The Wall Street Journal published an article titled Banks Fought Against Stablecoins. Now They Are Considering Launching Their Own. The news (i.e., factual) part of the story:

• JPMorgan discussed issuing its own stablecoin, on top of JPM Coin, its existing tokenized deposit product. There’s no active product underway, but the bank says it would evaluate its options as customer demand and regulation develop.

• A group of more than a dozen banks, including Bank of America, Wells Fargo, and Santander, is advancing a joint stablecoin venture covering the dollar first, with the euro and other G7 currencies to follow.

• Thirty-nine state banking associations representing about 3,000 banks unveiled the BankChain Alliance, a bank-owned blockchain platform targeting a 2027 launch to support both tokenized deposits and stablecoins.

My take: Banks haven’t been “fighting” stablecoins. They’re fighting stablecoins that pay yield, because a yield-bearing stablecoin is a direct assault on deposit pricing. That fight hasn’t ended. What’s changed is that banks now want to build the non-yield version themselves rather than cede the payment rail to Tether and Circle.

What Banks Are Fighting

The Journal’s “banks embrace stablecoins” framing as one decision is off-base.

A stablecoin that pays no yield is a payment rail. A stablecoin that pays yield looks a lot more like a deposit or savings account, except the issuer doesn’t carry the same reserve requirements, capital rules, or deposit insurance costs that banks do.

From a bank’s perspective, that’s the problem. The technology isn’t what threatens the deposit franchise; the economics are.

That’s why banks can live with the first and have started building it themselves. They can’t live with the second, which is why the yield question stayed in the Clarity Act fight rather than getting resolved alongside everything else.

Reporting the current wave of bank stablecoin activity as a reversal of the banks’ position on stablecoins misses the real story: banks embrace the payment rail as long as they can keep someone else from using it to reprice the deposit market.

The Interoperability Trade-off

Tokenized deposits are generally built around permissioned infrastructure, where participating banks can control who gets access and how transactions are processed. That works well when the objective is to make existing banking transactions more efficient, particularly around settlement and reconciliation.

Stablecoins solve a different problem. Their appeal comes from being able to move across public blockchain networks such as Ethereum and Solana and interact with parties outside a particular bank or banking consortium.

I’ve made this distinction before: tokenized deposits solve an efficiency problem—reconciliation. Stablecoins solve an effectiveness problem—payments. A bank that builds tokenized deposit infrastructure but ignores stablecoins risks modernizing its back office while leaving the broader payments battlefield to Circle and its successors.

Brookings’ Aaron Klein argued that tokenized deposits solve the programmable-money problem without requiring new infrastructure, and he’s not wrong about the narrow use case. If the money only needs to move among institutions participating in the same network, there’s a strong argument for keeping it inside the banking system.

The problem comes when the counterparty isn’t in the network. At that point, the walled garden that made the tokenized-deposit system attractive becomes a constraint. Banks needs a form of digital money that can move beyond its own infrastructure and interact with the broader ecosystem on networks it doesn’t control.

That’s why the BankChain Alliance is building support for both formats rather than picking one. Tokenized deposits can handle transactions that stay inside the banking system, while stablecoins can provide a way to move money outside it.

The two aren’t necessarily competing answers to the same problem; they’re complementary pieces of a broader payments infrastructure.

What This Means for Community Banks and Credit Unions

I wrote in June that the megabanks' Clearing House tokenized deposit network was a press release about a press release: no vendor chosen, no name settled, and "modest" client demand by Bank of America's own admission.

The new WSJ article doesn't change that verdict, but it sharpens the more important point underneath it: the big banks clearly see blockchain-based money movement as infrastructure worth owning, even if they don’t know which version will win.

JPMorgan can afford to evaluate a proprietary stablecoin on its own blockchain. Most community banks and credit unions can’t, and shouldn’t try. The more interesting opportunity is shared infrastructure that gives smaller institutions access to the same capabilities without requiring each one to become a blockchain company.

That’s what makes the BankChain Alliance worth watching. It’s essentially applying the shared-infrastructure model that the Federal Home Loan Bank system has already demonstrated can work across thousands of smaller institutions to a new piece of financial infrastructure.

Cari Network, led by former Comptroller of the Currency Gene Ludwig, is furthest along. It’s anchored to Ethereum, issues FDIC-insured deposit tokens, and selected Prividium as its infrastructure, a private permissioned chain built to let regulators audit activity without exposing transaction data to other participants.

Whether Cari, BankChain, or another effort ultimately becomes the dominant infrastructure matters less to a community institution today than the fact that the infrastructure layer is beginning to take shape.

Community banks and credit unions need to understand what participation in a tokenized deposit network would mean for their payments strategy, their deposit economics, and their ability to move money outside their existing networks.

In 2025, SoFi opened more new checking and payment accounts than every credit union in the United States combined. It ended 2025 with 13.6 million customers and $38 billion in deposits.

Community bankers react predictably: “We could be like SoFi if we had their _____,” with “digital app” or “marketing budge” (or both) filling the blank.

My take: SoFi isn’t winning because it has a better digital experience or more marketing dollars (although the latter doesn’t hurt). It’s winning because it prices the first product to get the customer, and then goes and gets the second one. It’s an acquisition-and-expansion machine. Credit unions are mostly acquisition-and-retention shops.

The App Isn’t the Strategy

The convenient story is that consumers want what fintechs offer: better apps, faster onboarding, higher rates, fewer fees. All true, but consumers aren’t really looking for a fintech. They’re looking for a better deal from whoever makes it easy to say yes.

A Cornerstone Advisors survey of SoFi customers found that those who switched to the fintech—I mean, bank—credited the mechanics of switching: linking an existing account, direct deposit setup, simple onboarding, and promotional incentives.

The app is an enabler of the acquisition strategy, not the acquisition strategy itself.

Banks and credit unions are obsessed with account opening, but are obsessed with how to make it easy for consumers to move their financial lives—not just open an account—from one institution to another.

Free Isn’t a Pricing Strategy

Somebody will object that the industry has been pricing for acquisition since the 1990s, and that it’s called free checking.

Not quite. Checking accounts got priced at zero because the institution down the street priced it at zero, and once everybody was there, nobody could raise the price without giving someone a reason to leave. There was no math about what “free” would buy, no product it was built to lead to, and no expectation of recovering it.

The problem: the industry doesn’t operationalize the first product as an investment with a defined payback path.

The industry’s cross-sell numbers tell the story. If free was the checking account’s right acquisition price, there would be something on the other end to show for it.

Pricing for acquisition means charging less than the product can earn on its own because you know where the money comes back. SoFi’s loan pricing was a bet on the deposit account, the credit card, and the investment account that came after it. Free checking was a bet that the customer wouldn’t leave.

The choice of first product also determines who walks in. Free checking brings in a customer with no balance, no loan, and no monthly reason to think about the institution. SoFi’s front door for most of its history was a loan (a student loan, at that) from someone with a balance, a payment schedule, and a reason to open the app.

Free checking asks: “How little can we charge without losing money?” Acquisition pricing asks: “What are we buying?”

SoFi Gets the Second Product

Eight in 10 SoFi customers have opened a second product, and one in five have three or more.

SoFi treats the second product as the reason the first product exists.

That difference shows up in the math on day one. If the second product is a hope, the first product has to clear the hurdle rate by itself. If the second product is the plan, the first product only has to bring in somebody worth expanding.

If a credit union evaluates a student loan solely on the economics of the loan, it might conclude it’s unattractive. But if the student loan produces a checking account, direct deposit, credit card, investment account, and eventually a mortgage, the economics of acquiring that customer have changed.

SoFi’s Rewards Program Is an Engagement Platform, Not a Rebate Plan

SoFi doesn’t just acquire customers. It pays customers to perform the behaviors that make expansion happen: funding the account, moving the direct deposit, opening a second product, and referring someone else.

Those are the behaviors that move a customer from being able to transact, to transacting regularly, to holding several products, to treating the institution as their financial home. Each one is a rung, and the program pays people to climb.

Most credit union loyalty programs pay for debit swipes. The points are redeemed for cash back or a gift card, and nothing about the relationship changes. The program is a rebate with a dashboard.

Members Aren’t Leaving You (But Their Deposits Are)

Seven in 10 of new SoFi checking account customers in 2025 moved money in from another institution, and 23% of the deposits came from credit unions.

Those consumers didn’t leave their credit unions. Somebody who moves 15% of their deposits to SoFi while keeping 60% at the credit union still shows up in the member count, still has an open account, and may still name the credit union as their primary institution.

But 15% doesn’t stay 15%. The direct deposit moves next. Then some savings. Then the investment account goes to SoFi as does the next auto loan.

Relationship attrition runs ahead of account attrition. Most credit unions measure the second one and have no metric at all for the first, so the retention dashboard reports good news right up until the day it doesn’t.

Stop! In the Name of Love

The good news: the game is far from lost for credit unions. Half of Gen Z and Millennials said they’d consider moving their primary banking relationship to a credit union that matched SoFi’s digital experience. And that’s not simply app design—it’s product design, pricing design, and relationship design.

That gets us to heart of many credit unions’ checking account acquisition challenge: it’s systemic inability to grow member relationships.

There’s a long list of things credit unions need to do to back into the game, but here are three things they need to stop doing:

• Stop evaluating the first product in isolation. If a student loan is only “worth it” on its own economics, you’ll keep saying no to the very product that brings in the customer who later gives you the checking account, the direct deposit, the card, and the mortgage.

• Stop paying for debit swipes. Pay them (and make it stupidly easy) to fund the account, move the direct deposit, open the next product, and bring someone else with them.

• Stop celebrating account retention. Especially when the deposits, the direct deposit, and the next loan are walking out the door.

For a deeper dive on this, see the new Cornerstone Advisors report How Credit Unions Can Steal SoFi’s Playbook, available here.

No, Jamie Dimon didn’t relinquish the CEO crown at JPMC. The most important job announcement in banking: M&T Bank named a new head of AI engineering. Finextra mentioned it in a 111-word article. M&T’s job description itself is nearly that long:

• Manages budgets across multiple cost centers;

• Oversees enterprise AI engineering and platform teams;

• Owns GenAI platforms and LLM integration patterns;

• Standardizes repositories and CI/CD pipelines;

• Embeds AI into development, testing, security and delivery;

• Establishes AI operating models;

• Operationalizes AI governance and guardrails;

• Embeds controls into platforms and pipelines;

• Manages audit and regulatory requirements;

• Optimizes AI platform cost, licensing and vendor spend;

• Oversees 50–100 employees

My take: This is a lot more important than you think—it’s a sign of a bank making a broad move from AI experimentation to AI transformation. This next stage of bank AI is about building the environment that allows AI to effectively perform work.

The Two AI Budget Buckets

Oversimplifying a bit (but not much), AI budgets fall into two buckets:

• Bucket 1: Access and experiments. This budget buys Copilot seats, ChatGPT Enterprise licenses, a vendor pilot in the contact center, a proof of concept in loan document processing, an innovation team with a mandate and a Slack channel, and maybe a chatbot with a name and a personality.

• Bucket 2: AI environment. This budget goes towards the development of logins for AI agents (with their own permissions), a log of what agents do and a way to undo it, account data reachable without clicking through a screen, and a coherent representation of the customer that AI can access across the core, card system, LOS, and the CRM.

I don’t have the numbers, but my strong bet is that, at most financial institutions (large and small), the lion’s share of the AI budget goes to bucket 1.

It’s easy to see why. Bucket 1 spending is easy to explain, easy to scope, easy to stop, and makes someone faster. A $40 per seat license that saves an analyst four to eight hours a week is a no-brainer.

Bucket 2 spending has none of those properties. You can’t demo it, the business case reads like a plumbing invoice, and the payoff is down the road.

Bucket 1 Only Goes So Far

AI does three kinds of work: 1) see (read, classify, extract, summarize); 2) say (draft, respond, explain); and 3) do (waive an overdraft fee, raise a credit limit, issue provisional credit, re-age a delinquent account, book/fund a loan).

PenFed Credit Union’s AI agents exemplify this:

• Wingman provides member service reps with profile summaries, email drafts, and accesses the institution’s knowledge base, cutting handle time by ~10% and back office support calls 30%.

• Max pulls payment histories and validates contact information so collections reps walk into member conversations prepared. Those are the assist and recommend rungs, where risk is contained and the plumbing requirement is modest.

• Penny is an IT service management agent that resolves routine requests and executes actions employees previously had to route to a human to get done. The AI agent enabled PenFed to discontinue its internal IT support phone line.

• Ace is the credit union’ first member-facing agent, sitting on the mobile app and online banking, answering questions grounded in a member’s own data and making payments, transfers, and recurring payment setups.

PenFed worked through governance, change management, and testing before it scaled anything, then shipped internally first.

Bucket 1 doesn’t necessarily mean “see and say” and bucket 2 doesn’t necessarily mean “do.” Bucket 1 can include an agent. But bucket 2 is what makes agents scalable, governable, and reusable.

Assessing the Impact of Your Bucket 2 Budget

Is your bucket 2 money having any impact? Here are 4 tests to determine the answer:

1. How many systems does an employee touch to answer a balance inquiry?

2. How does an agent get its own identity, separate from a borrowed employee login

3. Can a non-human take an action and undo that action?

4. Show me a log entry from an agent, not from a person.

If the answers are “six,” “it uses a service account somebody set up in 2019,” “no,” and “we’d have to check with the vendor,” your AI budget is likely going into bucket 1 and if there is anything going to bucket 2, it’s not having much of an impact.

What Building the AI Environment With Bucket 2 Buys

The payoff shows up as speed. Olivia Boles, who runs platform development and engineering at PenFed, described the hard part as ironing out security, privacy, trust, and governance ahead of that first member-facing agent.

Once that was done, her team could “freely add actions thereafter,” because every new one reuses the same secure framework. Ace is now the orchestrator for a set of secondary agents covering mortgages, consumer loans, IRAs, and investments.

The expensive work happens disproportionately up front. Once the environment exists, each additional capability gets cheaper and faster.

Without it, here’s what happens: a vibe coding product manager stands up a working prototype in an afternoon, and the build is nearly free. But everything downstream isn’t: the data, the permissions, security review, a path to production. The prototype sits in a demo environment for months and dies there.

When the environment exists, that vibe coded app has somewhere to go. The cost of trying something drops toward the cost of thinking of it, and the constraint on an institution stops being its technology budget and becomes the quality of its ideas.

OK, that’s a bit of hyperbole, but the constraint does shift from the cost of building something to the institution's ability to decide what’ s worth building.

M&T Moves From Experimentation to Transformation

M&T’s chief data officer has talked about deploying AI tools to 17,000 employees, which is a bucket 1 move. Hiring an engineering leader to own platforms, architecture, and governance is a bucket 2 signal.

