Something shifted while the ink was drying on the last issue. Agents stopped being things that fetch answers and became things that touch money. DBS switched on agentic banking for ten million customers - every action gated behind a login, but actions all the same. An onchain infrastructure firm shipped a suite that lets an agent hold, move and stream balances outright. And Goldman and JPMorgan, inside the same week, built baskets that let you trade the AI build-out itself - the debt behind the data centres, long or short, up to $250 million at a clip.

Last issue the moat was data - whose intelligence your agent stands on. This fortnight it moved one step downstream, to the wallet. And that step changes what’s scarce. Not the model, not even the data, but the quiet machinery that has to sit under the moment an agent spends: who it is, what it’s allowed to touch, and who answers when it gets it wrong.

That last one is the question nobody has finished writing. On-chain settlement has no chargeback. US law still hasn’t said who’s liable when an autonomous agent pays the wrong party. A regulator spent the fortnight warning developers that even shading an output away from accuracy could count as deception - a sign of how nervous the ground underneath all this has become. The capability shipped. The accountability didn’t.

Let’s get into it.

- The Editor

The Brief- Things Worth Knowing

DBS Took Its Agents From Answering To Doing - For 10 Million Customers, Behind A Login

DBS said its Gen-AI assistants - DBS Joy for corporate and SME clients, digibot for individuals - are going agentic across Singapore, Hong Kong and Taiwan, now reaching more than 10 million users and an expected 1 million-plus chats a month. DBS Joy turned fully agentic in Singapore this week: its roughly 350,000 corporate users can complete tasks inside a single conversation - asking whether a payment cleared, what they paid in fees, whether money went to a named company - instead of being pointed at the right screen. digibot, already resolving nine in ten retail queries digitally, gets agentic and voice capabilities in Q4 (checking card usage, tracking rewards, blocking or replacing cards). The guardrail is the headline: agentic actions switch on only after a customer logs into DBS IDEAL or digibank, the AI acts solely on an authenticated customer’s instructions, and a human handoff stays one tap away.

The Takeaway
The authority boundary here is authentication, not autonomy. DBS didn’t set an open agent loose - it gated every action behind a verified login and kept a human escape hatch. If you’re putting agents into customer channels, copy that sequencing: establish identity first, bound the action set second, preserve the handoff throughout. “Moving from asking to doing” is only safe when the system can say exactly who did the asking.

Goldman And JPMorgan Turned The AI Build-Out Into Something You Can Trade - The Same Week

Within days of each other, the two banks shipped products to dial exposure to AI-related debt up or down. Goldman launched a basket of bonds from 18 equal-weighted US high-yield issuers - CoreWeave, Applied Digital and others - with the desk pricing inquiries from $50m to $250m, tradable as the underlying bonds or as total-return swaps. JPMorgan put out three baskets the same week: long-dated investment-grade hyperscaler and project-finance names (Microsoft, Meta, Amazon, Alphabet, Oracle), a 15-issuer AI junk basket, and a 16-name semiconductor and hardware basket including Nvidia.

Read them together and the signal is that AI has become an exposure to be managed, not just a capability to be bought. The capex cycle now has a tradable wrapper - and the banks, not the model labs, are the ones packaging it.

The Takeaway

Watch where the smart money is hedging. When the two biggest dealers build baskets to trade AI-buildout risk in the same week, that’s a read on how concentrated and how debt-financed the AI cycle has become - useful intelligence whether or not you ever touch these instruments.

Zerohash Gave Agents a Wallet - And Joined The Standards Body The Same Day

Zerohash launched its Agentic Finance Suite - products that let AI agents hold, move and stream money programmatically, including stablecoin balances - and simultaneously joined the x402 Foundation, the open-standards body for machine-driven payments. Where Anchorage puts the agent inside a regulated institution, zerohash embeds money movement directly into whatever product a fintech is building. Same problem, opposite end: one hands you a rail, the other hands you an API.

