY2025: VAS up 23% nominal. Both numbers are from the 10-Ks, not the marketing materials. Most of that growth is in identity and fraud-adjacent categories.

On the issuer side, the product stack is identity-at-authorization. Visa Advanced Authorization scores transactions against cross-network behavioral data and returns a risk score in the authorization message. Visa Deep Authorization (launched March 2024) is a recurrent neural network trained on CNP sequences. Visa Risk Manager lets issuers build custom rules using network scores as inputs. Mastercard Decision Intelligence performs equivalent work on Mastercard rails and is now positioned as network-agnostic scoring for any transaction. Both networks market these to issuers as authorization-rate-lift tools.

On the acquirer and merchant side, the product stack is identity-at-credential and identity-at-checkout. Network tokenization replaces the PAN with a merchant-specific token that carries lifecycle data, including the provision date, reissuance history, and merchant context. Visa reports 13.7 billion tokens issued by mid-2025, roughly half of all global Visa e-commerce tokenized. Mastercard reports ~35% of total transactions are tokenized. Both networks sell tokenization at a 3-6% authorization-rate lift. Mastercard’s Payment Optimization Platform (launched in October 2025) sits between the acquirer and the issuer, enriching the authorization message in real time with cross-network data.

Here’s the structural point. Both product lines are built on the same underlying telemetry. The network sees both sides of every transaction, the merchant’s presentation, the issuer’s response, and the cardholder’s behavior across the ecosystem. That telemetry is the raw material.

What each side buys is a derived product. Issuers buy a risk score built from that telemetry and delivered in the authorization message. Acquirers and merchants buy a token lifecycle signal and an authorization optimization service. Both sides pay. Both sides get real value.

Neither side gets the other side’s derived signal. And neither side gets the raw telemetry on which both sides’ products are built.

The thing nobody is selling, and nobody is buying, is the coherence product.

There is no product that tells the issuer, “This acquirer’s fraud model scored this session as low-risk with these specific signals.” There is no product that tells the acquirer “this issuer’s cardholder passed SCA with a phishing-resistant authenticator six minutes ago on an enrolled device.” Those signals exist. The network sees them. They aren’t packaged.

The networks are the only party with the visibility to build cross-side identity coherence. They have no commercial reason to do so when they can sell two products to two buyers built on the same underlying data. This is not a criticism of the business model. It’s an observation about where the incentive architecture has settled.

Where This Leaves Us

Four identity decisions per transaction on the good days. None of them checks whether the others still hold. None of them travels with the message in any standardized way. None of them gets measured for whether it was right.

That’s not a technology problem. That’s an ownership problem.

The identity layer fails silently because silence is what happens when nobody is accountable for the first decision in the chain, and when nobody on either side of the authorization message can see the other side’s first decision. The fraud has learned to operate inside that silence.

The fraud isn’t getting through the risk stack. It’s passing between the layers that each side of the stack is built on.

That’s where Issue 2 ends.

Issue 3 takes the merchant layer apart. The richest identity context in the entire chain sits at checkout. Almost none of it reaches the authorization message. Next week, I’ll walk through what gets lost at the handoff, what survives, and why the signal quality gap between what the merchant knows and what the issuer sees is one of the most under-discussed structural problems in payments.

Thank you for reading.

This edition was made in partnership with Incognia, which provided its expertise, insights, and reports to help me better understand the challenges with Identity.

P.S. If reading this series made you wonder where your own authorization chain is leaking value, that’s exactly what I help institutions figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally, through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or, if you just want to keep fueling these deep dives, buy me a coffee.

That asymmetry is at the heart of everything that follows in this series. Not as a moral critique. As a structural observation with measurable consequences.

The Risk Stack, ranging from identity verification, fraud scoring, authentication, and credential security, has never been better funded or more technically sophisticated. However, Fraud losses have never been higher. Both of those things are true simultaneously, and the explanation isn’t that the technology is failing. It’s that each layer was built to solve its own problem, measured against its own metrics, and incentivized by its own penalties. Nobody designed them to work together. The seams between them are where the fraud gets through. They’re also where legitimate customers are turned away.

In this series, I will pull that apart, layer by layer. It starts with the number that frames everything else.

Let’s break it down.

The Numbers Nobody Wants to Compare

Let’s start with fraud, because it’s the most defensible figure in this space.

The Nilson Report is the closest thing this industry has to an authoritative source on card fraud. They survey more than 2,000 financial institutions globally, take no advertising, carry no vendor relationships, and have been publishing since 1970. Their January 2026 report put global card fraud losses at $33.41 billion in 2024. A trend that doesn’t seem to be flattening, as Nilson projects it reaching $41.06 billion by 2030.

Now set that against the other number.

In September 2023, Checkout.com commissioned Oxford Economics to measure false-decline losses across the US, the UK, France, and Germany. Four markets only, covering 2022 data, methodology built on a survey of 1,500+ enterprise merchants with willingness-to-pay extrapolation. Vendor-commissioned by Checkout.com, which sells payment optimization tools, the report is commercially driven. But the Oxford Economics methodology is sound, and the figure has held up against independent scrutiny. The result: $50.7 billion in direct merchant revenue lost to false declines, across four markets, in a single year.

The comparison is simple. Even narrowly scoped, even with the provenance on the table, the industry refuses more legitimate revenue than it loses to fraud.

Another big number that has been floating around for a while is the $443 billion figure originating from a Aite Group research commissioned by ClearSale, a fraud prevention vendor, which measures gross global CNP transaction value falsely declined, a fundamentally different metric that bundles in lifetime value estimates, competitive transfer costs, and future revenue projections. It’s not wrong, it just measures something different. The $50.7 billion is the cleaner, more conservative, more directly comparable number. Even if one wins the comparison.

What makes this more than an interesting data point is what sits underneath it.

No one can tell you precisely how large the false decline problem is, because the industry has never agreed on what a false decline is.

The Merchant Risk Council reports false-positive rates as a percentage of total ecommerce orders. Aite-Novarica surveys merchants for self-reported estimates. Some merchants proxy it through second-attempt approvals. The definitions don’t align, the methodologies don’t align, and the incentives to measure accurately are almost entirely absent. That definitional chaos isn’t incidental. It’s the mechanism. We’ll come back to it.