The pay range for the position is in the $300k range, with responsibility for budgets across multiple cost centers, AI-enabled SDLC transformation, guardrails, operational controls, and audit readiness.

Advisory roles don’t come with cost center ownership. This is a hire with a checkbook.

Governance Belongs in Engineering

Note where the governance sits in M&T’s job description: the engineering department.

Bank and credit union AI governance today is an approval body: a committee, an acceptable use policy, a model risk process, a form. Necessary, but backwards-looking.

The missing piece is AI deployment governance: who can ship what, to which systems, with which entitlements, and how it gets rolled back at 2am on a Saturday.

Putting governance under the person who owns the platforms means the controls get built into the pipeline instead of enforced at a meeting.

PenFed’s version of this was augmenting an already mature change management process with AI-specific tools before scaling, which is why its agents could expand scope after launch instead of stalling in review.

Who Pays For Bucket 2

Productivity gains don’t pay for transformation. They increase capacity. But they don’t—and won’t—deliver the promise and potential of AI to achieve radical compression of process cycle time and process cost reduction.

That begs a question: Who’s going to pay for bucket 2? It can come from one of three places: new money, reallocated money, or realized economic value.

Best of luck with the first two.

Where does realized economic value come from? It can come from revenue, losses, capacity, headcount, or technology spend. But the most immediate cost-reduction opportunities are in work currently performed by people or vendors.

That produces a Catch-22: You can’t reduce staff or reap tech spend reductions until after you’ve made investments in bucket 2.

But when you do make that investment, a slew of tech spend reduction opportunities presents itself: overflow and after-hours contact center, card dispute processing, document processing BPO, collections agency placements, standalone chatbots, and outsourced development.

Every one of those is a check written to a third party that an AI environment can begin absorbing. PenFed retiring an internal support phone line is a small, but real example of that conversion.

The cost of execution doesn’t disappear when AI gets cheap. It moves out of headcount and vendor invoices and into infrastructure, and the institutions that prebuild the destination get to move it deliberately instead of scrambling later.

The money to build the AI environment is a massive new vendor opportunity—and that's why the vendors are scrambling now.

The pundits will say this will happen in a 3 to 5 year timeframe. Too aggressive, and the major reason why is contracts. Even if you could build the AI environment to move everything, you’re locked into existing contracts.

AI transformation doesn't require waiting for every contract to expire. Institutions can (should) build the environment now while existing systems and contracts remain in place, then redirect work as contracts roll off.

Watch the Movie Credits

A “head of AI” can mean strategy, use cases, evangelism, and a roadmap. A “head of AI engineering” signals something different: ownership of the operating environment that gets AI into production—identity, platforms, pipelines, entitlements, controls, and deployment.

When the second title starts showing up across the industry, you’re watching balance sheets reallocate in real time.

Not every institution has to fund bucket 2 right now. You can sit it out and play catch up in the 2030s. The choice is yours.

Meta’s Mark Zuckerberg published an “encyclical” (whatever that is) under the title “The Future is for Everyone.” It’s a lobbying memo masquerading as public policy—speed up data center permitting, keep the silicon export controls, loosen the rules on training data and distillation, you know the rest.

The part that matters for banking is his claim that everyone will soon have a personal AI agent handling their “health, career, finances, hobbies, home management.”

The “finance” part of that implies that an agent will sweep balances to the best yield the moment rates move, claim every benefit, dispute every fee, and re-shop every renewal. Consumers will get a financial life that’s—finally—fully optimized.

My take: The money movement part is real (it's been happening for 10+ years without AI agents). The optimization part is hype—constant account shifts carry costs, and the gains are smaller than advertised. Better financial outcomes require people to change behaviors, not just accounts. Agents might optimize rates and fees, but they'll be doing it against institutions running their own agents—agents that have run the same negotiation a million times and know exactly how much they need to concede. Whether consumers come out ahead is going to be a hard question to answer.

The real contest isn’t just one agent against another over a single customer’s savings rate. It’s a battle for which side can scale to millions of negotiations the fastest. Meta and Google will be doing it across their entire user base. A $2 billion bank would be doing it with whatever its core provider ships, on the core provider’s timeline.

Moving Money is the Easy Part

AI agents will move money. I’m not arguing with that part (because it’s been happening for years).

Checking accounts have become paycheck motels. Money lands, sits for a few days, and moves on. $3 trillion was displaced from banks and credit unions to Robinhood, Fidelity, SoFi, and their fintech brethren with no AI agent involved. Autopay and card-on-file updates are real friction, but they’re speed bumps.

Personal AI agents will take care of the easy stuff. Opening new accounts? Piece of cake. Transferring balances? Child’s play. Figuring out who’s got better rates or lower fees? Easy-peasy.

But that doesn’t get consumers nearly as far towards “financial nirvana” as Zuck and the other AI overlords make a personal AI agent out to be.

Behavior Change is the Hard Part

Take a consumer with $6,000 revolving at 22%, costing them about $1,320 a year. The agent moves it to a card with an 18-month promotional rate and the interest goes away (or down). Real money, and exactly the kind of move agents will be good at. When the promotional period ends, $2,000 of the transferred balance is still sitting there at the go-to rate, and there's another $5,000 on the original card. The agent solved the balance but not the spending that produced it.

Financial performance runs on behavior. How much goes out against what comes in. Whether saving happens on a schedule or whenever there’s something left. Whether there’s enough cushion that a transmission repair doesn’t turn into a balance that revolves for three years. Product selection works around the edges of those decisions.

We’ve been through this before: PFM tools, budgeting apps, financial literacy programs. All fell short, mostly for one reason: they didn’t change people’s behavior.

Knowing what to do was never the problem. An AI agent solves information and execution—neither of which is the problem.

An agent that could change behavior would have to do things consumers won’t accept: decline the purchase, veto the vacation, invest money someone had other plans for. The moment a personal AI agent has enough authority to change financial behavior, it stops being a convenience and starts being your parent.

Optimization is a Pipe Dream

Optimization is a technical term and requires something called an objective function—a number to maximize. Consumer finance doesn’t have one, and its pieces conflict:

• Yield argues for locking money into term deposits, liquidity argues against it.

• Points argue for concentrating spend on one card, interest cost argues for whichever card is carrying a balance.

• Credit standing argues for old accounts left open and applications kept rare, bonus capture argues for opening something new every quarter.

Two big challenges here:

• The conversion rates don’t exist. A Membership Rewards point is worth about 0.6 cents against a statement credit and north of two cents against the right partner redemption, and which one applies depends on a trip the customer hasn’t booked on dates they don’t know. An agent choosing between a points card and a cash back card has to price that option using future behavior nobody can specify.

• The objective has to come from somewhere. Either the customer defines it, which means the customer is doing the hard part while the agent does arithmetic, or the vendor picks a default, which means millions of consumers get optimized against a definition of “better” that someone at an AI company chose.

An investment portfolio can be graded. There’s an index, a peer group, a risk-adjusted return, and a methodology for deciding whether the manager earned the fee.

A consumer household has none of that. You can’t run your financial life twice. There’s no version of the last three years where you kept the other card, stayed at the other bank, and left the money alone, sitting next to this one for comparison. There’s no base case, so the agent’s performance can’t be objectively verified. Nobody can prove it added (or subtracted) value.

Hard to defend for a product that promises optimization. Easy to live with for a product that wants to keep charging for it.

The Agent Won't Just Add Accounts, It Will Cycle Them

Say the objective ends up being yield, which is the likeliest default. Follow it where it goes.

Agents didn't create the fragmentation. A millennial couple like Cameron and MacKenzie already has ~40 financial relationships, and no AI agent helped.

What an agent changes is the rate of turnover. Rates move constantly and they don’t move together, so an account that’s best in March is fourth-best by June. An agent chasing the leaderboard doesn’t settle at 40 accounts. It runs 40 at any given moment and a hundred over the course of a year, opening and abandoning as the rankings shift.

Chasing the top rate beats sitting still, but the question is by how much.

Getting to real money takes real balances. At $25,000, a 100-basis point difference is $250 pre-tax, which is a nice bump. But a household capturing a full point was parked at 0.01% to start with, and fixing that is one move, not a hundred. Once the money is sitting at a competitive rate, the spread between best and decent runs 30 to 50 basis points—$75 to $125 on that same $25,000, and less after tax.

Top rates are also advertising. The leader is usually buying deposits with a promotional rate that steps down after four to six months, so the agent captures the differential for part of the year and then moves again.

The gain is real but small for most households. The costs come with every opening:

• Every account generates a 1099-INT, and the customer files the return

• Every institution holding the customer’s identity data is a breach target

• Every account is a fraud surface somebody has to monitor

• Every card opened for a bonus adds an inquiry and lowers the average account age

None of that shows up in the gains the agent reports.

The Quantifiables Win, the Intangibles Lose

The agent has to optimize something, so it optimizes what carries a number: rate, fee, cash back percentage, minimum balance. These are all published, structured, and comparable. What gets left out of the calculation are the intangibles:

• How good is the institution’s advice?

• How well does the institution personalize and know its customers?

• Does the institution do what’s right for the customer and not just its bottom line?

None of these can be evaluated from the outside. You find out after you're a customer.

Anyone who’s dealt with an insurance company after a car accident knows not to buy the next policy on rate alone, and knows why: what a provider is worth becomes clear at the moment you need something from it. A comparison can’t give you that.

The objection is that an AI agent can read Yelp, Reddit, and Trustpilot. Those are other people’s claims under other people’s circumstances, public because someone was angry enough (or incentivized) to post.

If personal AI agents treat reviews as an input to money movement, manipulating—or even falsifying—reviews skyrockets. The technology that makes the agent possible is the same technology that makes a convincing fake review easy to produce.

Agents will be good at the measurable part, and they'll report improvement. Consumers looking for optimization won't know the difference.

In Case You Hadn’t Heard: AI Can Make Mistakes

Agents might not be that good at the measurable part.

Every AI product on the market ships with a disclaimer that its output can be incorrect. I saw one when I asked Claude to look over this post and tell me what readers would argue with.

The disclaimer means the user has to do the verification. Checking the agent's work means doing the agent's work. A human capable of checking the recommendation doesn’t need the agent. And a human who can’t check it has no way to grade it.

The Coming Battle

Everything above assumes the products sit still. They won’t.

Zuckerberg’s thesis is that broadly distributed capability produces better outcomes than concentrated capability. That implies that if every consumer gets a personal AI agent, every institution will get its own agent—one that reprices continuously. That creates two problems for the personal agent:

• The arbitrage closes before it opens. An institution that can reprice individually doesn’t lose a customer over 30 basis points. It matches, right up to the point where keeping the money costs more than replacing it. The agent’s threat to leave stops being leverage and becomes an input to a retention model. What the agent captures is what the institution calculates it has to give up, not the entire gap.

• There’s nothing left to compare. Comparison assumes a posted price. Individualized dynamic pricing means the rate doesn’t exist until the customer asks, and what comes back depends on the institution’s read of that specific household. The agent’s core function is to compare and select. That doesn’t work when every quote is a negotiation.

Two machines at the table and one of them has the file.

Financial institutions will also be able to see who has an agent. Agents leave tracks: constant API queries against the account, retention offers answered in seconds instead of days, balances moving in round numbers on a schedule. A consumer who generates those signals is telling the bank they’re ready to walk, and gets priced accordingly. A consumer who generates none of them is telling the bank they aren’t going anywhere, and gets priced accordingly too.

That’s the opposite of what Zuckerberg is arguing. Handing agents to everyone doesn’t level the field. It sorts customers into the ones who’ll leave and the ones who won’t, and the second group pays for the first.

Insurance already ran this experiment. Carriers built models to identify which customers were least likely to shop, and charged those customers more. Regulators called it price optimization and killed it.

Deposit and fee pricing hasn’t faced that test—yet. An institution that quotes 4.3% to a household with an agent and 3.1% to an identical household without one is doing what insurers were told to stop doing. Elizabeth Warren will be all over this.

The real contest isn’t just one agent against another over a single customer’s savings rate. It’s a battle for which side can scale to millions of negotiations the fastest. Meta and Google will be doing it across their entire user base. A $2 billion bank would be doing it with whatever its core provider ships, on the core provider’s timeline.

What Gets Repriced

The profitability of a lot of financial products depends on customers leaving value on the table:

• Premium cards priced on the assumption that most cardholders don’t recoup the annual fee.

• Rewards checking paying a headline rate on the assumption that a share of accountholders miss the monthly qualifiers.

• Promotional CD rates that depend on customers letting the account renew at the default.

I wrote about that gap a few weeks ago. If every customer claims everything, those products don’t survive at today’s prices. They get repriced, and the arbitrage closes.

Zuckerberg says consumers come out ahead in aggregate. They don’t. If products get repriced until the arbitrage closes, the gains don’t accrue—they get redistributed. The customer who extracts $1,200 from a $695 card loses when it gets repriced. The net effect is a transfer from sophisticated customers to unsophisticated ones, plus whatever product variety disappears when the cross-subsidy stops working.

Agents are going to manage money. Measuring how well they do it is a problem yet to be solved, and I’d want it solved before I hand over the keys to my accounts to a personal AI agent. OK, before my wife hands over the keys.

Come into the Fintech Snark Tank and tell me where I’m wrong.

AI is driving the biggest investment boom in decades. Most banks will never finance any of it. Their real opportunity sits one layer down.

The Wall Street Journal reported that AI-related capital spending now carries a meaningful share of U.S. economic growth—roughly $1.5 trillion in annual investment across AI categories, with that spending and the associated equity gains accounting for about a third of recent GDP growth.

Data center construction has become the fastest growing segment of private construction, reaching a $68 billion annual rate in June, up more than $20 billion in a year, as other private construction categories declined by more than $101 billion.

Where is the money coming from? Hyperscalers are funding these facilities in the bond market and through their own balance sheets. Regional and community lenders are not in that capital stack—and won’t be.

My take: The banking opportunity created by the AI boom is banking the ecosystem that's building it. The capital behind the AI build-out is out of reach for community institutions. Businesses doing the work on-site are within reach, however. Community banks need to focus on where AI investment is physically showing up.

The Data Center Supply Chain Is the Opportunity

The media covers the build-out as a semiconductor story. A data center is a construction project before it’s a technology asset, however. Thousands of workers from hundreds of businesses—concrete, steel, electrical, cooling, equipment rental, logistics, and more—report to one address for three years.

The hyperscalers and national contractors already have their banking relationships. The opportunity for community banks sits further down the supply chain: the regional and local contractors, suppliers, and service businesses that provide the labor, equipment, and ongoing operations.

These businesses already know how to build. What they need are banks that understand growth: deposits, treasury management, equipment finance, working capital lines, and relationship lending that moves at the speed of a construction schedule.