The Takeaway

“Can your agent transact?” is quietly replacing “can your agent answer?” as the product question. If you build in payments or fintech, the decision in front of you is whether money movement becomes a native capability of your agent or a dependency you rent - and either way, the controls around what it’s allowed to move are now your problem.

The FTC Drew a Line On AI Accuracy - And The Comment Window Just Closed

The FTC’s proposed policy statement warns that altering an AI system’s outputs away from accuracy - even to comply with a state law - may count as consumer deception under Section 5 of the FTC Act. The public comment period closed July 31. It’s aimed at developers, but it lands squarely on any financial firm deploying AI in customer-facing decisions, where “correct, explainable and disclosed” is exactly the bar that gets a tool past risk review.

The takeaway: accuracy is becoming a compliance surface, not just a quality metric. Before you tune a model’s behaviour for any reason - fairness, policy, brand - document what the tuning does to accuracy and whether you’ve disclosed it. The regulator is signaling that a quiet trade-off is the risky one.

The Workflow

The execution story last issue was institutional - a bond desk letting an agent act inside bounded authority. This fortnight the same architecture showed up at the retail edge, which is where it gets uncomfortable.

Two platforms for individual investors launched days apart, both built on the identical spine: the user sets a strategy and hard risk limits, the agent monitors markets across asset classes continuously, and when a condition fires the agent forms and routes the trade - but only within the parameters the human set, and with execution passing through a broker’s API so the platform never holds the money. It’s the professional desk’s bounded-authority model, sold to someone who doesn’t have a compliance team.

Walk the diagram and the design logic is the same as the institutional version, with one difference that matters: the authority boundary isn’t a human clicking approve on each trade, it’s the parameter set configured up front.

Step 1 is the whole ballgame - the strategy and the loss limits are the consent layer, and everything downstream is bounded by them.

Steps 2 and 3 are what humans are not so good at: watching everything, all day, then acting the instant a condition matches.

Step 4 is the gate - the intended action is tested against the user’s hard limits before anything happens.

Step 5 keeps the structural guardrail at the money layer: execution routes to a broker, funds never sit with the platform.

Step 6 logs it all to one session memory - the audit surface a user, and eventually a regulator, will want.

The problem that hasn’t been solved is accountability at the consumer edge. There’s no card-network chargeback here, no established dispute path, and reliability that frays over long chains of tool-calls. The architecture works; the question is who’s responsible when it doesn’t. That’s not a reason to avoid these tools - it’s the reason the highest-autonomy setting should stay a deliberate choice, not a default.

The Toolkit

The Brief keeps circling one problem from three directions: an agent that can act needs somewhere safe to move value, and in the last two weeks three different answers shipped. They aren’t competitors so much as three places to put your trust boundary - and which one fits depends entirely on where your money-movement risk needs to live. Anchorage puts the agent inside a regulated institution. Zerohash lets you embed money movement into your own product.

The x402 Foundation settles it over a neutral open protocol no single network owns. The graphic below lays out what each is best for, what it buys you, and what it costs you - and the one thing none of them has solved yet.

The Takeaway

Map the three to your own situation before you pick a vendor. Need charter-level trust before an agent touches capital → regulated rail. Building a product and want money movement as an API → embedded infrastructure. Moving value across platforms you don’t control → open protocol. And whichever you choose, note the shared gap: when an autonomous agent makes an unauthorized payment, who’s liable is still being written.

Weekly Coverage - The Broader Signal

LSEG - The Grounding Number

On its H1 2026 call (July 30), LSEG raised its full-year outlook and said AI is driving demand for its data, with CEO David Schwimmer telling analysts that AI has enhanced the company’s value. The number that matters for this issue: tool calls through its MCP connector — which pipes LSEG data straight into customers’ own agents — rose nearly fivefold from May to June across 200-plus clients, and agents pull roughly ten times the data a human does through that channel.

The Signal: last issue’s data-moat thesis just showed up in an earnings call. Grounding stopped being an argument and became a revenue line - and the agents doing the consuming are hungrier than any human user ever was.

Two weeks ago the frontier was grounding - whose data your agent stands on. This fortnight it moved one step downstream, to the moment the agent stops answering and starts doing: allocating exposure, holding a balance, routing an order, moving value.