What the Incentive Architecture Actually Rewards

The Risk Stack doesn’t produce this outcome because it’s broken. It produces it because it’s working exactly as designed.

I’ve walked issuer risk teams through this arithmetic more times than I can count. The response is almost always the same sentence: “We know.” The incentive structure is not invisible to the people inside it. What’s invisible is the cost of the false declines they’re generating, because that cost never lands anywhere they can see it.

Here’s why.

For approving fraud, the consequences are severe, visible, and directly attributable:

• Every dollar of fraud costs US financial institutions $4.41 in total losses once investigation, recovery, and legal costs are included, per LexisNexis’s 2024 True Cost of Fraud Study

• Each disputed transaction costs issuers between $9.08 and $10.32 to process, per Mastercard’s 2025 data

• Regulatory penalties for fraud-related compliance failures surged 522% in 2024, per Alloy’s 2025 State of Fraud Report

• The worst-case outcome a MATCH listing, effectively ends a merchant’s ability to process cards

These numbers appear on dashboards. They trigger risk reviews. They end careers.

For refusing a legitimate customer, the consequences are: nothing.

• No regulatory penalties.

• No network monitoring programs.

• No standardized reporting requirements.

The cost of a false decline lands entirely on the merchant’s income statement; lost revenue, lost customer, absorbed quietly. It never appears on the issuer’s balance sheet. It never shows up in the weekly risk review.

The network monitoring architecture makes this asymmetry structural rather than incidental. Visa’s VIMP, the Visa Issuer Monitoring Program, monitors issuers for excessive fraud approval and penalizes them when fraud ratios breach certain thresholds. It pushes issuers to decline more. There is no program that pushes them to decline less. Both feet on the same pedal.

The April 2026 VAMP change further sharpened this.

Visa’s Acquirer Monitoring Program (VAMP) went live on April 1, 2025, consolidating 38 prior remediation processes into a single framework. On April 1, 2026, now in force, Visa tightened the merchant Excessive threshold from 220 basis points to 150 basis points, a 32% reduction in fraud tolerance across the US, Canada, the EU, and the Asia Pacific. The formula counts TC40 fraud reports plus all TC15 disputes, divided by settled transactions; a single fraud-related chargeback counts twice in the numerator. Penalties reach $8 per transaction in the Excessive tier, with MATCH listing as the terminal outcome.

Acquirers don’t wait for Visa’s threshold. They run their own internal buffers, typically 30 to 40 basis points below the network limit, giving themselves room to recover. Tighter network threshold, tighter acquirer buffer, tighter merchant pressure.

The April 2026 change widened the false decline incentive gap in five directions simultaneously:

1. Tighter fraud tolerance forces merchants and acquirers to reject more marginal transactions to stay safely clear of the threshold

2. TC40+TC15 double-counting means every fraud incident that becomes a chargeback costs twice in ratio terms

3. Acquirer buffers cascade tighter below Visa’s threshold, compounding the effect before any formal penalty is triggered

4. The cost of an approved fraud just increased; the cost of a false decline remained exactly zero

5. No countervailing mechanism exists at any network, not Mastercard, not Amex, not Discover

Visa and Mastercard do offer tools that help with the false decline side: Visa Advanced Authorization evaluates 500+ risk attributes per transaction, Mastercard Decision Intelligence scores transactions in real time, and network tokenization consistently delivers 3–6% improvements in the authorization rate for participating merchants. Both networks have been growing value-added services revenue at 22–28% annually, from roughly $5 billion each in fiscal 2021 to $10.9 billion (Visa) and $13.3 billion (Mastercard), respectively, by fiscal 2025, per their 10-K filings.

These products work. But they’re sold, not mandated. VAMP is a compliance obligation with financial penalties. Visa Advanced Authorization is a subscription. Networks monetize false decline reduction as a profit center while mandating fraud reduction as a compliance obligation. That’s not a criticism of the network business model. It’s a structural observation that explains where the incentive to fix this actually lies.

The Spiral the Dashboard Doesn’t Show

The chargeback feedback loop is where the problem compounds silently.

When a cardholder disputes a transaction, whether it’s genuine fraud or a purchase they regret and misrepresent, that dispute gets coded as fraud in the issuer’s training data. Ethoca, a Mastercard subsidiary, has documented this mechanism directly.

Friendly fraud, chargebacks filed on transactions the cardholder actually authorized, is now the majority of disputes in digital goods categories and a significant share in most others. When friendly fraud is coded as genuine fraud, it drives model tightening. Tighter models decline more legitimate transactions. Some of those customers leave. The fraud-to-legitimate ratio in the training data worsens. The model tightens again.

Fraud losses look fine throughout. Chargeback ratio: clean. Regulatory posture: solid. The spiral is invisible in every metric that gets reported in the weekly risk review.

The damage accumulates in cardholder engagement and merchant relationships, in numbers that nobody has built the reporting to track.

McKinsey’s 2025 Guardrails for Growth report quantifies the gap between good and average. Average institutions run false positive rates in the high 90s. Best-in-class run the same metric in the 60s. That’s a gap of more than 30 percentage points, achieved not through different tools but through different measurement and organizational accountability.

The institutions performing in the 60s have made one structural choice that the majority haven’t: they put false decline rate on the risk team’s scorecard alongside fraud loss rate. They retrain models on downstream outcomes, not just the fraud flag at transaction time, but the chargeback result 60 days later, the retry outcome, the cardholder attrition over the following quarter. The model learns from the full consequences of its decisions, not just the moment they were made.

PULSE Network’s 2024 Debit Issuer Study puts a harder number on the organizational failure: only 35% of issuers simultaneously achieve both high authorization rates and low fraud losses. The trade-off most institutions accept as inevitable, that approving more means more fraud, is a fiction. Those 35% have already disproved it. The other 65% are optimizing against a constraint that doesn’t exist because the reporting that would show them otherwise was never built.

The false positive spiral is insidious because it feels like the system is working. Fraud losses are contained. The dashboard is green. The merchants churning quietly in the background are someone else’s problem. But those merchants and the cardholders they’re losing aren’t abstract. They’re the revenue base the entire ecosystem depends on.