The AI boom is concentrating thousands of commercial banking prospects into a handful of zip codes for two to three years at a time.

A Data Center Announcement is a Market Selection Study

Banks spend a lot money researching and answering four questions:

• Which industries should we target?

• Which counties should we expand into?

• Where should we hire lenders?

• Where should we put a loan production office?

The typical inputs for those decisions are lagging indicators—deposit market share, census projections, three years of past loan performance, a consultant’s MSA screen.

A data center announcement answers all four questions at once, before construction begins. Every announced data center comes with a pre-built commercial prospect list and a geographic expansion roadmap.

The Relationship Outlasts the Project

A data center employs thousands during construction and a few dozen after commissioning. The lifetime value isn’t in the site—it’s in the businesses that win the work.

The electrical contractor, concrete supplier, equipment rental company, and service businesses involved in the project experience a growth event. They need more equipment, more working capital, more sophisticated treasury services, and a banker who understands contractor cash flow and can make decisions quickly.

The opportunity doesn’t end when the construction crews leave—data centers aren’t one-off projects. Campuses are built in phases, and completed facilities create a permanent vendor ecosystem—security, HVAC service, electrical maintenance, landscaping, and other services that continue as long as the facility operates.

The advantage goes to the bank that knows the local construction market, understands contractor cash flow, can make equipment decisions quickly, and can underwrite the business.

The construction phase creates the relationship. The operating phase creates the lifetime value.

Bottom line: Community banks won’t finance a $40 billion data center. They can finance the businesses that pour the concrete, wire the substations, rent the cranes, fuel the trucks, and maintain the facility after the construction crews leave.

The AI boom may not create a new generation of AI firms for community banks to bank, but it could create a new generation of conventional businesses with unconventional growth opportunities.

Cerulli Associates projects that $124 trillion in wealth will change hands through 2048, ~$105 trillion going to heirs and ~$18 trillion to charity. Boomers and the Silent Generation account for roughly $100 trillion of that, or 81% of the transfer.

That number ripples across financial services: wealth managers worry about retaining heirs, banks and credit unions worry about winning them, and fintechs use it to justify their addressable market.

My take: Cerulli's number could be right—I have no reason to doubt it. But banks' obsession with the "great wealth transfer" is causing them to focus on the wrong opportunity. For millions of households, the defining financial challenge over the next 25 years will be spending down wealth to pay for aging and long-term care.

Visa’s Vexing Verdict

Visa’s economists published their own version of the math. They started with the $93 trillion Boomers are sitting on and subtracted the top 1%, liabilities, retirement spending, taxes, and charitable giving.

What’s left: ~$36 trillion transfers over the next 20 years, and only $8 trillion of it gets spent. The rest sits in heirs’ savings and investment accounts, because 3/4 of them are already doing fine.

That’s less than a third of the number the industry keeps quoting. And I think Visa’s estimate is generous.

It’s Too Late Baby Now

Even if wealth makes it to heirs, it's less valuable to banks than the headlines imply:

• Half the transfer comes from the wealthiest 2% of households. Their money isn’t at a community institution now, and it won’t go there after an inheritance.

• It goes sideways before down. A big chunk of it moves to a surviving spouse before it goes anywhere near a younger generation.

• Only ~30% of American households will receive an inheritance. For everyone else, the great wealth transfer is something that happens to other people.

• Inheritances arrive in heirs’ 50s. Parents die in their 70s and 80s, so heirs are ~40 to 60 years old. Millennials planning to inherit $46 trillion will be AARP members by the time it shows up.

• By then, banks’ opportunity has passed. An inheritance at 55 just patches a retirement account that was already behind. More importantly, a 55-year-old has decades-old banking, investment, and lending relationships that aren't easily displaced by a larger balance.

What Visa Didn’t Envision

After the money moves sideways to a surviving spouse, it goes next to assisted living homes, nursing homes, and memory care units.

Assisted living runs ~$75k/year, nursing homes ~$130k, and memory care costs more than either. Fidelity’s just-released estimate says a 65-year-old retiring this year will spend $185k on out-of-pocket medical costs, and that figure excludes long-term care.

The median household headed by a 65- to 74-year-old has a net worth of ~$410k, and a lot of that is the house. Do the arithmetic on a few years of care and the estate is gone.

This isn’t speculation. A JAMA study tracked 200,000 Medicare beneficiaries who entered nursing homes paying privately. Among those still there four years later, 62% had burned through their assets and gone on Medicaid.

That’s the wealth transfer for a huge share of the middle of the distribution. The money goes to a care facility, the house gets sold or clawed back through estate recovery, and the “heirs” get the funeral bill.

Three Strategies That Don't Require Anyone to Die

Bank and credit union execs are dreaming up ideas—or worse, listening to Gen Z consultants tell them how to reach consumers their age—in order to benefit from a ‘great wealth transfer’ that simply isn’t going to happen.

The transfer will occur over such a long period that it’s foolish to put a number that has a trillion after it in a strategy deck. That doesn’t mean banks and credit unions should do nothing. What they can/should do:

• Model deposit runoff from care, not from heirs. The industry’s stated fear is that heirs will move inherited money to a competitor. The bigger and nearer risk for a bank with an aging deposit base is that the money leaves to pay a care facility and doesn’t land at any financial institution. It’s forecastable, but few run that model.

• Enable family financial caregiving. Many adult children pay their parents’ bills, handle fraud alerts, coordinate with caregivers, and exercise power of attorney. Permissioned access, tiered controls, trusted-contact workflows, and fraud monitoring designed for cognitive decline remain under-built. This is one of the few ways to establish a relationship with the eventual heir before the estate settles.

• Sell de-cumulation, not accumulation. Home equity conversion, structured drawdown, long-term care funding, annuitization, Medicaid-aware planning. Most community banks and credit unions offer none of it. Most fintechs are still building for the 28-year-old opening a Roth IRA. The elder financial services market has the advantage of consisting of people who have money right now.

Baby Boomers: The Next Gen of Banking Consumers

$124 trillion is a projection about gross asset movement over 24 years, most of it originating with households community-based institutions don’t serve.

The real number is smaller, it lands later, and it stops at a care facility along the way.

In January 2020, in a piece titled Baby Boomers: The Next Generation Of Banking Consumers, I argued that the next generation of banking consumers wasn’t Gen Z.

I listed 4 things aging Boomers needed from their banks: 1) protection from fraud and predatory practices; 2) permissioned account access for family members and trusted advisors; 3) integration with estate planning; and 4) help managing healthcare costs.

My read at the time was that nobody in the industry offered it. Six years later, that’s still mostly true. I also predicted that Big Tech or HMOs would get there first. Wrong, but the window has stayed open because no one has come through it.

Bottom line: Institutions that serve aging consumers and the adult children managing their money will own the relationship long before an estate settles. The delusion isn't the $124 trillion. It's believing it could end up in your institution.

BTW, that noise you hear is my wife’s iPhone buzzing from the texts she’s getting from our daughters, wanting to know what the hell dad just posted on his blog.

Revolut told its employees last week that their shares are worth $2,017 apiece, making the company worth $115 billion. That makes the 11-year-old fintech—oops, bank—more valuable than U.S. Bancorp and PNC, and in line with Capital One’s valuation.

The latest deal moves at least $750 million of stock—which means two-thirds of 1% of the equity sets the price of the other 99.3%. Every company would love to name its own share price. Only a few—like Revolut—get to.

My take: This valuation is insanity (but we do live in an insane world). It rests on one big assumption: that Revolut will/can become one of the largest financial institutions in the world without ever being priced like one.

Nobody in Public Markets Pays 65x For This

Revolut’s 2025 results deserve their due: revenue up 46% to £4.5 billion, pre-tax profit up 57% to £1.7 billion at a 38% margin, retail customers up 30% to 68.3 million. A $115b valuation translates to 19x trailing revenue, 50x pre-tax profit, 65x net profit.

No large financial institution trades anywhere near there. The best-run banks sit in the low-to-mid teens. Barclays, which Revolut just passed in market value, generated 4.7 times Revolut’s net profit last year and trades at 11 times earnings.

Scenarios: Take Revolut’s $1.76 billion of net profit, compound it four years at three growth rates, and apply two exit multiples.

Just one cell in that grid clears today’s price. Reaching it takes four straight years of 40% profit growth and a 20x multiple—that no large bank sustains—and it pays about 4% a year for the trouble. At a 15x multiple, the bank’s—oops, fintech’s—valuation is down 12% from today’s.

If growth slows to 20%—which would still make Revolut one of the fastest-growing institutions on the planet—a 20x multiple costs the new investors half their money.

The Buyers Can Do This Math Too

Nobody writing a check at $2,017 is confused about the multiple. Four reasons why they’re writing checks:

• They’re underwriting the next buyer. Revolut has told investors it’s targeting $150 billion to $200 billion at IPO. Buy at $115 billion, sell into that book in 2028, collect 30% to 74%. That works whether or not 65x trailing earnings is defensible, right up until the public market declines to pay that number.

• Fund math beats asset math. A $200 million check in a $10b fund is 2% of the portfolio. The upside dwarfs the downside, and a stake in Europe's most valuable startup looks good in the fund's marketing deck whether or not it ever pays off.

• The mark is part of the product. Funds holding Revolut at $45b or $75b just got written up. The buyer at $115 billion supplies the price that validates the seller’s book, which flows into reported NAV, IRR, and the next fundraise. There’s a constituency for higher prices that has nothing to do with Revolut’s earnings.

• Some of them believe the thesis. The honest bull case is that Revolut’s cost to serve sits permanently below any incumbent’s, so bank multiples don’t apply. Nubank is the live test of that argument, and the market’s answer is 23x.

None of these is a claim about Revolut’s worth—they’re reasons to own the position.

Nubank is the Better Comparison Point

The WSJ article contrasts Revolut and Barclays, but Nubank is a better reference point. At a $67 billion valuation, Nubank closed 2025 with 131 million customers (135m by the end of Q1 2026), $6.99 billion of revenue, and $2.87 billion of net income.

Profit per customer is nearly equal—~$22 at Nubank vs. $26 at Revolut—so nobody gets to explain the gap with superior unit economics.

Compared to Revolut, Nubank has double the customer base, 63% higher profit, but just 58% of the valuation. Nubank now trades at 23x earnings and 9.6x revenue. Apply those multiples to Revolut and you get a valuation between $40 billion and $58 billion.

Lending: Where Bank Multiples Are Born

Revolut’s profit comes from subscriptions, interchange, FX spread, trading commissions, and float on £50b of customer balances. The loan book is £2.2b, ~4% of those balances.

Most of Revolut's earnings still come from fee businesses rather than lending, which helps explain why investors haven't yet applied traditional bank multiples.

The company is changing that (on purpose). Credit cards, unsecured personal loans, and overdrafts are rolling out in the UK, and the US national bank charter application was submitted in March. Strategically that’s the right call, and it’s probably the only route to the growth this valuation requires.

It’s also the move that turns Revolut into the kind of institution that trades at 11x earnings.

Lending brings durable spread income and scale. It also brings capital requirements, a loss curve that surfaces 18 to 30 months after origination, and an earnings stream that investors know how to discount.

One concern: the credit cycle hasn’t tested Revolut’s loan portfolio yet.

Bottom line: Buyers at a $115b valuation are paying 65x earnings for illiquid common stock in a company whose CEO has said it won't list before 2028, on the theory that a bank which just started lending will never be valued like a bank. Sign me up (not)! To justify this valuation Revolut has to become a giant bank and never get priced like one.

Is Fintech Vs. Banks Still a Valid Comparison?

A broader question: Revolut is a bank—why is it still getting fintech multiples? After all these other fintechs applying for bank charters get one, will we start counting their revenue and valuation growth as “bank growth” and not “fintech growth”?

Fintech cheerleaders have been running the same comparison for years: “Revolut grew revenue 46%! Nubank grew net income 46%! The banking industry grew low single digits! Fintech is winning!”

Just one problem here: That comparison has been banks versus banks for a while now.

Revolut now operates as a licensed bank in 30+ markets and picked up its full UK license in March. Nubank is a regulated bank in its core markets. Every deposit they gather is a bank deposit. Every loan they book is a bank loan.

Fold their numbers into the industry aggregate and the story changes shape. Banking didn’t lose share to some outside force. A handful of new banks grew fast and most old banks didn’t, which is what happens in every industry where a young competitor works off a small base.

The category is doing the work here. Grow fast and you’re a fintech. Grow slowly and you’re a bank. If a chartered, deposit-taking, lending institution with 68 million customers doesn’t count as a bank, the word has stopped meaning anything except “an institution that doesn’t grow fast.”

That’s the same sleight of hand holding up the $115 billion.

In Cornerstone Advisors’ 2026 What’s Going On in Banking report, I equated bankers in 2026 to Dorothy and her Wizard of Oz pals heading into the woods, arm in arm, chanting “AI, crypto, and fraud, oh my.”

The study found that going into 2026, financial institution execs were:

• Optimistic. Credit union execs more so than bankers.

• Growth-focused. New customer growth jumped to the top of the concern list.

• Deep into AI. Gen AI deployment among credit unions hit 59%. Agentic AI: 17%.

• Talking tokenization. On the agenda at 71% of boards, but 0% deployment.

Halfway through the year, nobody’s been eaten. But the trees look different than they did at the beginning of the year, and a few things came crashing out of the underbrush that nobody had anticipated in January.

What did bankers spend the first half of 2026 learning that they didn’t expect to learn?

1. Competition changed faster than expected (fintechs, checking, Bizumers).

2. Technology changed faster than expected (AI plumbing, agentOS, Plaid, Anthropic).

3. Distribution didn’t change the way people think (the Chase branch myth).

4. Money is becoming more mobile than banks assume (deposit mobility, tokenization, stablecoins).

The economy didn't collapse, but it didn't provide much help either. Growth softened, confidence weakened, and uncertainty stayed elevated.

My take: The banking industry spent the first half of 2026 adapting to structural shifts that had little to do with GDP or interest rates: fintechs kept taking checking and payment account share, AI infrastructure accelerated, and the assumptions underlying deposit strategy continued to erode.

Fintechs Ate the Checking Account

Big fintechs—PayPal, Block—hit 80% on the “significant threat” list going into 2026, up from 47% three years ago. Challenger banks like Chime posted the biggest jump of the group, from 21% to 70%. Crypto providers entered the chart at 29%.

The fear is well-founded: Fintechs opened 56% of new checking and payments accounts in 2025, up from 36% in 2020. Megabanks fell from 24% to 13% over the same stretch. Credit unions and community banks are mired in single digit market share.

Who’s doing the damage: PayPal (16% of new accounts), Chime (10%), Venmo (10%), with SoFi and Cash App climbing through the 5-6% range.