Every story here is a different group racing to build the thing that makes that moment safe - a regulated rail, an embedded API, an open standard, a regulator’s line on accuracy.

The uncomfortable part is what still isn’t built. When an autonomous agent moves money it shouldn’t have, the question of who answers for it doesn’t have a settled answer yet - not in the card networks, not in the protocols, not in US law. The institutions that win the next phase won’t be the ones with the most autonomous agents. They’ll be the ones who can say, precisely and on the record, who set the limits and who’s accountable when the limits are tested.

Build the accountability before you need it.

Until we meet next!

Best,
Swati S
Editor-in-Chief
Finance×AI Lab

Share Finance×AI Lab

That pause is this entire issue.

The autonomy debate is over. We spent a year asking whether an agent should be allowed to act - price the loan, flag the transaction, move the money. It acts now. In production, at institutions you’d recognise, in workflows that used to demand a signature.

So the question moved, and it landed somewhere far less comfortable than can it act? The moment an agent acts on someone’s money, a new one opens underneath it: when it gets a decision wrong, and it will - can you explain what happened, undo it, and prove you were in control the whole time? “The model was tested” is not an answer a regulator accepts. “A human can intervene” is not a control if nobody can find the button. Trust, it turns out, was never something the model produced. It’s something the product around the model either has, or doesn’t.

And that quietly splits FinTech into two kinds of team. One treats trust as a message - a line marketing adds once the product is built. The other treats it as architecture: decisions made on purpose, at the level of each individual thing the system is allowed to do. The first kind demos beautifully. The second kind survives a real customer, a real complaint, and a real audit.

This issue is a field guide to becoming the second kind - where autonomy actually belongs, how to explain a decision to the person it hit, how to make a system throttle its own power when the ground moves, and why controlled autonomy, not maximum autonomy, is what ends up winning.

Let’s get into it.
- The Editor

FinTech products are built on trust. Customers trust a payment product to move their money correctly, a lending product to assess them fairly, an investment product to act responsibly, a fraud system to protect them without blocking their real activity.

When AI enters these products, that trust gets harder to manage. The system analyses more data, decides faster, personalises, and can act without waiting for an employee. Greater intelligence and speed raise one question that matters more than any capability demo:

What should the system be allowed to decide on its own?

You can’t answer that with broad statements about responsible AI, human oversight, or regulatory compliance. You answer it at the level of individual financial decisions - how the product decides, how much independence it gets, how decisions are explained, and what happens when something goes wrong.

Financial Products Don’t Make One Decision - So Stop Treating Autonomy As One Switch

Teams love to debate autonomy as a single toggle. Should the assistant be autonomous? Should there always be a human in the loop? Can the model make the final call? For a financial product, these questions are too blunt.

A single AI feature makes many decisions across the journey. A lending product might organise application data, spot missing documents, estimate risk, recommend a credit limit, flag a case for review, and communicate the outcome. Those decisions don’t carry the same weight. Some only inform. Some influence an employee. Some directly affect the customer. Some reverse easily; others hit someone’s access to money right away.

The first move is a decision inventory: list every decision the feature makes, give each its own row, and examine each on its own terms.

The Trust & Autonomy Canvas

The playbook turns that discussion into a documented, reviewable product artefact. One pass through these eight dimensions per decision, and “is it autonomous?” becomes a much more useful conversation.

Let Reversibility and Value at Risk Set the Ceiling

In financial products, the consequence of a wrong decision matters as much as how likely it is. A reversible action with limited impact can justify more autonomy. A hard-to-reverse action with real customer impact demands stronger evidence and tighter controls.

So make both explicit. For each decision, estimate the maximum financial impact on a single customer if the system gets it wrong. You’re not chasing a perfect prediction - you’re making the potential consequence visible before you decide how much independence to grant.

Skip this, and you’ll quietly hand the same autonomy to decisions with wildly different risk. A feature that organises information is not the one that moves money. A feature that recommends to an employee is not the one that executes for a customer. The interface can make both feel seamless - their controls should not be identical.