Why the Fraud Problem Keeps Winning

Here’s what most fraud prevention analysis doesn’t reach.

The Risk Stack was built to detect fraud at the point of transaction. That’s where the investment is, where the monitoring programs sit, where the commercial products operate. It’s sophisticated, well-funded, and increasingly effective at doing exactly what it was designed to do.

The problem is that the fraud problem has moved upstream.

By the time a fraudulent transaction reaches the authorization decision, the fraud has often already succeeded at a layer the transaction-level stack was never designed to see.

Mule accounts pass identity checks because they use real identities. The account was opened legitimately, KYC was passed, and the profile is clean. The mule was created after onboarding, either by recruiting the account holder or taking over a dormant account. By the time a mule transaction arrives at the authorization engine, the identity layer has already cleared it. The fraud wasn’t at checkout. It was at the account opening.

Synthetic identities survive credit checks because the data is internally consistent. A fabricated identity assembled from real data fragments can pass the kind of verification the transaction-level stack performs, because the stack checks transactional signals rather than the coherence of the underlying identity. According to the Federal Reserve’s research on synthetic identity fraud, it’s the fastest-growing type of financial crime in the US. The fraud happened at onboarding, months or years before the transaction that triggers the chargeback.

Account takeover fraud clears fraud scoring because the behavioral patterns match those of the real cardholder. The fraudster has the credentials, the device may be recognized, and the transaction history matches. The signal the authorization engine looks for, behavioral consistency, is exactly what a sophisticated account takeover replicates. There’s no way to distinguish the cardholder from the fraudster at the transaction layer, because the impersonation already happened at the credential layer.

Compromised tokenized credentials produce clean authorization signals because the token infrastructure doesn’t distinguish theft from legitimate use. The token is valid. Merchant relationship checks out. Transaction flows. The chargeback arrives 60 days later and shows up in Nilson’s next survey as part of the accelerating trend.

The Nilson trajectory of $33.41 billion in 2024, heading toward $41.06 billion by 2030, isn’t evidence that the Risk Stack is failing at its own logic. It’s evidence that the problem has moved to a higher layer. By the time the authorization request arrives, the trust decisions that should have stopped the fraud have already been made or silently failed at identity, onboarding, or credential.

This is the structural problem the industry hasn’t fully named. The Risk Stack optimizes at each layer and produces a worse aggregate outcome than the sum of its parts. Each layer makes individually rational decisions within its own data, incentives, and mandate. Nobody is accountable for the seams. And the seams are exactly where the fraud is living.

The Measurement Vacuum That Keeps It All in Place

Both problems, the false decline problem and the fraud growth problem, share a common enabling condition: neither is measured in a way that creates institutional accountability.

Fraud gets close. Nilson’s annual survey, VAMP, VIMP, and regulatory reporting requirements, there’s at least a framework for measuring and penalizing fraud approval. Imperfect and gameable, but it exists.

False declines have nothing comparable. The Merchant Risk Council measures false positive rates as a percentage of e-commerce orders. Aite-Novarica surveys merchants for self-reported estimates. Some merchants proxy it through second-attempt approvals. The definitions don’t align. The methodologies don’t align. And the incentives to measure accurately are almost entirely absent.

The reason is structural, not incidental:

Issuers have the data but not the incentive. They know when a cardholder calls to dispute a decline. That signal exists in their systems. But because there’s no penalty for false declines, there’s no organizational pressure to act on it, aggregate it, or report it systematically.

Merchants have the incentive but not the data. They absorb the revenue loss and the customer attrition. But the response code they receive when a payment fails is typically “Do Not Honor”, three words that convey nothing about whether the decline was legitimate, why it happened, or what could be done differently. PULSE Network’s own guidelines suggest that generic declines, such as Code 05, should represent no more than 5% of all denied transactions. In practice, they account for 30–50% across the industry. The code exists precisely because specificity creates accountability. And accountability is uncomfortable when the decision might be wrong.

Networks have the visibility but have chosen to sell it rather than mandate it. They can see the patterns that no individual institution can. That visibility is packaged into VAA, Decision Intelligence, and authorization optimization products, growing 22–28% annually, not built into compliance obligations.

The measurement vacuum is the mechanism. You can’t penalize what you can’t measure. You can’t measure what nobody with data has an incentive to report. And no one with data has an incentive to report, because the penalty architecture runs in only one direction.

What This Series Is About

The false decline problem and the fraud growth problem have the same root cause: the layers of the Risk Stack were each built to solve their own problem. None of them were designed to work together.

The merchant knows the customer, device, session, behavioral history, and account tenure. That knowledge doesn’t survive the handoff to the acquirer intact. The acquirer’s fraud model concludes that a transaction is low risk. That conclusion doesn’t reach the issuer. The issuer’s authentication layer tests an identity claim without knowing what the merchant knew about the person behind it. The credential layer validates a token without knowing whether that token was compromised three months ago at the identity layer. Each layer operates on a different slice of reality. None of them can see what the others saw.

The seams between them are where the fraud gets through. They’re also where legitimate customers are turned away.

Fixing either problem at the transaction layer, where the tooling, monitoring programs, and commercial investment are concentrated, is solving the symptom. The problem starts earlier. It starts at the identity layer, where trust is either established on solid ground or silently assumed based on inputs that have already been compromised.

Over the next four issues, we pull this apart layer by layer:

• Issue 2 — The Identity Layer: Device intelligence, behavioral biometrics, location signals, and why mule accounts are the clearest current evidence that the foundation of the Risk Stack is failing silently

• Issue 3 — The Merchant Layer: What signals merchants generate or suppress before a transaction ever reaches the issuer, and why the richest context in the chain almost never travels downstream

• Issue 4 — The Authentication Layer: ACS quality, 3DS exemption strategy, and why the gap between authentication deployed as a trust tool versus a compliance checkbox is measured in hundreds of millions in fraud losses

• Issue 5 — The Credential Layer: Dynamic CVVs, network tokenization, and what credential security means when the entities presenting credentials are increasingly not human

The Risk Stack isn’t failing because the technology is inadequate. The fraud numbers keep on growing because the layers of the Risk Stack were built, measured, and incentivized in isolation. The investment keeps going in. The fraud keeps going up. And the legitimate customers keep getting turned away.