Why this is happening: This is a product-market fit story. Fintechs are winning because they're designing accounts around the financial lives people lead—digital-first, multi-provider, increasingly entrepreneurial, and less loyal to a single institution.

AI: The Battle Lines are Getting Clearer

The first half of 2026 produced some important AI infrastructure announcements. Plaid launched a financial foundation model. FIS partnered with Anthropic. Fiserv introduced agentOS. These weren’t announcements about better chatbots. They were announcements about who intends to own the operating system for banking AI.

These moves change the face of the vendor landscape in banking:

• Plaid expands more broadly into credit underwriting and scoring, offering a(nother) legitimate alternative to FICO.

• The big 3 cores take on big AI lab partners providing access—albeit not very direct access—that their client base could never get on their own.

But here’s what’s getting lost in all the excitement over Claude, ChatGPT, Gemini, and Llama: the strategic battle is over the orchestration layer.

Vendors love to talk about “model neutrality.” It’s become the AI equivalent of “open architecture.”

But model neutrality is just marketing until it’s written into the contract. The vendor that controls how AI agents authenticate, retrieve data, invoke systems, and hand work off to other agents owns the customer relationship regardless of which foundation model happens to sit underneath.

Banks think they're selecting AI models. In reality, they're selecting the platform that will determine which models they can use, how easily they can change them, and who captures the value created by AI.

As foundation models commoditize, competitive advantage shifts upward. The model becomes a component, not the product. The orchestration layer—the software that decides which model to call, when to call it, what context to provide, and what actions to take—becomes the strategic control point.

It’s Not About Stablecoins, It’s About Deposit Mobility

NO—ChatGPT did not write that headline for me.

The GENIUS and CLARITY Acts moved the conversation from whether digital dollars belong in the financial system to how they’ll operate. But bankers are asking the wrong question.

The biggest strategic mistake banks can make is thinking the issue is stablecoins. It isn’t. The issue is deposit mobility.

The concept of “core deposits” no longer holds. There are no core deposits. They’re all movable. That trend didn’t start with stablecoins, and it won’t end with them. Stablecoins simply accelerate a shift that’s already underway.

The question banks need to answer: “What happens when deposits become dramatically more mobile?” Today’s deposits are surprisingly sticky. Payroll lands every two weeks. Bill pay follows predictable cycles. Moving money between institutions takes enough time and friction that inertia works in banks’ favor.

Programmable digital dollars change that equation. Money that can move 24/7, settle in seconds, and be embedded directly into software becomes easier to redirect.

Treasury systems, AI agents, and business applications won’t care whether value moves via ACH, RTP, FedNow, tokenized deposits, or regulated stablecoins. They’ll choose whatever rail is fastest, cheapest, or built into the workflow. That’s the issue.

Stop managing deposits as if inertia is a strategy. For decades, friction was one of banking's most profitable products. Deposits stayed because moving them wasn't worth the effort. Programmable money eliminates the friction.

Branches Aren’t Dead (Yet)

Conventional wisdom: branches are back, and Chase is proof.

Reality: Chase is proving that Chase can make branches work. Marianne Lake, who ran Chase’s consumer and community bank, said:

“Increasing branches is directly correlated with deposit growth, with new branches accounting for 40% of JPMorgan Chase’s new deposit share gains.”

That stat has been mangled by the trade press and quoted as 40% of its deposits or 40% of its deposit growth. That’s not what she said.

Chase's own data tells a different story.

Back in 2021 it said that half of its new consumer accounts were opened digitally, and that number hasn’t gone backward. But its share of new retail checking accounts fell from 5.2% to 3.7% in a single year.

The buildout is about small business relationships, wealth deepening, and planting a flag in markets Chase didn’t serve: relationship managers on the ground, referrals from accountants and attorneys, advisors chasing affluent households. It’s a real strategy with real ROI—for Chase.

The question for smaller financial institutions: is your distribution strategy focused on winnable customers? Even Chase’s small business bet rests on an assumption that Brex, Ramp, and Mercury are working to invalidate.

The Rise of the Bizumer

The classic W‑2 consumer—one employer, steady paycheck, predictable cash flow—is no longer the dominant growth customer.

Today’s young consumers are Bizumers—half consumer, half small business. Personal and business cash flows blur. Credit needs straddle consumer and business products. The bureau file wasn’t built for these people. Nor are traditional checking accounts.

The Fed pegs roughly 32 million American adults as effectively un-scoreable. For plenty of them, the signal is sitting right there in the file. The models just don’t go looking, because they’ve barely changed in how they read the data they already have.

Fix the model and a big, underserved, entrepreneurial market snaps into focus. SoFi’s push into small business lending is a bet on that market.

Prebuilding the Future

The first half of 2026 was when long-building trends became impossible to ignore. If there’s one word worth using going into 2027 strategy conversations, it’s prebuilding.

Two quotes make the case:

“The risk of under-investing is significantly greater than the risk of over-investing.” (Sundar Pichai, 2024)

“The very worst case would be that we have just prebuilt for a couple of years.” (Mark Zuckerberg, 2025)

Prebuilding means laying track before the train shows up: the data infrastructure, the governance, and the experiments that tell you what’s possible before you commit real money. Five places to start prebuilding:

• AI. Management owns three questions: 1) how does AI change how people work? 2) how does it change who does the work? and 3) How fast can you get—at what rate of change, at what cost, and at what level of operational risk?

• Data. Grade your data quality function by function, put the functional heads on the hook for the score, and get it AI-ready. It’s unglamorous, and it gates everything else on this list.

• Governance. What’s missing: an AI deployment policy that controls how AI systems run in the tech environment, who built them, how they’re tested, and what they do with data.

• Tokenization. Questions to address: 1) what happens ALM models when money can leave the balance sheet 24/7 with minimal friction; 2) what’s the case for stablecoin payments/cash management; and 3) who builds the digital asset stack?

• Products. Banking products are evolving from informational to active and autonomous, fueled by AI agents. Which market segments offer growth opportunities, and what products will meet the needs of those segments?

Get Up Offa That Thing

So where does this midpoint leave you? The winners will react, adjust, and prebuild. The losers will play “wait and see.” James Brown nailed it 50 years ago when he sang “get up offa that thing.” Whatever you do for the rest of 2026, don’t do nothing.

A LinkedIn post from Backbase CEO Jouk Pleiter caught my eye. On top of a large image showing the logos of the largest bank core providers and the largest AI labs was the headline: “The core is set.”

Jouk’s post started off: “Good example of AI THEATRE in banking.” He went on to state his position:

“Model Neutrality. Running the right model for the right job. Mixing open and closed models. Deploying models inside your own governance framework. Improving them over time with your own data, policies and banking expertise.”

My take: Jouk is spot on about model neutrality, but is wrongly accusing the big 3 cores on model lock-in. The more important point: neutrality isn't real unless your vendor supports the model you pick, not just permits it.

What the Big Cores Announced

In a seven-week stretch this spring, the “big 3” core providers put an AI lab’s name in a press release:

• FIS went first. On May 4, FIS announced a partnership with Anthropic and a Financial Crimes AI Agent built on Claude. Anthropic embedded its Applied AI team and FDEs (forward-deployed engineers, duh) inside FIS to co-design it. FIS’s roadmap adds agents for credit decisioning, deposit retention, onboarding, and fraud. The stock jumped 5%.

• Fiserv went next. On May 14 Fiserv announced agentOS and a strategic partnership with OpenAI covering: 1) first-party agents on agentOS; 2)AI applied to core conversions and modernization; 3) banking-specific AI capabilities; and 4) cybersecurity. Four Fiserv agents and nine third-party agents at launch. General availability expected by August 2026.

• Jack Henry went last. On June 25 Jack Henry expanded its Google Cloud relationship, adding Google’s agentic defense products, Mandiant Consulting, and the Gemini Enterprise Agent Platform, to build a proprietary AI security platform for its roughly 7,400 bank and credit union clients.

These Aren’t the Same Deal

Despite some similarities in language, the three deals are very different:

• FIS bet on one model and kept everything else in-house. Claude is described as powering the reasoning across the strategy, not just the inaugural agent, and client data stays inside FIS-controlled infrastructure on FIS’s own data and AI platform. More dependency on one lab, less dependency on others’ plumbing.

• Fiserv went the other way on infrastructure and stayed quiet on the model. OpenAI is co-developing some of the first-party agents. The platform's model layer was named separately: agentOS runs on Amazon Bedrock and Bedrock AgentCore. Less dependency on one lab, more on someone else's plumbing. AWS is arguably the more consequential partner in that announcement.

• Jack Henry barely made a model bet at all. Its announcement is a cloud, security, and enterprise-agent-platform deal, extending a four-year-old relationship, with Mandiant doing a lot of the work. Gemini is in there, but the focus is internal productivity and cyber defense.

The important considerations/clarifications here:

• Neither FIS nor Fiserv disclosed contract terms. FIS made an affirmative claim about breadth. Fiserv didn’t make one either way.

• None of the three announced exclusivity. Claude runs on AWS, Google Cloud, and Azure. Bedrock is multi-model by design.

The Cores’ Case

Two things make these moves compelling for the cores and financial institutions:

• Scarce resources. The scarce input in enterprise AI right now is engineering. Community financial institutions will never work directly with the big AI labs. Through their core, though, they rent a sliver of it.

• Compliance. FIS CEO Stephanie Ferris has argued that FIS’ moat is decades of regulatory and compliance experience, not just technology. She’s got a point.

In creating SR 26-2—which replaces SR 11-7 and the 2021 BSA/AML model risk statement—the regulators explicitly put generative and agentic AI outside its scope. Traditional statistical models, machine learning in credit and fraud, all still in. The newest systems—the ones every one of these announcements is about—out.

Institutions now set the standard for governing agentic AI and supervisors will grade against that standard (because the regulators declined to write one).

For a bank with no AI governance function, buying a governed agent architecture from a vendor that has been examined by the agencies for 50 years is a defensible strategy.

Agents that can write to the ledger need access to the ledger. The cores have it—real capability transfer, governance value, and structural advantage.

The Orchestration Layer is Still Up for Grabs

Ferris said something else worth repeating:

”You’re not going to innovate around us.”

Fiserv makes a same claim, that the defensible business is the orchestration and control layer between frontier AI and the bank.

The problem is that the orchestration layer is exactly what Bedrock AgentCore, Gemini Enterprise Agent Platform, and OpenAI’s agent tools are trying to become.

Fiserv calls its product an “operating system” for agentic AI in banking, but it runs on Amazon’s operating system for agentic AI. It still requires a banking finish-out, but that’s where the value is: identity, regulatory reporting, and an examiner-acceptable audit trail. The hyperscalers won’t build this for the community banking market.

A more defensible version of Ferris’ statement: “you're not going to replicate 50 years of compliance experience overnight.”

What bankers won’t like: Agent workflows built inside agentOS or FIS’s platform make switching cores even more expensive and time-consuming than it already is. Why? Prompt libraries, agent configurations, workflow logic, audit history, etc., aren’t portable. Deconversion of agentic workflows isn’t in today’s contracts (it could/should).

The Cloud is the More Important Layer

The parties worth scrutiny sit one layer below the one that got the headline.

FSOC has flagged that regulators’ authority to supervise third-party service providers varies by agency and asked Congress to fix it. The providers that gap exposes are the hyperscalers.

Two of the big cores named their cloud partners: Fiserv/Amazon, JHA/Google. With FIS, the evidence points to a multi-cloud strategy.

The cores can change labs: Anthropic advertises Claude as the only frontier model available on all three leading clouds, which tells you that model portability exists and that it's the exception rather than the norm.

Changing clouds is a rebuild. That dependency is examinable by no banking regulator, it predates these deals by years, and the announcements deepened it while pointing the press somewhere else.

Anthropic and OpenAI are the only parties in this picture with no contract with the bank, no examination exposure, and no ability to move the roadmap on their own—but they’re the ones whose names run in the headlines.

Neutrality is a Marketing Claim

“Model agnostic” is easy to say but hard to deliver. The key question: does swapping models mean re-validating every workflow? If it does, the claim holds no water.

Smaller (I didn’t say “small”) core and digital providers have an opportunity here: contract terms. A defined deconversion path for agentic workflows, portable agent configurations, exportable audit logs and prompt libraries, bring-your-own-model as a contractual right.

Corelation, CSI, COCC, DCI, and the digital providers that don’t own the ledger can offer all of that tomorrow, and none of it requires knowing Sam or Dario.

Neutrality only counts when it’s written into the contract.

Jouk is right about model neutrality, but the graphic he posted is misleading. "The core is set" points at the model layer, and that’s the easy part to change—no core signed an exclusive deal, and Claude alone runs on all three clouds.

Lock-in is real—but it's in the cloud and the contract.

The Bottom Line

When FIS, Fiserv, and Jack Henry all ship AI agents to their customers, no one gets a competitive advantage from the agents. Everyone gets the same expense and process time reduction on roughly the same timeline.

That’s the historical pattern for everything they offer (it’s a feature, not a bug). Broad distribution of a capability is how a utility works. It’s why buying your core’s agent is a cost decision. Treat it as a strategy decision and you’ll waste 18 to 24 months.

The advantage is in the layer above: the workflows you choose to automate, what your data says that your competitor’s doesn’t, what you do with the capacity you free up. If you haven’t figured that out, it doesn’t matter which AI lab your core aligns with.

Stripe is offering $53 billion to acquire PayPal.

Predictably, much of the analysis has focused on checkout. Stripe is one of the world’s leading merchant payment platforms. PayPal remains one of the strongest consumer payment brands. Combine the two and you get a payments powerhouse.

My take: Banks spent years treating PayPal like a payments competitor. Stripe is valuing it like a deposit franchise. Processing more transactions helps Stripe, but isn’t worth spending $53 billion for. Owning more of the customer relationship—and the financial activity around it—is.

Five numbers explain this deal:

• $1.23 billion: the value of money that doesn’t move. PayPal currently holds ~$38 billion of assets backing customer balances, and, in 2025, generated $1.23 billion in interest income from those balances. Stripe Link is a credential vault—it reduces checkout friction, but the money remains outside Stripe's ecosystem.

• $121: the price per account. PayPal has spent years getting consumers to open accounts, connect funding sources, store credentials, and leave money in. Banks pay hundreds to get a new account and then never do what PayPal has done. $121 gives Stripe room to spend to broaden/deepen its new financial relationships.

• 5 to 7: Stripe’s window. Stripe is renting a mechanism it doesn’t have: it has to pay cash, cash at this size requires debt, and debt at this size requires a sponsor. Enter Advent. The challenge: Advent wants an exit in 5 to 7 years and gets there by cutting costs. Stripe’s plans will conflict. Stripe found the right asset, but did it find the right balance sheet to buy it with?