You can’t fix a trust problem by optimizing a fraud score.

That’s where this series starts.

Thank you for reading.

If you’re working on risk infrastructure at an issuer, acquirer, or network, or trying to understand why your fraud losses and your false decline rates are moving in the same direction at the same time, this is the series for it.

If this resonated, forward it to one person who should be reading it. The conversation this series aims to start benefits from having more practitioners in the room.

P.S. On 23 April, I'm speaking at The Fraud Forum in London, UK, a small, off-the-record afternoon with 25 senior fraud and risk leaders, hosted by SafeCypher. No panels, no product demos, just an honest conversation about where the industry is heading. Attendance is free, but space is limited. Register your interest here.

What happens next takes less than two seconds. Four institutions will touch this transaction. Each one will make a decision, or influence one, based on a different slice of reality. None of them will see what the others saw.

Over the past four issues, we’ve dissected what each layer of the authorization chain does independently. The issuer’s paradox. The acquirer’s divergence. The network’s advantage. Separately, each one made sense. Together, they reveal something that no single-layer analysis can show.

This series started with a number. The basis-point gap between best-in-class and average authorization performance. Four issues later, the source of that gap is visible. It isn’t in any single layer. It’s in the seams between them.

The transaction is still waiting. Let’s follow it through.

What the Acquirer Knows (And What It Can’t Say)

The PSP’s system picks up the transaction and immediately starts building context.

• Device fingerprint: recognised, same browser, same operating system, consistent with her previous sessions.

• IP address: residential, matches her billing region.

• Account age: fourteen months, three prior purchases, no disputes.

• Cart: single item, women’s outerwear, consistent with past category behaviour.

The PSP’s ML model, trained across billions of platform transactions, scores it. Low risk. High likelihood of approval.

The acquirer selects the optimal route. Local acquiring licence in the issuer’s market. Visa rails. Network tokenized credential. Every decision so far has been informed by the richest data in the entire chain.

Then the acquirer builds the authorization message.

This is where the compression begins. The message format that carries card transactions, ISO 8583, was built in 1987. Major processor implementation guides, updated as recently as March 2025, still reflect interpretations of that original standard. The updated version, ISO 8583:2003, hasn’t achieved wide adoption. Nearly four decades later, the industry’s most critical message format remains fundamentally unchanged.

In that message, the merchant’s identity travels as a four-digit MCC code, one of roughly 800 categories for the entire global economy, and a card acceptor name crammed into 40 fixed characters.

• The device fingerprint doesn’t travel.

• The session behaviour doesn’t travel.

• The IP address, the account age, the cart contents, the customer’s purchase history, none of it has a standardised field in the authorization message.

The acquirer populates certain fields because interchange qualification requires it, not because it helps the issuer make a better decision.

I’ve spent years working across these handoffs, and this is the part that still strikes me. The acquirer’s model concluded this was a low-risk, high-value returning customer with a clean device and a consistent purchase pattern. That conclusion doesn’t travel with the message. The evidence that supports it doesn’t either. The acquirer did its job. The message just can’t carry the proof.

This is the acquirer divergence in mechanical terms. Modern PSPs like Adyen and Stripe invest heavily in pre-authorization intelligence. Adyen’s Uplift, Stripe’s Payments Foundation Model, Worldpay’s Revenue Boost. But the structural constraint is that the output of that intelligence gets compressed at the handoff. The acquirer shapes the signal. The signal then narrows.

PSD3 and the Payment Services Regulation, now working through European legislative process, explicitly call for businesses to share more environmental and behavioural data, location, device, session history, spending patterns, to support issuer decisions. The fact that regulators are mandating richer data sharing tells you how wide the current gap is.

The message leaves the acquirer. It’s already a fraction of what the acquirer knew.

What the Network Adds (And What It Can’t See)

The message arrives at VisaNet. Within a millisecond, Visa Advanced Authorization scores it against more than 500 risk attributes. VAA has been in production since 1993. It’s deployed at over 8,000 financial institutions across 129 countries. It produces a numeric risk score, on a scale of 0 to 99, and attaches it to the authorization message before it reaches the issuer.

What VAA sees that the acquirer can’t: this cardholder’s spending patterns across every merchant on the Visa network. The issuer’s historical approval tendencies. The merchant’s fraud profile across every bank. Cross-border patterns that no single acquirer has visibility into. That’s the structural advantage of sitting at the centre of 257.5 billion annual transactions.

What VAA doesn’t see is everything the acquirer just compressed out of the message. The device context. The session behaviour. The customer’s account tenure with this specific merchant. VAA fills that gap with its own cross-network behavioural models, which are enormously powerful at scale. But it’s a different kind of intelligence, not a richer version of the same intelligence. The network adds signal. It doesn’t recover the signal that was lost at the previous handoff.

Nobody outside Visa knows precisely how much of VAA’s scoring power comes from the acquirer’s message fields versus Visa’s own derived models. Feature attribution is proprietary. What public sources tell us is that the score is delivered through the authorization stream, the issuer receives it alongside the original message, and it’s designed to support approval lift, particularly for card-not-present transactions.

Mastercard’s Decision Intelligence (DI) does the same thing on its rails, scoring transactions in real time and delivering the score through the authorization message. Both networks now offer their scoring as a network-agnostic product: Visa expanded VAA to non-Visa card payments in March 2024, and Mastercard positions Decision Intelligence as scoring transactions processed by any network. They’re competing to be the intelligence layer for all card payments, not just their own.

On Mastercard rails, the network can go further. The Payment Optimization Platform, launched in October 2025, enriches the authorization message itself, evaluating over a trillion combinations of data elements to construct the optimal request before the issuer sees it. On Visa rails, Real-Time Account Updater fixes stale credentials inline during the authorization process. In both cases, the network is intervening between the acquirer’s message and the issuer’s decision.

Our transaction is on Visa rails. VAA has scored it. The score, along with the acquirer’s (compressed) message, moves to the issuer.

What the Issuer Decides (And What It Never Saw)

The message arrives at the issuing bank. It carries the original ISO 8583 data, a network token, and VAA’s risk score.

The issuer has something neither the acquirer nor the network has: this cardholder’s full account history. Current balance. Available credit. Spending limits. Prior fraud incidents. Payment behaviour over years. That’s valuable context that no other party in the chain can replicate.

But the issuer doesn’t have the acquirer’s device-level scoring. It doesn’t have the session context. And whether it meaningfully integrates the network’s score depends entirely on its internal architecture. No credible public statistic exists on what percentage of issuers actively gate decisions on VAA or DI scores versus treating them as advisory inputs versus ignoring them. The public material is overwhelmingly “issuers can.” Not “issuers do.”

What I’ve seen across dozens of issuer implementations over the years is a system that layers ML models on top of rules accumulated through every fraud cycle the institution has lived through. A velocity rule from a 2019 card-testing attack. A geographic block from a 2021 breach response. Manual overlays nobody has audited. Rules added for defensible reasons. Rules that are almost never removed.

Our €197 transaction. The acquirer scored it as low-risk. VAA scored it as low-risk. The issuer’s fraud system flags it anyway. Three transactions from the same IP address this evening. Two of them were from a different cardholder, her partner, using a different card on the same home WiFi. The acquirer knew they were different sessions with different device fingerprints. That data didn’t survive the handoff. The issuer sees one IP, three transactions, and applies a velocity rule.

Declined. Response code 05. Do Not Honor.

Close to half of all decline responses across the industry use that code. Three words. No explanation. No feedback loop. The merchant doesn’t know why. The customer doesn’t know why. The issuer’s dashboard records a successful risk intervention.

As we covered in The Issuer Paradox, the penalty structure makes this entirely rational. Visa monitors issuers whose fraud ratios breach certain thresholds. Fines up to $100,000 for non-compliance. No equivalent programme penalises issuers for high false decline rates. Not one. The cost of this decline lands on the merchant’s income statement. The issuer’s dashboard stays green.

Best-in-class issuers run false positive rates in the 60s. Average institutions run them in the high 90s. A gap of more than 30 percentage points, per McKinsey’s 2025 analysis. That gap doesn’t come from access to different tools. It comes from how the same tools are integrated, what feedback loops exist, and what the organisation is actually measured on.

The Basis Points Are in the Seams

Three layers of intelligence. Each one did its job.

The acquirer built a clean message, scored the transaction correctly, selected the optimal route. The network added cross-network intelligence the acquirer couldn’t generate, attached a risk score that flagged no concerns. The issuer applied its portfolio-level risk rules using the cardholder context only it has access to. Every institution behaved rationally within its own data, its own incentives, its own KPIs.

The transaction was declined anyway. The customer doesn’t retry. She closes the tab, finds the same coat on another retailer’s site, and pays with a different card. The merchant absorbs a €197 loss it will never know about. The issuer records a clean fraud metric. The chain worked exactly as designed. That’s the problem.

If you want to understand where the authorization rate gap lives, this is it. Not in any single layer’s underperformance. In the seams between layers where data is lost, intelligence is compressed, and incentives don’t align.

Three structural forces keep those seams open.

The data compression problem. The richest context exists at the edges. The merchant side holds device, session, and behavioural data. The issuer side holds cardholder account history and balance. The authorization message that connects them is a standard from 1987 that compresses merchant identity into four digits and 40 characters. The network fills the gap with cross-network intelligence, but that’s not the same data. It’s a different lens on the same transaction. ISO 20022 won’t solve this for card authorization any time soon. Visa’s own materials note the “limited commercial case for adoption for payment systems already operating on ISO 8583.” The handoffs don’t add up to a complete picture anywhere in the chain.

The incentive misalignment.

The acquirer is paid on volume processed. The network is paid on switched transactions. The issuer is penalised for fraud but not for false declines.

Only the network’s economics are fully aligned with approval. That’s not a criticism. It’s how the model works. And it explains why the network is investing the most in making approval happen.

Which brings us to the money.

Visa’s value-added services revenue grew from approximately $5 billion in fiscal 2021 to $10.9 billion in fiscal 2025. From roughly 21% to 27% of net revenue.

Mastercard’s value-added services and solutions grew from $5.4 billion in fiscal 2020 to $13.3 billion in fiscal 2025. From 35% to 41% of total net revenue.

Both figures are from SEC filings, audited. Both are growing materially faster than core payment processing revenue.

That money comes from selling intelligence back to the institutions on either side of the network. Risk scoring. Tokenization. Identity solutions. Consulting. Authorization optimization products. The layer investing the most in closing the gap between layers is the one whose revenue depends on it. Mastercard CEO Michael Miebach has framed it explicitly: more digital payments create more opportunities for value-added services, more services drive more digital payments, and 85% of that VAS revenue is recurring.

Capital One understood this equation well enough to pay $35.3 billion for it. The Discover acquisition, approved in April 2025 and closed in May, wasn’t about processing efficiency. It was about owning both the issuing and network layers. Seeing both sides of the transaction. Internalizing the intelligence and the economics that currently sit with Visa and Mastercard. Capital One projected $175 billion in purchase volume migration to the Discover network and $1.2 billion in network synergies by 2027. That’s a direct bet on the thesis that the value is in controlling the intelligence at the point of decision, not in renting it from someone else.

For Visa and Mastercard, every transaction that moves off their networks isn’t just a fee that shrinks. It’s training data that disappears. The moat gets a little shallower with every card Capital One migrates.

Five Issues. One Number.

This series started with a customer reaching checkout, card declined, and closing the tab. The merchant absorbed the loss. The issuer recorded a clean metric. Nobody was accountable.

Four issues later, the mechanics behind that moment are visible. The issuer decides with the least context. The acquirer shapes the signal but can’t carry the proof of its own intelligence across the handoff. The network fills gaps with cross-network data and is increasingly stepping into the decision itself. Each layer optimises rationally within its own view. The outcome is still irrational.

The authorization rate isn’t a metric you optimise at your layer. It’s an outcome of how well the layers in your chain work together. And right now, the layer investing the most in making them work together is the one capturing the most value from the result.

If you’re an Issuer: your authorization rate is increasingly a function of how well you consume and act on intelligence generated outside your own system. The question isn’t whether your fraud models are good. It’s whether you’re building the capability to integrate network-level and acquirer-level signals into your decision, or paying someone else to make that decision for you, one VAS fee at a time.