• $3.7 trillion: the scale opportunity. Stripe+PayPal=$3.7T annual payment volume. Every scale deal claims better risk pricing, smarter routing, and stronger fraud models. This is different: Stripe sees the merchant end, PayPal sees the consumer end. Owning both creates the possibility of underwriting the full commerce relationship instead of just the merchant relationship.

• $60.50: the price that doesn’t close the deal. The market values PayPal as a mature payments company. Stripe is valuing it as a consumer financial platform. Those two views produce different valuations. Stripe can’t argue that PayPal is a slow-growth payments company and pay a payments company price. Stripe’s own thesis is the strongest argument for a higher bid.

Bottom line: Stripe + PayPal ≠ Really Big Stripe. Stripe + PayPal = Different Stripe. With PayPal, Stripe doesn’t just move money—it starts holding it. Holding is a different business with a different regulator and different economics.

Which is why the charter question matters. If Stripe’s strategy is to own the relationship rather than simply process the transaction, it runs into the banking questions: who holds the balances, makes the loans, and controls the economics?

PayPal saw this coming. In December, PayPal applied to the Utah DFI and FDIC for an industrial loan company charter. The proposed bank would support merchant credit, merchant acquiring, interest-bearing FDIC-insured savings, and direct access to card networks. The application remains pending.

The Stripe/Advent bid probably kills it. Put a private equity firm on half of a de novo insured depository institution with $50 billion of acquisition debt above it, and the FDIC’s source-of-strength analysis isn’t a close call.

Stripe’s stack doesn’t fill the gap. Bridge holds an OCC trust charter, and national trust banks don’t take deposits. So it becomes a different company without getting the charter that would make the difference pay. The structure that buys the asset is the structure that caps it.

PayPal holds $38 billion of its customers' money, pays them nothing for it, and just drew a $53 billion offer.

Ain't singin' for FIS
Ain't singin' for Fiserv
I don't sing for nobody
Unless it’s well deserved
This note's for you
—with all apologies to Neil Young

The demo always looks the same. There’s a planner agent at the top, a few solver agents in the middle, a reviewer agent that checks the work, and a manager agent orchestrating the whole thing. It looks like a team.

In fact, it looks like your team—and the org chart they show you makes the AI feel supervised, accountable, and grown-up.

My take: Don’t buy it. New research on how AI agents organize suggests the “agentic team” being sold may be the least efficient way to deploy the technology. The org chart is the sales pitch—but the research says that’s often the worst part of the design.

Teams of Agents Are Often No Better Than a Single Agent

The research compared seven different ways AI agents could be organized—from a single expert, to hierarchies, committees, pipelines, and markets—first in thousands of simulations and then with real AI models performing real tasks.

Findings: Most of the forms that copy a human org chart—pipelines, hierarchies, and committees—underperformed a single agent doing the work by itself. The committee was the standout failure. Why? Because more agents ≠ better thinking.

The Handoff Tax

Why do extra agents cost more than they add?

The answer has a clunky name—contextual transaction cost—but a simple meaning: Every time work moves between agents, context has to be packaged, compressed, and handed off. Every handoff is a chance to drop something that mattered.

A handoff is a tax on the system. When the tax runs higher than what the extra agents contribute, you’ve engineered an expensive way to get a worse answer.

In an earlier post, I made the case that AI’s job is to reduce “coordination”—the handoffs and translation meetings that exist so one function can explain itself to another. A team of AI agents does the opposite: it creates new handoffs between agents, the same overhead AI stripped out when the coordinators were people.

A single agent beats the team on a self-contained job because it doesn’t have to pay a handoff tax. In the research, a single agent posted the highest efficiency in the commercial batch because it skipped the overhead entirely.

You’ve flattened the human org chart. Now someone wants to sell you a new one built of bots.

Five Clones in a Trench Coat

Vendor pitches for AI agent teams imply: more agents=more perspectives=better decisions. It’s supposed to be “collective intelligence.”

Putting five AI agents together doesn't make them smarter any more than hiring five accountants makes your accounting department more intelligent. The value is in how information moves, not the headcount.

If all five agents are built on the same model and fed the same evidence, agreement ≠ independent confirmation. It's just the same opinion repeated five times.

Agent diversity has to be engineered through different models, tools, and sources of evidence.

Dealing With the AI Agent Vendor Pitch

When a core provider or a fintech shows you an “agentic” fraud review, an “AI team” for loan processing, or a “multi-agent workflow” for customer servicing, the org chart language is meant to imply governance.

The AI “org chart” tells you almost nothing about whether it works. Adding more agents rarely fixes the problem. That doesn’t mean you should kick a team-selling vendor out the door. If a vendor pitches you a team, make them show you:

• Diversity. Agents should run on different models, tools, and/or evidence, not five clones echoing each other.

• Evidence mobility. When one agent hands to the next, the facts and their sources move with it, not just a summary.

• Independent verification. The reviewer agent checks against source data, not the same context that produced the answer, otherwise it just inherits the mistake and blesses it.

• A human gate. A person who can see the trail, override the call, and own the result.

If the vendor can show you this, keep talking. If they don’t, keep shopping.

Not Every Vendor Is Selling an Org Chart (Kudos to Fiserv)

I won’t name vendors selling the org chart, but I’ll name one that isn’t.

When Fiserv launched agentOS in May, it called its offering a “control plane” that handles identity, policy, data masking, audit log, kill switch, and orchestration. That’s the layer, not the org chart. The agents that run on top are tasks, not roles.

Example: commercial loan onboarding agent. Getting a commercial loan onto the core is the kind of work a single AI agent shouldn’t swallow whole, and Fiserv’s design doesn’t ask it to. It splits the work of creating the loan package into two stages: 1) checking the documents for problems, and 2) booking the loan to the core.

In stage 1, the agent scores each document, flags what doesn’t match across the loan package, and points to the exact page and clause for every flag. The second stage stays locked until the first one clears. Nothing gets booked until the documents check out.

A banker’s name sits on the file. A human in the loop clicks to generate a discrepancy report before anything moves forward. Visible evidence, nothing booked until it clears, a human name on the result—the design principle suggested by the research.

Can’t I Just Replace the Whole Process With One Agent?

If a single agent often beats a team, why not point one agent at an entire business process? Because the unit of design isn’t the process—it’s the step.

Loan origination, dispute resolution, and account opening all have control points where evidence must be verified, decisions audited, or humans sign off. Those checks aren’t overhead—they’re governance.

They’re the controls the process was built around, and—as your compliance team will tell you—some are required by regulation.

Trying to collapse an entire process into one agent doesn’t eliminate those check points—it hides them inside the model, making them harder to inspect and audit.

The challenge is defining which control points do real work, which are required, and which are just coordination overhead you should eliminate.

Beware: Vendors will pitch the process then show you the task (or step). That’s not a bad thing (I just argued it’s a good thing), but you don’t want to set expectations (with the exec team) that an entire process will be done by a team of AI agents on day one.

Bottom Line

Too many vendors sell a metaphor—the org chart. Don't buy it. Buy the architecture underneath it, and the audit trail that lets a human stand behind the result.

Come into the Fintech Snark Tank and tell me where I’m wrong.

Klarna applied for a bank charter to give Americans a “fairer, more transparent” bank with no hidden fees. 𝗠𝘆 𝘁𝗮𝗸𝗲: Yeah, right. Klarna applied to become a Utah-chartered industrial bank (ILC) for 7 reasons:

1️⃣ 𝗨𝗻𝗰𝗮𝗽𝗽𝗲𝗱 𝗶𝗻𝘁𝗲𝗿𝗰𝗵𝗮𝗻𝗴𝗲. Today, Klarna splits interchange revenue with WebBank. A Klarna-owned ILC, kept under $10 billion, lets Klarna keep the whole thing. That’s recurring, credit-risk-free revenue that scales with card usage, not with lending. Klarna has 700,000 US card customers and a waitlist in the millions (or so they say). The card is the prize, and the charter is how Klarna owns its economics. It also imposes a constraint: Klarna has to keep the US bank’s balance sheet below that $10 billion line to keep the exemption alive.

2️⃣ 𝗦𝘁𝗼𝗰𝗸 𝗿𝗲𝗽𝗿𝗶𝗰𝗶𝗻𝗴. Klarna's stock price has dropped roughly 60% since its IPO. A BNPL shop living on merchant fees and a partner bank trades like a feeble fintech. A deposit-funded, chartered digital bank with recurring interchange and net interest income trades like a durable financial institution. It gives the “we’re a bank, not a BNPL app” pitch teeth.

3️⃣ 𝗖𝘂𝘀𝘁𝗼𝗺𝗲𝗿 𝗿𝗲𝗹𝗮𝘁𝗶𝗼𝗻𝘀𝗵𝗶𝗽𝘀. Pay-in-4 is a commodity. De-commoditizing means converting BNPL users into primary banking customers: checking, debit, savings, card, and P2P. You can’t be someone’s primary provider if you’re renting rails from a partner bank.

4️⃣ 𝗗𝗮𝘁𝗮. As bank of record, Klarna sees transaction and deposit data end-to-end instead of through a partner’s filter. That’s the fuel for its underwriting models and its merchant ad network—a high-margin business that depends on knowing what customers buy and how they pay for it. Owning a bank means owning the first-party data that makes both the ads and the AI defensible.

5️⃣ 𝗗𝗲𝗽𝗼𝘀𝗶𝘁𝘀. Today, Klarna funds US receivables through WebBank, warehouse lines, and capital markets—all of which reprice with rates and can seize up in a stress scenario. Klarna’s model is spread on receivables, and FDIC-insured retail deposits are the best funding available. The further Klarna moves from interest-free Pay-in-4 toward interest-bearing credit (Fair Financing) the more important a cheap deposit base becomes.

6️⃣ 𝗙𝗲𝗱𝗲𝗿𝗮𝗹 𝘀𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗼𝗿𝘆 𝗶𝗻𝗱𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝗰𝗲. A conventional bank charter turns the parent into a bank holding company under the Bank Holding Company Act. The ILC puts Klarna outside Fed supervision and free to run its non-bank businesses: the ad network, the AI commerce platform, the merchant relationships.

7️⃣ 𝗣𝗮𝗿𝘁𝗻𝗲𝗿 𝗶𝗻𝗱𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝗰𝗲. Depending on a single partner is a single point of failure: regulators can throttle Klarna’s roadmap without warning. Bringing the bank in-house reads cleaner to investors and to Klarna’s examiners-to-be.

Bottom line: Put it together and the strategy is clear: Klarna wants the parts of banking that print money—uncapped interchange, cheap deposits, net interest margin, first-party data. The ILC structure lets it keep those while shedding Fed holding company supervision and the revenue split with WebBank.

Everything in the announcement about “reliability” and “hidden fees” is puffery.

The fintechs doing this have figured out what a charter really is: cheap(er) funding, fat interchange, and rich data—all in one wrapper. Klarna wants to own—not rent—that wrapper. The institutions that have held that wrapper for decades should ask themselves how much of that value they’re actually capturing.

SoFi announced it’s entering the small business lending market. It’s offering fixed loans ranging from $2,500 to $250,000, with no application fee, no origination fee, and no prepayment penalties, funded as soon as 24 hours after approval.

According to SoFi CEO Anthony Noto:

“For many of our members, their financial lives do not stop at personal goals, they also include the businesses they are building.”

On the surface, it’s another example of a consumer fintech adding a commercial product. Plenty of firms have tried to bridge those two worlds.

My take: SoFi’s move into SMB lending is recognition that the boundary between consumers and businesses has collapsed. The bigger story is the consumers whose financial lives have outgrown the consumer/business divide.

The Rise of the Bizumer

For decades, banks organized around a simple assumption: consumers are consumers, businesses are businesses, and you staff, build, and price for each one separately. That assumption made sense when most people had one employer and one paycheck.

It makes a lot less sense today.

The traditional credit borrower—W-2 income, stable employer, predictable cash flow—is giving way to a workforce defined by side income, freelancing, and on-demand work. This isn’t a fringe phenomenon. It is a structural shift in how Americans earn.

Block calls them “Modern Earners”—consumers who earn income beyond a traditional paycheck. And there a lot of them. Half of Gen Z and Millennials have a side hustle, half of Gen Z wants to become influencers.

Some make big money. One Gen Z influencer makes $250k a year working 6 hours a week. I hate that guy—I have to work 12 hours a week to make that kind of money.

Young consumers have become “Bizumers”—half consumer, half small business.

They represent a significant underwriting challenge: Their income blends personal and business cash flows, their expenses are partly deductible, and their credit needs span both consumer and commercial products.

They’re developing the financial behavior of an entrepreneur while carrying the credit profile of a consumer. Traditional underwriting models and bureau data weren’t built for that combination.

SoFi’s Bizumer Members

SoFi’s membership reflects that shift more than most financial institutions. SoFi ended 2025 with 13.7 million members, $37.5 billion in deposits, and $3.6 billion in revenue. Last year, it opened more new checking and payment accounts than all credit unions and all community banks in the US.

A recent Cornerstone Advisors study of SoFi members found that:

• 17% own or operate a solo business

• 16% said gig work generates at least a quarter of their income

• 14% said freelance or project work accounts for at least half of their income

That’s 2x to 3x higher than the overall US adult population. The categories overlap, but point to the same conclusion: SoFi has a high concentration of members whose financial lives span both consumer and business banking.

The Credit Gap Bizumers Face

Bizumers run into a wall when they look for credit. Traditional underwriting prioritizes steady, verifiable income, which these folks don’t have. A bank built around W-2 income has no clean way to underwrite them even if it wanted to.

These aren’t unbankable borrowers—they’re misread borrowers. The signal is sitting in the bureau data. Legacy scoring models just weren’t built to find it.

SoFi has something community institutions don’t: an existing transaction history on these bizumers. Nearly half of SoFi members do direct deposit with the fintech, and among that group, 70% have moved deposits, savings, or investment funds over from other institutions, and a third keep half or more of their deposits there.

That’s real income and cash flow moving through SoFi. A bureau file can’t see that. SoFi can and does for a good share of the members its business loan is aimed at.

Banks Built Walls, Consumers Walked Through Them

There's a reason banks built a wall between consumer and business banking: different regulatory requirements, risk models, and sales forces. The org chart isn't crazy—it just maps to a distinction bizumers don't experience, moving between a paycheck and a side business in the same account without noticing the line banks organized around.

Banks drew that line between consumer and business, and the customers living across it never had a reason to see it. SoFi’s “business loan” is a smart recognition of the market it already sits on.

What This Means for Everyone Else

Organize around the consumer/business divide, and you’re optimizing for a world that no longer exists. A couple of credit unions saw this coming:

• Vantage West Credit Union built HUSTL, a digital brand aimed squarely at freelancers and side hustlers, with tools to auto-save for taxes and track gig income across platforms.