If you’re an Acquirer: the optimization you’re building is valuable, but the intelligence that makes it valuable gets compressed at the handoff. Your differentiation lives in the merchant-side context that the network can never see. The acquirers winning this are the ones whose intelligence complements the network’s, not the ones duplicating what the network already does better at scale.

If you’re a PSP or PayFac: you’re sitting between the merchant and the acquirer, which means you’re the first place context gets created and the last place it exists in its richest form. The question is whether you’re building your own authorization intelligence or inheriting your acquirer’s. The PSPs that invest in their own scoring, their own data infrastructure, and their own issuer relationships will compound that advantage. The ones that treat acquiring as a pass-through will find themselves commoditised from above by networks selling POP and from below by merchants demanding orchestration. The middle ground is disappearing. Partner with acquirers who share data and treat authorization as a product discipline, or build the capability yourself. Waiting isn’t a strategy when the layers above and below you are both moving.

The gap between the best and the rest isn’t closing. It’s compounding. The institutions that invest in intelligence at their layer are generating better data, training better models, producing better outcomes, and attracting more volume. The flywheel doesn’t wait.

The authorization chain has one dimension these five issues haven’t addressed head-on. Every optimization, every model, every score operates under a constraint that nobody in the chain fully controls: the fraud decision. That’s where this goes next.

Thank you for reading.

P.S. If reading this series made you wonder where your own authorization chain is leaking value, that’s exactly what I help institutions figure out.

20+ years in payments data and strategy. From being the First Data Scientist at Adyen, to being the First VP of Data Science & Analytics at Checkout.com, to helping over 50 of the top 150 acquirers and issuers globally, through my consultancy.

For Advisory. Speaking. Consultancy. Email me or DM me to set up a call.

Or if you just want to fuel the next series (which will be about Risk Intelligence), buy me a coffee.

The issuer gets the score. At most institutions, that score lands in a rules engine built fifteen years ago, where it sits alongside manual overlays nobody has reviewed since a fraud spike in 2019. The issuer makes its call.

In a growing number of cases, the issuer doesn’t make a call at all. It’s handed the decision to the network.

For thirty years, the four-party model assigned the authorize-or-decline decision to the issuer. On paper, that’s still how it works. In practice, the intelligence that drives the decision has been migrating to the network layer for over a decade. In 2024 and 2025, the pace of that migration picked up sharply.

In the second edition of this series, Issuer Paradox, we went inside the issuer: the institution that holds the final decision but makes it with the least context. In the third edition, Acquirer Divergence, we looked at acquirers and how the best ones shape what the issuer ever sees. This issue goes inside the layer between them. The networks themselves. The party that increasingly knows more about the transaction than either side of it.

Let’s get into it.

257 Billion Reasons They Know More Than You

The network advantage isn’t that Visa and Mastercard built good AI models. Lots of institutions have decent models. The advantage lies in what those models are trained on.

Visa processed 257.5 billion transactions in fiscal year 2025. That’s across $14.2 trillion in payment volume, roughly 160 currencies, and 329 billion total Visa-branded transactions. VisaNet switches about 78% of those directly. Mastercard processed 175.5 billion switched transactions on $10.6 trillion in gross dollar volume that same year. Here’s a detail that doesn’t get enough attention: Mastercard’s 10-K discloses it switches about 70% of all Mastercard and Maestro-branded transactions. That’s up from 55% in 2020. Mastercard has been systematically pulling more volume onto its own network for years, and the remaining 30%, roughly 52 billion transactions, still generate assessment revenue but don’t feed its AI with the same depth.

American Express plays a different game entirely. $1.67 trillion in billed business across 86.6 million proprietary cards. AmEx doesn’t disclose transaction counts the way Visa and Mastercard do, but the structural difference matters more than the numbers. AmEx is simultaneously the issuer, the acquirer, and the network on every transaction. Where Visa and Mastercard see what passes through the pipe, AmEx sees both ends of it. Cardholder income, credit behaviour, full spending history, linked to specific merchant profiles, business details, and settlement patterns. In open-loop networks, issuers are reluctant to share cardholder data with the network. AmEx simply doesn’t have that problem.

But scale isn’t just a bragging right. It’s what makes the AI work.

A plain way to say it: more training data makes models more accurate. That’s how supervised machine learning works. And the network sits at the only point in the chain where cross-issuer, cross-merchant, cross-border patterns are visible simultaneously.

An individual issuer sees its own cardholders transacting at a given merchant. The network sees every issuer’s cardholders at that merchant. When a merchant is compromised, the network detects the pattern across dozens of banks before any single bank has enough signal to notice it. When a new fraud typology emerges in one market, the network detects it propagating across borders while issuers in the next market are still reviewing their own portfolios.

No single institution can build that dataset. Not even the largest global bank. The moat isn’t the model. It’s the data the model trains on. And every transaction that flows through the network makes it deeper.

What the Models Actually See

Each network has built a distinct AI architecture. They’re all designed to do the same thing: score transactions in real time and influence the authorization decision. The differences matter if you’re deciding what to integrate and how deeply.

Visa runs a three-layer stack.

Advanced Authorization, or VAA, is the longest-running neural network in payments, in production since 1993. VAA scores every transaction on VisaNet against more than 500 risk attributes in under a millisecond, producing a score from 1 to 99 that goes to the issuer. It covers two years of cardholder history, merchant behaviour, device and channel signals, and geography. It’s deployed at over 8,000 financial institutions across 129 countries.

On top of VAA, Visa launched Deep Authorization in March 2024: a recurrent neural network built specifically for card-not-present e-commerce, trained on petabytes of contextual data. Where VAA scores broadly across all transaction types, VDA scores sequences and context for the channel where approval rates are lowest and false declines are highest.

Below both sits Visa Risk Manager, a configurable rules engine that allows issuers to build custom rules using network scores as inputs. One Latin American issuer that moved to deep VRM integration saw a 7% increase in approval rate and a 90% drop in false positives within eight months. That’s vendor-reported, so discount it appropriately. Even halved, the gap between shallow and deep integration is commercially decisive.