• Members Choice Credit Union built a Side Hustle Banking account that gives early access to direct deposits and discounts loan rates for members who also hold a consumer checking account there.

The playbook isn’t complicated, building it is: cash flow tools that make sense of multiple income sources, tax and accounting help built into the account, underwriting that reads gig platform history instead of requiring a W-2, and credit products sized for a business that doesn’t have a corporate entity yet.

Whoever builds that ends up owning the checking account, and the loan follows.

SoFi’s announcement isn’t just about business lending—it’s about the rise of the bizumer, and about how much of the market already lives there. The same opportunity is in every bank's and credit union’s deposit base.

Come into the Fintech Snark Tank and tell me why I’m wrong.

News: Sony Bank got conditional OCC approval for Connectia Trust, a national trust bank that’ll issue a dollar-backed stablecoin. $40 million in capital, targeting a 2027 launch, wholly owned by Sony Bank.

Conventional wisdom: “Sony gets into crypto.”

My take: Sony’s stablecoin is a rail for spending inside its ecosystem: PlayStation, games, anime, subscriptions. It’s an interchange-avoidance strategy: a way to move billions of content spend off the card networks and onto plumbing Sony controls.

You Get a Charter, and You Get a Charter

Here’s what makes the OCC national trust charter the hottest ticket in fintech right now: A national trust bank can custody assets and act as a fiduciary.

What it can’t do is take deposits and it can’t make loans. Because it doesn’t take deposits, it’s generally not FDIC-insured, the capital requirements are lighter, and the chartering process is faster than a full bank charter.

For a stablecoin issuer, that combination is close to ideal: federal supervision and credibility, the ability to custody your own reserves instead of using a third-party custodian, eligibility to apply for a Fed master account, and an exit from the state patchwork of money-transmitter licenses.

In December, the OCC conditionally approved five charters: Circle, Ripple, BitGo, Fidelity Digital Assets, and Paxos. Peter Thiel-backed Augustus landed a full-service version. Wise and Kraken’s Payward subsidiary are in the pipeline.

Not everyone gets the golden ticket: Coinbase and Stripe’s Bridge subsidiary both got left out of the December batch.

Charter Strategies

Everybody keeps putting every applicant into a “stablecoin issuer” bucket when they’re playing four different games:

• Infrastructure incumbents (Circle, Ripple). The firms in this group issue their own flagship tokens. They want the national charter to bring reserve custody in-house, get a shot at a Fed master account, and pre-comply with the GENIUS Act. The charter is infrastructure and regulatory positioning—a better foundation under a product they already sell. No new product required.

• Regulated plumbers (Paxos, BitGo, Fidelity Digital Assets). These players converted from state trust companies and plan to issue stablecoins under the federal charter. They’re the regulated back-end for everyone else’s tokens. The charter makes them a federally supervised counterparty that risk-averse institutions can sign with. Custody-and-issuance-as-a-service is the product.

• Brand builders (Sony). Consumer brands with a huge captive customer base want a token to spend inside their ecosystem. The motivation is control of the payment experience and the interchange they hand to the networks and processors. Sony is the first big consumer brand through the door. It won’t be the last.

• Maximilists (Augustus). Augustus won conditional approval for a full-service national bank—deposits, lending, the works—aimed at AI-powered payments and stablecoin settlement.

They’re sorted by two questions: 1) who’s the customer? and 2) what does the charter let them do?

Sony in the Middle, Banks at the Ends

Sony’s model previews what happens when large consumer brands decide they’d rather run their own payment rails than keep paying yours.

Every retailer, platform, and marketplace with a big enough customer base and enough transaction volume can now look at the trust-charter path and run the math on cutting the card networks out.

When they do, the spending that used to cross your rails—along with the interchange and the deposit balances riding on it—starts flowing somewhere you don’t touch. That’s the deposit-displacement story wearing a Sony logo.

Displacement isn’t destiny, though, and the opportunity is more specific than “partner with fintechs.”

Picture the payment in three pieces: 1) where the money comes from; 2) the moment it changes hands; and 3) where it lands. Sony’s token is built to own that middle piece—the purchase itself and the interchange that used to ride on it.

That’s the part banks can’t win back, because a closed loop the brand controls will beat your card every time. The money banks can still hold sits at the two ends:

• Front-end. A Sony wallet is built for spending; paychecks still land somewhere else, and money still rests somewhere else. Every dollar that funds that wallet passes through the account it lands in first. As long as banks own the funding instrument they own a data and relationship footprint Sony never sees.

• Back-end. Every stablecoin dollar Sony pays out—to the studios, the indie developers, the creators earning inside its ecosystem—has to turn back into real dollars in a real bank account. Those are small businesses that need operating accounts, cash management, and credit.

There’s a third play: being the regulated partner for the brands that won’t charter themselves. Most retailers and platforms won’t spend two years and eight figures in front of the OCC the way Sony did—they’ll want a bank to hold reserves, run the fiat on- and off-ramps, or sponsor issuance.

That’s real fee income for an institution with the compliance muscle to do it right. Holding the reserves behind a major issuer at scale, though, is a money-center game.

None of this is about the token itself. The plays are the same ones that have always separated the banks that keep their customers from the ones that rent them: own where the money comes from, and bank where it has to turn back into dollars. The brands can have the middle; banks need to hold on to the ends.

Bottom line: don’t read charter approvals as “another crypto company gets blessed.” Ask: 1) who’s the customer? and 2) what does the charter let them do? Most are infrastructure plays by firms most banks and credit unions won’t compete with directly. The Sony story is the one that reshapes who owns the payment rail and who holds the deposit behind it.

The Wall Street Journal reported that JPMorgan, Bank of America, Wells Fargo, and PNC have all held conversations about buying a debit network from Fiserv. None of them really wants to be in the network business—what they want is a way to get around the Durbin Amendment’s cap on debit interchange.

My take: Big banks are circling Fiserv’s debit networks to avoid a fee cap they’ve spent 15 years fighting. Capital One established a regulatory precedent. For everyone else, the math…doesn’t math.

Why Fiserv, Why Now?

Fiserv owns STAR and Accel, two of the legacy PIN-debit networks. It’s also a company sharpening its focus.

A legacy PIN-debit network that has lost ground to Visa for years is a logical candidate when a company is narrowing to core banking and merchant acceptance. For a megabank, the network’s appeal is largely regulatory: the Durbin exemption that comes with owning the rails, more than the processing volume itself.

If It Worked For Capital One…

When Capital One acquired Discover, it combined the card issuer and the network operator under one roof. Under the Federal Reserve's interpretation of the Durbin Amendment, the interchange cap applies when a network sets interchange to compensate a separate issuer.

Fold the issuer and the network into a single entity and that cap no longer applies. Capital One has said that migrating its debit portfolio onto Discover’s network could generate more than $1 billion annually in additional interchange revenue.

The large banks ran the same arithmetic. That’s why they’re shopping. But this is where the Capital One comparison starts to break down.

Discover controls the brand on the front of the card. Capital One can move its debit portfolio onto Discover’s network and capture the economics end-to-end.

STAR and Accel, by contrast, are back-of-card PIN-debit networks. Under Durbin's routing rules, the merchant—not the issuing bank—chooses which eligible network processes a debit transaction, typically based on the economics, processor defaults, fraud performance, and other operational considerations.

Two questions to address:

1. Legal: would a bank-owned PIN network qualify for the same regulatory treatment that benefited Capital One?

2. Economic: does the bank has enough market power to force merchants to route transactions over the network(s)?

According to Andrew Dresner, author of the Payments Full blog, JPMorgan and Bank of America each control something like 15% of national debit volume (even more in some regions). A merchant can decline Capital One–Discover economics and barely notice. Declining Chase’s debit base is a different decision.

According to the WSJ article, several of the banks that looked have already decided to pass. The math just doesn’t math.

What Durbin Taketh Away, It Giveth Back

The richest part is that STAR and Accel exist in their current form largely because of Durbin. The amendment’s routing mandate forced every debit card to carry two unaffiliated networks, which is the only reason these regional PIN networks stayed relevant after Visa and Mastercard consolidated the front of the card.

Durbin kept the patient breathing. Now the big banks want to buy the patient to escape Durbin.

The larger problem with spending billions to escape the debit cap is that nobody knows whether the cap survives. Two forces are pulling in opposite directions:

• The Fed’s 2023 proposal would cut the base rate from 21 cents to 14.4 cents which makes escaping the cap more valuable and sharpens banks’ incentive to own rails.

• In August 2025, a federal court in North Dakota vacated Regulation II's fee standard, finding the Fed had overstepped its authority. The ruling is stayed pending the Fed's appeal to the Eighth Circuit, and a Kentucky court reached the opposite conclusion weeks later, so nothing's settled.

The direction is what matters: the ND plaintiff is a merchant who won by arguing the cap is too high, so if it holds, the Fed has to reset the cap lower rather than scrap it. Either way the incentive to own rails that sit outside the cap only grows.

Son of Zelle

Fiserv’s networks route debit for thousands of community banks and credit unions. If STAR and Accel end up owned by a megabank—or a consortium of them—a lot of smaller institutions will end up sending their debit traffic across pipes owned by their biggest competitors.

The Bigger Prize

The bigger story is about who controls the transaction layer as payments fragment. The same banks exploring this deal are building tokenized-deposit systems and eyeing stablecoin rails.

Owning a network is a way to own the pipe as commerce moves toward agentic checkout and programmable money. Durbin arbitrage is the near-term justification—control of the rails is the actual ambition.

Banks spun Visa and Mastercard out of their own consortia because the network layer looked like a low-margin utility. Fifteen years later they want back in, because that’s where the value—and the regulatory optionality—migrated. Everything old is new again, this time with a Durbin wrapper.

The Capital One playbook worked because Discover was a closed-loop signature network and Cap One owned the whole stack. A back-of-card PIN network, run over rails the merchant gets to route around, is a thinner edge. The banks that walked away did the math. Those still at the table are considering whether they have enough debit share to make merchants accept a network they'd prefer to not use.

This Deal Ain’t Happening

I don’t see JPMC (or BofA) acquiring either/both of Fiserv’s networks for 3 reasons:

• Fiserv won’t sell them. The debit networks are highly profitable and (can/could) play a key role in improving the value that Fiserv provides to its bank clients. Selling the networks would tick off a lot of financial institutions and I can’t see Fiserv putting itself in that position right now.

• Merchants will play chicken. The logic of the deal (from the megabank perspective) is that the bank’s debit volume would give it market power and merchants would have to kowtow. I don’t see Walmart and Target backing off so quickly. They’ll play chicken—decline Chase debit transactions (for a period of time), betting that consumers will use alternative forms of payment.

• The Visa angle. Visa bundles credit and debit economics, and the DOJ's monopolization suit alleges Visa has used credit-side discounts to defend debit routing. Chase is an enormous Visa credit issuer, so a debit move that trips a volume commitment could put credit incentives—far larger dollars than debit interchange—in play. The escape doesn't just cost Chase some debit incentive income. It puts the economics of the portfolio Chase cares about most at risk.

And if you need a 4th reason, I don’t think this gets regulatory approval.

Come into the Fintech Snark Tank and tell me why I’m wrong.

Plaid launched a “sequential” foundation model that reads financial behavior over time instead of transaction by transaction. It’s the latest piece of the “intelligence layer” the company has been assembling.

A conventional risk model reads a consumer’s financial life the way Vonnegut wrote Slaughterhouse Five: unstuck in time, events pulled out of sequence, no narrative. They collapse transaction history into summary statistics like rolling averages, fee-to-deposit ratios, and days since last cash advance. That’s useful, but the model strips out the order things happened in, which is where the story lives.

In a proof-of-concept measured against LendScore, Plaid's own cash flow credit score, the sequential model reduced default risk by 13.6% at a hypothetical 70% approval rate. In early pre-production testing on ACH payment risk, it flagged 26.5% more of the dollars headed for returns without flagging more transactions.

Plaid spent a decade building the rails that let consumers route around the credit bureaus. Then, to sell credit intelligence, it registered as one. Plaid Check is a consumer reporting agency under the same law as Equifax and Experian. The company’s latest announcement makes the most sense looking through that lens.

The Glide Path and the Rocky Road

The sequential model will fold into Signal at no extra cost—it’s in testing now, with production rollout in the coming months. Same per-transaction pricing, better model, more volume through a pipe customers already run.

The payoff is a stickier Signal and an easier pitch to fintechs not yet using it. The path into the bank market is smoother here too, because payments risk is a “lightly” regulated area: declining to front a risky transfer doesn’t trigger the adverse-action and fair-lending requirements that apply to credit decisions. (If the scores are used to decide who gets an account, FCRA kicks in).

The bigger—and harder—move is selling the new model into credit decisioning.

How a Sequential Model Improves Credit Decisioning

The best conventional risk model can make bad lending decisions. Plaid's illustration: two applicants with near-identical 90-day summaries, same income, average balance, rent, overdraft count, spending level, and risk score, but:

• Applicant A receives his paycheck, pays his bills (which clear), discretionary spending follows, one overdraft shows up after an unusual expense, and the account recovers on the next paycheck.

• Applicant B receives his paycheck and, within 24 hours, a short-term loan repayment and a credit card payment drain the account. Small transfers bridge the gap. An ACH debit hits before the next payroll. Overdraft fee posts. Repeats two weeks later.

One consumer absorbed a shock and recovered. The other runs a recurring liquidity deficit that looks stable in the aggregate but isn’t. The ledger’s the same, but the story is different. A conventional model approves both—a sequential model doesn’t.

The Questions Bank Credit Risk Committees Will Ask

Every credit risk committee that hears “13.6% reduction in default risk” will ask the same five questions, so I put them to Plaid:

1. Compared to what baseline? LendScore, Plaid’s own cash flow score, with a baseline default rate of 13.3% (measured as 90+ days delinquent within 12 months). A 13.6% relative reduction takes modeled defaults to roughly 11.5%.

2. On what loan population? A generic sample—primarily credit cards, installment loans, and auto loans. Not any bank’s own portfolio.

3. Is 70% our approval rate or a hypothetical? A hypothetical. Plaid also frames the result the other way: hold losses constant and approval rates can rise 25%—a view lenders will find more interesting.

4. How does it interact with bureau data and internal scorecards? It doesn’t. The test ran in isolation.

5. Has it cleared fair lending analysis and model validation? Not yet. Plaid says it excluded transactions that correlate with protected classes when building the proof-of-concept, and that production deployment—planned for LendScore this fall—goes through full fair-lending review and model validation first.

Two of those answers deserve a closer look:

1) The isolation result cuts both ways. A 13.6% improvement produced without the bureau data and internal scorecards a bank underwrites with doesn’t tell that bank how the model performs inside its real decisioning stack. The lift could shrink when bureau data absorbs some of the signal or grow where bureau files are thinnest.