Visa now offers VAA as a network-agnostic product. Issuers can use Visa’s AI scoring on non-Visa transactions. That positions Visa’s intelligence as infrastructure beyond its own rails, and captures data from competitor-branded volume in the process.

Mastercard’s approach centres on Decision Intelligence.

Decision Intelligence uses predictive analytics to score transactions in real time, incorporating account data, merchant information, device signals, and geolocation. The capability was significantly enhanced by Mastercard’s acquisition of Brighterion, an AI specialist, and, like VAA, it is network-agnostic.

In February 2024, Mastercard launched DI Pro, which adds generative AI through a proprietary recurrent neural network with transformer architecture. The innovation: instead of analysing text like a traditional large language model, DI Pro analyses relationships between merchants in a cardholder’s transaction history. It scans over a trillion data points per assessment in under 50 milliseconds.

Mastercard claims DI Pro reduces false positives by more than 85%. I want to be precise about what that means. Mastercard’s own press release qualifies this as “shown in our own analysis.” No independent third party has validated it. And it measures false positives, legitimate transactions incorrectly flagged, not authorization rate improvement directly. The two are related. They’re not the same thing.

In March 2026, Mastercard announced a foundation model for payments, signalling that DI Pro’s approach is being scaled into a platform-level capability. That’s worth watching.

AmEx’s Gen X is the tenth generation of its scoring model.

The current architecture combines gradient boosting machines, a forest of more than 1,000 decision trees, with LSTM deep neural networks running on NVIDIA infrastructure within a strict two-millisecond latency requirement. It processes roughly $1.2 trillion in annual charge volume. What makes Gen X different isn’t the model architecture. It’s the data. That closed-loop visibility means AmEx trains on variables that open-loop networks can only infer. AmEx has maintained the lowest fraud rate among major networks for over a decade, and the data advantage is a large part of why.

Here’s the honest caveat on all three: no independent study has ever compared these systems head-to-head. We’re working from vendor disclosures and structural inference. What we can say with confidence is that the cross-ecosystem data these models train on gives them a structural advantage that no single institution’s model can replicate.

The Token Trojan Horse

Tokenization started as a security feature. It has become the primary mechanism for networks to capture data that didn’t exist in the card-number era. And that data feeds the AI models we just discussed.

Visa has provisioned more than 16 billion tokens as of September 2025. The growth curve tells the story: the first billion took five years, from 2014 to 2019. The most recent billion took about a single quarter. Half of all digital Visa transactions are now tokenized. Over 30% of all Visa transactions, including in-person, carry tokens.

Mastercard reports about 40% of all its transactions are tokenized, up from 30% in late 2024, with 50% year-over-year growth. Both networks have committed to 100% e-commerce tokenization. Mastercard has set 2030 as the deadline for eliminating manual card entry entirely.

Why does this matter for authorization? Because tokens capture information that card numbers never did.

When a network provisions a token, it records data that was invisible in the PAN era:

• Device type and operating system

• Wallet (Apple Pay, Google Pay, etc.)

• Channel: in-app, card-on-file e-commerce, contactless NFC, or IoT

• Domain controls restricting where the token can be used

• Verification method: biometric, one-time password, or other

The full token lifecycle, provisioning, activation, card updates, suspension, and deletion, creates a longitudinal behavioural dataset that simply wasn’t available when transactions carried raw card numbers. Networks now track credential continuity across card replacements, device migration patterns, cross-merchant token usage, and authentication methods over time.

The impact of the authorization rate is real, though the exact magnitude depends on which figure you trust. Visa reports a 3 to 6 percentage-point improvement in e-commerce transactions. Mastercard’s figures range from 2.1% to 3-6 percentage points depending on the document.

Tokenized transactions carry a richer, fresher credential signal. Issuers trust them more. Approval rates go up. Juniper Research projects network token transactions will roughly double from 283 billion in 2025 to 574 billion by 2029.

Here’s where the flywheel kicks in.

1. More tokenized transactions mean more training data for the network’s AI models.

2. Better models mean higher approval rates.

3. Higher approval rates incentivize more merchants to adopt tokens.

4. More adoption means more data.

The network sits at the centre of that loop. Visa even built a product, Provisioning Intelligence, that uses machine learning to score token provisioning requests themselves, catching 14 times more fraud than token requestor scores alone. The token issuance process itself became training data.

Tokenization isn’t a security upgrade. It’s a data-capture infrastructure that compounds over time. And the networks control it.

From Scoring to Deciding

For most of the network era, the relationship was straightforward. The network provided a score. The issuer made the decision. That’s changing.

Having worked across these systems for more than two decades, I can map the progression clearly. It happened in four steps, each one feeling incremental at the time. The cumulative effect is structural.

• Step one: passive fallback. Traditional Stand-In Processing. When an issuer’s system went down, the network handled transactions using static, portfolio-level rules. Blanket spending limits are nothing smart. Issuers experiencing outages sometimes saw 0% approval rates because the fallback rules were so blunt.

• Step two: scoring as a service. VAA launched in 1993. DI followed. The network started providing intelligence, not just routing messages. Most issuers still think this is where the relationship sits.

• Step three: modifying the question. This is where it gets interesting. Mastercard launched the Payment Optimization Platform in October 2025. POP sits between the acquirer and the issuer, enriching the authorization message in real time using Mastercard’s global transaction data. It evaluates over a trillion combinations of data elements to construct the optimal request before the issuer ever sees it. Early pilots with Adyen and Worldpay reported 9-15% increases in conversion rates. Those are vendor-reported pilot numbers. But the structural point matters: the network is editing the question before the issuer answers it.

Visa’s Real-Time Account Updater works from a different angle. It updates card credentials in real time during the authorization process, before the issuer applies its decision logic. The issuer doesn’t see the stale credential. It sees the current one. The network fixed the input.

• Step four: making the decision. Mastercard’s On-Demand Decisioning, launched globally in October 2025 (excluding India), lets issuers define authorization criteria directly on the Mastercard network. It operates in two modes. In proactive mode, Mastercard responds to transactions without real-time involvement from the issuer. In review-and-modify mode, Mastercard reviews the issuer’s decision and can change it before it reaches the merchant.

Porto Bank, an early adopter, described the experience plainly: the solution required no implementation effort and delivered results with minimal operational risk.

Think about what “no implementation effort” means. The decisioning logic runs entirely on Mastercard’s infrastructure. The issuer’s legacy stack, built twenty years ago and layered with overlays from every fraud cycle since, stays exactly where it is. The quality of the authorization decision is separated from the limitations of the system used to make it.

Visa’s Smarter STIP takes a different path to a similar destination. It uses deep learning to emulate what the issuer would likely have decided during an outage, with a claimed 95% accuracy (vendor self-tested). It demonstrates that Visa can reconstruct issuer-level authorization logic using network data alone.

I want to be clear about how far this has gone. Networks have not become the default decision-maker for authorization of normal traffic. Issuer authorization remains the standard operating model. But the direction is unmistakable: from routing to scoring to modifying to deciding. Each product in that sequence is live and deployed.

And each one is aligned with a straightforward incentive. Transaction processing assessments, the fees that generate core network revenue, are driven by the number of successfully switched transactions. The network makes more money when more transactions get approved. That’s not a criticism. It’s how the model works. And it explains why every product in this section exists.

Follow the Money

If you want to know where a company is heading, look at where it’s earning.

Mastercard’s value-added services revenue, the category that includes authorization intelligence, risk scoring, tokenization, and identity solutions, hit $13.3 billion in fiscal 2025, up 23% year-over-year. That’s roughly 40% of total revenue. Core payment network revenue was $19.5 billion, or about 60%. Five years ago, VAS was a much smaller share. It’s growing faster than the payments business itself. Eighty-five percent of it is recurring.

Visa’s VAS revenue reached $10.9 billion in fiscal 2025, up from $8.8 billion the year before. That’s 25 to 26% growth in constant dollars, significantly faster than core processing. At their February 2025 Investor Day, Visa identified a $520 billion total addressable annual revenue opportunity for value-added services.

Neither company publicly discloses per-transaction pricing for AI authorization products. Those numbers are buried in bilateral contracts with rebates and incentives. But the macro picture is clear. The fastest-growing revenue line at both networks is the one that includes selling intelligence back to the institutions that generate the data.

Mastercard CEO Michael Miebach has framed the strategy explicitly: more digital payments create more opportunities for value-added services. More services drive more digital payments. When you hear a CEO describe a “virtuous cycle” between core payments and intelligence services, believe the revenue mix. It says what the strategy deck can’t.

The networks are becoming intelligence companies that happen to run payment rails. The rails are the distribution channel. The intelligence is where the margin is growing.

Who Competes With the Network?

The networks aren’t building this unopposed.

On the acquirer side, the platforms we covered in The Acquirer Divergence hold data that networks don’t have access to. Stripe’s Payments Foundation Model, Adyen’s Uplift, and Worldpay’s authorization optimization all leverage merchant-side context: device fingerprints, session behaviour, IP addresses, and cart-level detail. That data gets stripped or compressed before it reaches the issuer through the network. The acquirer has already scored the transaction. It just doesn’t tell the issuer what it found. Those acquirer models and the network models sometimes conflict. POP enriching a message might not align with what Adyen’s per-issuer tuning has already optimized. The layers of intelligence can work together. They can also get in each other’s way.

Visa made a significant move in December 2024, acquiring Featurespace for roughly $950 million. Featurespace built the ARIC Risk Hub, processing more than 100 billion payment events annually across 180-plus countries. What made the acquisition strategic was what it enabled: Visa Protect for A2A Payments. In a pilot with Pay.UK, Visa analysed billions of historical UK bank transactions and detected 54% of total fraud value beyond what the banks’ own systems found, with a 40% reduction in false positives. That’s Visa’s authorization intelligence expanding entirely beyond card rails.

The most consequential competitive development might be the Capital One-Discover merger, approved in April 2025 at $35.3 billion. Capital One CEO Richard Fairbank framed the logic explicitly: owning a network lets you deal directly with merchants, create more value, and capture the economics of vertical integration. By routing volume to Discover’s network, Capital One can approximate AmEx’s closed-loop advantage by seeing both sides of transactions while setting its own interchange.

For Visa and Mastercard, this creates a problem that goes beyond fee revenue. If Capital One redirects increasing volume off their networks, it’s not just the transaction fees that shrink. It’s the training data that feeds the AI. Every transaction that moves off-network makes the moat a little shallower.

What Happens When the Agent Has No Fingerprint?

Everything in this issue assumes a human at the checkout. The behavioural signals that current authorization models depend on, device fingerprint, session history, typing cadence, and mouse movement, are all generated by people. AI agents produce none of them.

The networks are already positioning for this. Visa launched the Trusted Agent Protocol in October 2025, an open cryptographic framework that gives agents a verifiable identity and, importantly, gives the network visibility into pre-transaction behaviour: browsing, product comparison, and shopping intent. Data that payment networks have never seen before.

Mastercard launched Agent Pay in April 2025, taking a different approach. Its Dynamic Token Verification Code formats agent credentials to match standard card payment fields, allowing AI agents to submit payments through existing checkout forms without any merchant code changes. In March 2026, Mastercard released Verifiable Intent, an open-source framework co-developed with Google, that creates tamper-resistant records of what a user authorized an agent to do.

Both networks are co-chairing the FIDO Alliance Payments Working Group to define how these new credentials interact with existing authentication standards. If agent identity and payment credentials are network-controlled tokens, the networks become the trust layer for a transaction type that no current issuer model can evaluate independently.

Agentic commerce doesn’t weaken the network’s position. It strengthens it.

In the final issue, we will put the full chain together. Issuers hold the authority. Acquirers shape the signal. The networks are becoming the intelligence layer between them. The question for every institution in that chain is the same one it’s been since the start of this series: are you building the capability to participate in that intelligence, or are you waiting for someone else to make the decision for you?

Thank you for reading.

P.S. If you are looking for a Payments Strategist/Data Scientist to help you figure out how to start leveraging data as an asset or like to educate your organization or audience through an event or webinar based on 20+ years of experience, don’t hesitate to email or DM me to set up a call and discuss how I can help.

Or if you just want to show your appreciation for my work, feel free to buy me a coffee, as it helps fuel the next editions.