2) The fair-lending answer is the harder one. What people buy, where, and when—the details that make a sequential model powerful—can create proxy variables for protected characteristics.

Plaid has built mitigations into LendScore: feature-level exclusion of variables that correlate with protected classes, five SHAP-based reason codes delivered with every score for ECOA adverse-action compliance, and a third-party fairness audit, which found LendScore’s features relatively uncorrelated with protected class attributes.

That’s more compliance architecture than most cash flow scores ship with. But excluding the obvious proxies isn’t the same as proving no disparate impact which is why the model still has to clear fair-lending review before it decides who gets a loan.

And putting a new model into the underwriting stack runs into ECOA, adverse-action rules, and—depending on the charter—OCC, Fed, or FDIC model-risk expectations.

Plaid’s Rocky Path Up the Credit Decisioning Stack

From Plaid’s perspective, the credit decisioning market is comprised of 3 segments:

1. Fintechs. Includes near-prime installment shops (OppFi, Oportun), cash advance and earned wage apps (Dave, MoneyLion, EarnIn), and BNPL players (Affirm, Klarna, Afterpay). Data science in-house, and no examiner requiring model risk validation (even though many do it anyway). Many already use LendScore, which the sequential model will feed in production this fall.

2. Banks. Most banks access Plaid-derived cash flow intel indirectly, embedded in decisioning tools they already use like Experian's cash flow products, UltraFICO, or Prism Data's CashScore. They’ll turn on a cash flow attribute inside FICO or Experian as a second look on borrowers they were about to decline.

3. Credit unions. A greenfield opportunity. The mission fit is real: cash flow underwriting is how credit unions say yes to young thin-file consumers that bureau data can't score. The data science bench isn’t there, so they adopt via Zest AI or Scienaptic, not a raw Plaid API. For Plaid, credit unions are a channel play.

What all three segments have in common: the path runs through FICO and VantageScore. They are the consumer scoring market, there's no third name. Between them, they sit under most mortgage, card, and auto decisions in the country.

This market moves at a geological pace.

The models mortgage underwriting runs on—“Classic FICO” versions 2, 4, and 5—predate the financial crisis. Fannie and Freddie only started accepting VantageScore 4.0 this spring—nearly 9 years after the score launched and 8 years after Congress mandated credit score competition. VantageScore has matched or beaten FICO on predictiveness for much of that time, and still fights for share.

Plaid enters from an unusual spot: competitor and partner. FICO partnered with Plaid to feed real-time cash flow data into the UltraFICO score. Plaid’s own guidance tells lenders cash flow data should complement traditional scores, not replace them.

Plaid distributes its scores through its consumer reporting agency, Plaid Check, which makes it a supplier of inputs to decisions others make, not the decision-maker (for now, at least).

This points to what Plaid is really selling (for now): the best input into the decision, rather than the decision itself. Cash flow intelligence that feeds the score, not the score that replaces FICO.

The LendScore groundwork is real, but to move past best input, Plaid needs three things at production scale it doesn’t have yet:

1. independent validation across diverse loan populations;

2. a regulatory track record built in production, not a proof-of-concept; and

3. lender adoption at enough scale to build the feedback loop.

The Bigger Picture in the Credit Market

What’s really going in the credit market? Three things:

• The summarization era is ending. For decades, risk intelligence meant collapsing transaction history into interpretable, adverse-action-friendly statistics. That era ends fastest where governance is lightest—payments risk, cash advance scoring, fintech underwriting—and slowest in the regulated core. Plaid leaning on “explainable” in its own marketing tells you it knows where the friction sits.

• The risk-model gap is widening. Banks building models can only tap into their own portfolio history. Plaid trains across thousands of institutions and half the country’s banked adults (its claim)—a spread of behavior across income levels, geographies, and account types no single institution can replicate. Better feature engineering doesn’t close that. It’s a different class of problem.

• Network data is becoming the moat. APIs and UX are commodities and anyone can copy the architecture. The data—and what you learn from it at scale—is the differentiator. Plaid isn’t the only one who’s figured this out. Mastercard, Visa, Stripe, Nubank, and Revolut are all building their own versions.

Open banking was supposed to commoditize access to the network. The 1033 rule is being rewritten, and banks are already moving to charge for the access the original rule made free. But access was never the moat.

The moat is the labeled outcomes—who defaulted, which transfers returned—accumulated across the network. Hand a competitor full open-banking access tomorrow and it still starts at zero on outcomes, with no scale.

Access can be regulated into a commodity—a decade of labeled behavior can’t.

Owning the richest training data—and the outcome labels to go with it—goes a long way to winning the intelligence layer. It’s an advantage that API connectivity doesn’t deliver. But it doesn’t take anyone all the way there—just ask VantageScore.

What Plaid is Selling

What Plaid's selling, then, is two things: a payments-risk upgrade it can book in the short term, and a claim on the underwriting stack it'll take years to earn.

Both are real and make sense once you recall where this post started. The company that built its business as the alternative to the credit bureaus is now one of them.

Come into the Fintech Snark Tank and tell me where I’m wrong.

The value a financial product delivers isn't what it advertises—it's what the consumer actually captures. They’re different numbers, and the gap between them is an under-examined problem in retail banking product design.

That’s my take after reading a couple of studies from Alexander Nayvelt, founder of WalletSavvy: The Best Card Paradox: Why the Card You Choose Is Almost Never the Card You Use and The Annual Fee Inversion: Why a $695 Credit Card Can Cost Less than a Free One, and Why the Market Cannot Tell You in Advance (yeah, that’s a long one).

Together they make an argument that goes to the heart of how financial institutions should think about product design and recommendations.

The argument starts with a paradox. The Amex Platinum card has a $695 annual fee. Its advertised benefits—airline credits, hotel credits, lounge access, Uber credits, etc.—totals more than $1,500 at face value. For a consumer who captures the full set of benefits, the card is a great product. For the cardholder who realizes just $200 worth of benefits in a year, it’s a $495 mistake.

Same card, same price, two different economic outcomes.

Nothing in the market—comparison sites, issuers’ disclosures, not even a cardholders own self-assessments—reliably sorts consumers into the right group before they commit.

Nayvelt calls this the Annual Fee Inversion: the underutilizing majority subsidizes the high-realizing minority. By Nayvelt’s reading of Bankrate, JD Power, and CFPB survey data, 55% to 70% of premium cardholders fail to recoup their annual fee each year.

The mechanism is identical to what one study documented in gym memberships 20 years ago: the upfront commitment is paid by everyone, realized use is skewed below breakeven for most people, and the firm’s profitability depends on keeping it that way.

Recommendation Platforms Optimize the Wrong Variable

Nayvelt frames consumer value capture as the product of two distinct stages: selection (which card you choose) and realization (how much of its value you actually capture).

He formalizes this as V = S × R, where S is the theoretical value of the card to your spending profile and R is a realization coefficient reflecting how much of that theoretical value you’ll actually capture.

Recommendation platforms—NerdWallet, CardRatings, The Points Guy—have spent decades getting good at estimating S. Spending profile inputs, demographic variables, stated preferences, attribute-weighted scoring.

Nobody has built the ability to measure R, the realization coefficient, however. This score is the product of three factors:

• Attention: how many benefits a cardholder is aware of when action is required. Low attention reduces R regardless of how well the card fits a cardholder’s spend.

• Activation: how many benefits are claimed. A cardholder who knows a bonus exists but doesn’t use it is no different than a cardholder with no access to it.

• Execution: how well spending matches what the benefit requires. A 5% category bonus capped at $1,500/quarter is useless if a cardholder’s category spend is $400.

Why doesn’t the recommendation industry address this? Because they’re not stupid. Comparison platforms get paid per-acquisition by issuers. Steering someone from a $695-fee card to a no-fee alternative forfeits the high-bounty commission.

Issuers, for their part, have every reason to leave the realization gap exactly where it is.

This Is a Widespread Banking Challenge

The cross-subsidy structure Nayvelt describes isn’t unique to premium credit cards. It shows up anywhere a consumer pays for a product whose value depends on future behavior they’re likely to overestimate. That fits a lot of what retail banking sells:

• Rewards checking accounts. The headline earn rate requires minimum monthly debit transactions, direct deposit, and a minimum balance. Most accountholders qualify some months and miss the threshold others. The bank prices the account assuming a realization distribution. Consumers assume they’ll always qualify.

• CD specials and promotional rates. Consumers who rollover a CD at maturity and capture the promotional rate each cycle are different than those that let it sit at the default renewal rate. Same product, different realized economics.

• Relationship bundles and fee waivers. The realized fee depends entirely on consumers’ follow-through, yet many consumers still end up paying fees they could have avoided.

• Loan-linked benefits. Examples: rate discount for autopay enrollment or holding a qualifying deposit account. The discount is real if claimed. It requires action the consumer intended to take at origination, but often doesn’t.

None of this is predatory in the legal sense. Terms are disclosed and the math works out as advertised for consumers who execute. But the product economics depend—in ways that banks understand but most consumers don’t—on a realization distribution that falls below consumers’ self-forecast.

The Opportunity for Banks and Credit Unions

Nayvelt’s proposed solution: a recommendation engine that estimates consumers’ realization coefficient based on behavioral data—transactions, redemption patterns, category spending distribution—rather than from stated preferences.

The CFPB’s Section 1033 personal financial data rights rule provides the legal foundation for consumer-authorized data sharing that would make this possible. For banks and credit unions, however, Section 1033 cuts both ways:

• As a data source, it creates the technical possibility of actually knowing how a consumer uses financial products before recommending new ones.

• As a competitive exposure, it means fintechs with a better recommendation infrastructure can use a consumer’s transaction history—from accounts held at the institution—to make a more compelling case for switching.

Nayvelt is careful to note that issuers aren’t being dishonest. The contractual terms are accurate, the fee is disclosed, the benefits are real. The gap between marketed value and realized value is a structural feature of how the product is designed and how consumer psychology works, not a disclosure failure.

A common complaint (from consumer advocates and politicians) is that the product economics are built around a realization gap that consumers can’t observe and incorrectly forecast. Wrong.

For banks and credit unions, there's a real strategic choice embedded in Nayvelt's framework.

The realization gap isn't just a consumer problem—it's a design decision, and institutions have historically benefited from it. Section 1033 changes the competitive math: fintechs can now use a consumer's own transaction history to recommend a better-fitting product from somewhere else.

The institutions that build the realization infrastructure first earn something that's genuinely hard to replicate—the customer's trust in the recommendation itself.

Building a “Realization Infrastructure”

A realization infrastructure is the combination of data, modeling, and customer communication that closes the gap between what a product offers and what a customer will actually capture. It has three components:

• Behavioral baseline. A pre-recommendation read of how a customer uses a product, built from transaction history rather than stated preferences. What categories does spending concentrate in? Which existing benefits does the customer redeem consistently? Where does the customer’s behavior suggest they’ll miss thresholds or fail to activate? This is where Section 1033 becomes significant: it makes consumer-authorized transaction data portable, which means an institution can potentially build this baseline even before a customer holds the product in question.

• Matching model. Most recommendation logic today answers the question “which card maximizes theoretical value for this spending profile?” Realization infrastructure answers: “which card maximizes expected realized value given what we know about this customer’s redemption behavior?” This produces different recommendations for a meaningful share of customers.

• Post-acquisition monitoring. This is the infrastructure that surfaces missed value in real time. Benefit utilization alerts, threshold warnings before quarterly caps reset, proactive reminders when a category bonus is underperforming. This isn’t new as a concept; banks have the transaction data to do it today. What’s missing is the organizational commitment to treat realization as a product metric worth tracking, not just an outcome customers are responsible for on their own.

Institutions that build this aren’t just improving product recommendations—they’re changing the implicit contract with the customer. Instead of “we disclosed the terms,” the value proposition becomes “we’ll help you capture what you signed up for.” That’s a durable competitive position that comparison platforms, by design, can’t replicate.

Come into the Fintech Snark Tank and let me know where I’m wrong.

Google launched a standalone Android app for Google Finance last week. The app delivers watchlists, real-time market data, a live news feed, and an AI feature called “Key Moments” that explains why a stock moved.

Portfolio tracking, advanced charting, live earnings calls, and AI-powered scheduled briefings are live on the web, and will be coming to mobile “in coming months.” iOS is coming “later this year.”

My first take: Oh geez, what is this, like the 734th time Google has relaunched its finance app?

My considered take: Three structural shifts at Google since 2021 (when Google killed Google Plex) create a different environment for Google Finance to survive—and potentially thrive—in ways that no previous Google financial product ever had.

Why Google Kept Killing Its Financial Products

Google’s “finance” history:

• 2006: Google Finance launches with interactive charts, portfolio tracking, and historical data.

• 2008: An update adds real-time prices from NYSE and NASDAQ.

• 2012: Google closes the Google Finance Blog after posting stopped in 2011.

• 2017: The “redesign” strips portfolios, historical data tables, advanced charting, and sector summaries. Users are furious. Google’s explanation: we’re making the product “more accessible.”

• 2019: Google announces Plex, a checking account embedded in Google Pay, co-branded with bank and credit union partners. A 400,000-person waitlist builds.

• 2021: Google kills Plex before it launches, notifying regulators before notifying its own bank partners.

Google has struggled with “Finance” because of it’s ad revenue-dominated culture.

The original Google Finance didn’t generate ad revenue and neither did Plex. Products that didn’t drive ad revenue were starved of resources, lost their executive champions, and eventually got killed.

The question for Google Finance version 734: Has anything at Google changed? The answer is yes.

What’s Changed at Google

Google 2026 is different from Google 2021 in three meaningful ways:

• Revenue has diversified. When Plex died, ad revenue was the gravitational center around which every product decision orbited. That’s no longer true. Google Cloud hit $20 billion in Q1 2026, up 63% year-over-year. Pichai told investors that revenue from gen AI models grew nearly 800% year-over-year. Alphabet now has 14 product lines each generating over $1 billion annually. For the first time, a product that doesn’t drive ad clicks has somewhere to live organizationally.

• The antitrust ruling changed the economics of user acquisition. Google can no longer pay its way to user relationships—it has to earn them product by product. A federal judge ruled that Google illegally maintained its search monopoly through exclusive default agreements with Apple, Samsung, and others. A financial app that users open daily is the kind of direct relationship Google needs now. Google Finance is infrastructure for Google’s post-antitrust user acquisition strategy.

• Gemini changed the organizational math for Google Finance. Pichai called scaling Gemini on the consumer side Google’s “biggest focus.” Every core feature of Google Finance—Key Moments, the AI research tool, scheduled briefings—is a Gemini showcase. Plex was built on Google Pay, a product already losing leadership support when Plex launched. Google Finance is tied to the product Pichai is most publicly committed to. That’s organizational protection.

What’s New in Google Finance

The “new” product’s features are worth taking seriously:

• Key Moments gives users an explanation for why a stock moved—earnings miss, guidance cut, macro event—without requiring them to scan headlines.

• Portfolio tracking is back with a better import experience: users can upload CSVs, PDFs, screenshots, or just describe their holdings and let AI go from there.

• AI research tool answers questions about a user’s portfolio—e.g., sector gaps, allocation impact on long-term growth, etc.

• Scheduled briefings let users define recurring AI research tasks and receive them on a custom schedule.

Google Finance still doesn’t offer trading. It’s a research and information layer, not a brokerage—a deliberate positioning given the Plex history.

The Call

I called Plex a winner because the product concept was genuinely good. aBut concept quality was never a problem at Google. The problem was whether the company’s culture would sustain a financial product that didn’t fit the advertising model.

The three shifts described above—revenue diversification, loss of default search distribution, Gemini as the strategic flagship—don’t guarantee success. But, for the first time since 2006, there are structural reasons to believe a Google financial product has durable organizational support.

I’m cautiously optimistic.

For banks and credit unions, the question is one I’ve been raising for years: who owns the financial context layer for your customer? A consumer who tracks their portfolio, gets daily AI market briefings, and researches stocks inside Google’s ecosystem has established a financial information habit that no bank app currently competes for.

Financial institutions have spent the past decade worrying about who owns the payment. The next 5 years will be a battle for who owns the financial context.

Come into the Fintech Snark Tank and tell me why I’m wrong (again).

Backbase bought Kasisto on June 23 and folded its team, its agentic platform, and its “financial services intelligence” into the Banking OS.

The most informative line in the Backbase-Kasisto announcement, however, is the one that isn’t there: the price. When a deal is a win, the acquired can’t wait to tell us the multiple. When the number’s “undisclosed,” it means the math ain’t mathing. No mention on what changed hands in this deal.

Let’s look at what each side brings to the altar.

Kasisto: $90 Million In, Not Much Out

Kasisto spun out of SRI International in 2013—the same lab that gave the world Siri. Over the next decade it raised roughly $90 million across nine rounds from a who’s who of investors: Oak HC/FT, Westpac, FIS, and others. The KAI platform landed marquee logos—J.P. Morgan, Standard Chartered, TD, Westpac.

Independent estimates put Kasisto’s 2023 revenue around $10 million. If that estimate is anywhere in the ballpark, it’s a red flag signaling a company not on the right track. Another warning sign: headcount declined from about 80 employees in 2023 to ~50 by early 2026.

The leadership carousel is another signal. Co-founder Zor Gorelov stepped out of the CEO seat in 2024 and, by April 2026, was running a different AI startup entirely. The firm’s former Chief Product & Marketing Officer left in 2024 to run a digital banking shop, about the time the Gorelov handoff was happening. A Managing Director of North America lasted 11 months between mid-2022 and mid-2023.

In October 2025, Kasisto relaunched as “agentic”—KAIops, agents, near-zero-incident operations. A conversational-AI vendor discovering it’s an agentic-AI vendor at the moment the market stopped paying for chatbots tells you this was a survival strategy.

Backbase: Bootstrapped and Backed Into a Buy

Backbase had an impressive run, growing to more than $200 million in revenue before raising outside capital. In 2022, Motive Partners put in €120 million at a €2.5 billion valuation, Backbase’s first external money.

At the time of the raise, TechCrunch reported that the company “will not be using the funding for acquisitions—integrating is just much more painful than building in an integrated way from the ground up.” Two acquisitions later—Nucoro in 2023, Kasisto now—and it’s clear that plan didn’t survive its first brush with private equity.

Two months before buying Kasisto, Backbase launched its “AI-native Banking OS” with layers it named Intelligence, Nexus, and Sentinel—the last an authority layer built to check every action against bank policy. Governance was on the whiteboard. Launching that OS and then buying an AI company eight weeks later still suggests it needed more than it shipped with.

The harder requirement sits downstream: use cases governed and documented well enough to clear a bank’s risk and procurement review in a live deployment, not on an architecture diagram.

What Changed? Chatbots Got Commoditized

Both companies sell the same thing under the agentic label: software that mediates the conversation between a bank and its customers. The challenge: large language models commoditized the interface. What Kasisto built is now just a model-access-and-compliance-review away.

What isn’t commoditized is everything behind it—integration with core systems, controls, governance. That’s where the value moved, and where deployments stall.

The repositioning follows the commoditization. Chatbots that answered became agents that resolve, and Backbase’s “conversation to resolution” framing is the whole thesis: the AI won’t just surface a balance, it’ll execute the work across the back office.

That may prove out. But the resolution claim isn’t new, and the distance between an agentic demo and agentic production in a regulated environment is where a lot of community bank and credit union AI budgets stall.

Agentic AI Creates Governance Challenges

The distinction between genuinely agentic and conversational AI under a new banner creates governance challenges.

Agentic AI sets goals dynamically, decides on actions from context rather than rules, handles cases outside its training, and orchestrates tools on its own.

Unfortunately, a lot of what’s sold as agentic today is closer to intelligent RPA: document classification, data extraction, rule-based decisioning, straight-through processing with an LLM in the loop. It’s been in production for years, and governed very differently. Calling those workflows “agents” is marketing.

Consumer appetite isn’t the constraint. A 2025 Commonwealth survey found 79% of chatbot users trusted the information their bank’s bot gave them. The constraint sits upstream: can a bank explain the thing to the regulators.

Take fee reversals. An agent that waives an overdraft charge on its own is a clean demo. In production, the bank still has to set the agent's authority limit, show it treats similar customers consistently to stay clear of UDAAP exposure, and produce an audit trail for why this fee and not that one. The demo is the easy part.

Backbase has an answer to this—on paper. Its authority layer, Sentinel, checks every action against bank policy and logs it. Backbase pitches it as enterprise AI a regulator can trust. The open question is whether that holds in a live deployment.

The harder vendors lean on “agentic,” the more governance questions it invites, and the further its outcome-focused messaging sits from the answers. Vendor pitches say:

• The agent completes end-to-end workflows with little human effort. Examiners, however, will ask where human review happens and how the bank shows humans are meaningfully in the loop rather than rubber-stamping.

• Decisions run dramatically faster. Examiners ask what data the model uses and how it’s validated.

This Deal is a Template

The conversational-AI vendors that raised through the last cycle are sitting on commoditized products and thin revenue, and the platform companies need an AI story to carry into renewals.

Expect more conversational AI providers absorbed by the banking platforms, rebranded agentic on the way in, at prices that stay undisclosed for the reason this deal did.

Alkami spent $135 million on Segmint and $400 million on MANTL to assemble its sales-and-service stack, and a conversational layer is an obvious next piece. Q2 sits in the same position. Candescent partnered with Kasisto and now needs a replacement conversational layer.

One tier up are the cores—Jack Henry, Fiserv, and FIS, the last of which was a Kasisto investor and partner and just watched the asset get bought out from under it.

Three consolidation candidates:

• Posh sells website and phone bots to credit unions and community banks and sits in almost exactly the size and customer band Kasisto did, which puts a Q2 or Alkami tuck-in at the top of the watch list.

• Interface.ai took a $30 million round from Avataar in 2024 and calls itself the most valuable AI company in banking. It’s the scaled independent in the category—an eventual prize for a core or platform, not a quick pickup.

• Eltropy is already playing consolidator, having absorbed Lexop and POPi/o. It and interface.ai planted competing “industry-first” agentic flags within weeks of each other in late 2025, which signals both intend to do the buying.

Buying and Selling an Adjective

“Agentic” is to 2026 what “digital” was to 2015 and “omnichannel” was to…whenever, before that.

Vendors sell it, banks think they want it, and technology analysts forecast it without anyone agreeing on what it is. The Backbase-Kasisto deal runs on that ambiguity—conversational AI folded into a platform and relabeled a category, because the category is what the market is paying for this quarter.

As my colleague, Cornerstone Advisors Managing Director Chris Miller, said:

“Backbase isn’t buying a buzzword; it’s buying pre-fabricated AI use cases that can survive procurement, risk, and integration without setting off governance alarms, and that’s the part no LLM API has commoditized yet.”

The pattern repeats every few years. The industry picks a word that names an outcome instead of a mechanism, prices it, and trades on it until the next word shows up. “Digital transformation” sold a decade of platform deals before anyone defined what got transformed. “Agentic” is further along that curve than its users admit.

Come into the Fintech Snark Tank and tell me where I’m wrong. :)

A recent HBS/Wharton field experiment put 776 Procter & Gamble employees—split randomly into four groups: one person teams with no AI, one person teams with AI, two-person teams with no AI, and two-person teams with AI—through real product development work.

The AI was OpenAI’s GPT-4 (GPT-4o in the later sessions). Everyone in the AI groups got an hour of training and a cheat sheet of prompts.

Each person or pair took on a real business problem from their own unit and wrote up one proposal: the idea, the solution, the rationale, and a plan to build it. A panel of 22 experts scored every proposal on a scale 1 to 10, and every proposal got ~3 reads each.

What the study found: Proposals from 1-person teams with AI scored:

1. significantly better than 1-person teams without AI;

2. about the same as 2-person teams without AI; and

3. marginally (i.e., not statistically significant) worse than 2-person teams with AI.

From a speed and volume perspective:

• Proposals without AI ran 380 to 410 words. With AI, it was closer to 900 words, with fuller rationale and more detail on how to build the thing.

• One-person teams with AI worked 16% faster than 1-person teams without AI. Two- person teams with AI were 13% faster than the 1-person teams without AI. In other words: both AI groups finished faster than the lone worker with no AI, and adding a second person without AI saved no time.

My take: Speed and volume are the easy headlines, and “AI replaces workers” is the easy fear. There’s a more important finding that should come out of this, however:

AI doesn’t eliminate people, it replaces coordination—the hand-offs, translations, and meetings that exist so one function can explain itself to the others.

Silo-Breaking Is the Real Story

Without AI, P&G’s R&D people pitched technical solutions while the commercial people pitched commercial ones. Typical silo behavior.

With AI, that split vanished. The R&D people started producing commercially sharp ideas and the commercial people started producing technically sound ones, and the quality held up either way.

One person, with AI, covered both sides of a divide that normally takes two people in a room. What that exposes:

A lot of “cross-functional collaboration” isn't collaboration.

Count how many of your meetings exist so one function can explain its language to another. Product sits with Compliance to learn how a feature reads against the reg. Marketing pulls in IT to find out what the systems will allow. Lending waits on Risk to turn a policy into something it can offer. Those meetings exist to translate.

AI speaks enough of both sides. AI doesn’t know Compliance the way your head of Compliance does, but it knows enough Compliance to enable Product to avoid having to schedule a meeting with them.

Experts still know more than the model—but you don’t have to route work through that expert to get a first draft that accounts for what they know.

AI did the same for work outside someone’s specialty. People who were put on a task they didn’t normally handle, working with AI, scored as well as a team that had a specialist on it. Same mechanism: AI handed one person the specialist’s framing instead of making them go find the specialist.

When a Second Person Still Pays Off

AI absorbs the coordination, but it doesn’t absorb everything. A second person’s value depends on the kind of work.

The first kind is the routine stuff your shop turns out by the ton: the loan memo, the product write-up, the marketing brief, the board deck. On that work, the second person stopped adding anything once AI was in the mix. One person with AI scored the same as two. The extra set of hands was wasted.

The second kind is the home run: the genuinely exceptional idea, the one in ten that rises above everything else on the table. There, two people with AI pulled ahead, producing a standout about twice as often as one person with AI did.

One person with AI was barely better at it than working with no AI at all. The pair’s edge is real, but slim. Two people with AI didn’t make the typical proposal any better. They produced more of the exceptional ones.

My take on why: a second person pushes back, offers another angle, and argues with the AI’s first draft instead of running with it.

For the bulk of what your shop turns out day to day, one person with AI now matches what a pair used to produce. When you need a breakout—a new product, a new marketing campaign—put two people and AI on it. That combination roughly doubles your odds of a standout (according to the study).

The Morale Surprise

In the study, participants rated their mood before and after the task. The AI users came out more enthusiastic and less anxious than the people working alone. Two people with AI had the best mood swing of any group, ahead of the human pairs without AI. And even one person with AI beat the human pairs on mood.

One surprise: AI users rated their own work lower than it deserved, and that gap was widest among the 1-person with AI teams. Their proposals got better while their confidence in them slipped.

What This Means for Banks and Credit Unions

Translate the P&G test to banking and what they tested were loan committee memos, product proposals, marketing briefs, compliance write-ups, and board decks. What bank and credit unions execs should take away from the Harvard study:

• Stop treating meetings as the default coordination mechanism. The first thing AI removes is translation, not labor. A lot of coordination overhead exists only to translate between functions, and that translation layer is what AI removes first.

• Put people on work outside their lane sooner. Someone who doesn’t normally do a task, working with AI, matched a team that had a specialist on it. Rethink how long you make people wait before handing them unfamiliar work.

• Stop defaulting to two-person staffing. One person with AI now matches the quality a cross-functional pair used to produce, whether the pair shared a job or not. Save the pair for the work that has to be exceptional.

• Don’t staff everything for the home run. For most of what you produce, one person with AI is enough. Reserve the human-plus-AI pair for the few projects where a solid result isn’t good enough.

• Your best people may undersell their work. Track output alone and you’ll miss why adoption stalls. The people who don’t believe their AI-assisted work is better will go back to doing it the slow way. Manage the belief, not just the output.

• The mood findings point to the need to coach people. People who don’t trust the AI-assisted version of their work drift back to the old way. That’s a management problem to get ahead of to ensure adoption of AI tools don’t go by the wayside.

What About Zero-Person Teams?

You might be wondering, “did the researchers test the output of just AI, with no human intervention?” They did—after the study.

When the researchers ran the AI by itself, its proposals came back looking alike, mostly variations on a single answer. The people were what pushed the solutions apart, which helps explain the home-run result: the exceptional ideas came from a human, and a second human with AI produced more of them.

Bottom Line

Regarding the academic study, someone is bound to say “this was a one-day workshop—that’s not how real projects work.” Fair point. Real projects involve rework, politics, changing priorities, and all the other joys of corporate life.

But that’s not a reason to ignore the finding. It’s a reason to ask how many of those activities exist because people need to translate between functions in the first place.

For decades, banks solved coordination by adding people—another specialist, another committee, another meeting to get the functions in a room. The Harvard study points the other way. Give one person enough of the next function’s thinking, and much of that coordination stops being necessary.

Much of your org chart exists only to move information between people who can’t speak each other’s language, and that’s the part AI is coming for first.

Come into the Fintech Snark Tank and tell me why I’m wrong.