A Financial Product Should Know When to Reduce its Own Autonomy

A trustworthy system knows not only when it can act, but when it should stop. Every autonomous feature needs an escalation trigger - a defined set of conditions that automatically dials autonomy down.

The trigger might be input-data quality or freshness, a shift in model performance, unusual activity, a spike in human overrides, or something product-specific. The exact condition varies; the principle doesn’t: if the conditions supporting an autonomy decision change, the autonomy level changes with them.

That needs a circuit-breaker specification, not a vague promise to “escalate uncertain cases.” Document what trips it, what happens when it does, who receives the case, and how normal operation resumes.

You also need a real override path. Customers and operators must be able to stop, challenge, or escalate a decision - fast enough to matter and visible enough to be used. An override that technically exists but is buried is not a control.

Explainability isn’t a Compliance Checkbox - It’s The Customer Experience

Explainability gets filed under “technical” or “regulatory.” In FinTech it’s also product. A financial product should explain an individual automated decision in plain language:

• What information influenced the decision?

• What did the system decide or recommend?

• Why did it reach that outcome?

• What can the customer do if it looks wrong?

A customer shouldn’t need to understand model architecture to understand what happened to them. Financial decisions are personal - an aggregate accuracy metric does nothing for the individual who just got an unfair or unexpected outcome.

Internally, you should be able to reconstruct the decision: the audit trail has to capture the relevant inputs, reasoning, tool calls, model version, and output. Without that, you know a decision happened but can’t establish how or why.

Poor explainability is also expensive. Customers contact support because they don’t understand an outcome. Support can’t investigate. Compliance can’t reconstruct the case. Complaints drag. Reviewers burn time decoding model behaviour. Explainability is part of how the product operates - not an add-on bolted to the model.

Compliance Starts Before You Build - Not The Week Before Launch

The classic FinTech failure is treating compliance as a review near launch. By then the product has already locked in decisions about data, architecture, experience, model design, and autonomy. A serious issue found late means redesigning the product, not editing the documentation. So run compliance across the whole lifecycle.

Stage 1 - Discovery & Design

Purpose: decide whether the product can be built responsibly in its intended form. These questions should shape whether you build it, not just how you document it later.

Stage 2 - Build & Validation

Purpose: evidence that the model, controls, and decision process behave as intended - reviewable proof, not general assurance.

Stage 3 - Pre-launch

Purpose: test the controls under realistic conditions. A product shouldn’t go live just because the model performs well - escalation, rollback, explanation, and oversight have to work in practice.

Aggregate Performance Can Hide Real Customer Harm

A product can look great overall and still produce poor outcomes for one group. That’s why averages aren’t enough. Examine override rates across significant segments. Review fairness metrics across relevant groups. Compare performance to the validation baseline and investigate drift instead of waving it off as normal variation.

Complaints deserve the same treatment. A rise in AI-attributed complaints can reveal what standard metrics miss - decisions that are hard to understand, an escalation path that isn’t working, or a pattern that needs a closer look. Trust can’t be measured by average accuracy alone; you need to know who is receiving poor outcomes, which decisions get overridden, and whether the reasons are shifting over time.

Monitoring is Part of The Product, Not a Launch Formality

Shipping doesn’t finish the trust work. Customer behaviour shifts, input data changes, the model gets updated, new segments and markets arrive, and the system is gradually allowed to take more consequential actions. Controls have to evolve with the product - on a defined cadence, not the occasional review.

Periodically, pick one real case and reconstruct the whole decision. Can you identify the inputs? The model version? The reasoning? The guardrails that applied? The output and the action that followed? A system that can’t reconstruct its decisions can’t meaningfully claim control over them.

Trust Needs Evidence, Not Reassurance

FinTech teams reach for comforting lines: the model has been tested, the product is compliant, a human can intervene, the system is monitored. On their own, none of these hold. Each important claim should produce an artefact someone can actually review:

Decision inventory · reversibility rating · value-at-risk assessment · evidence checklist · circuit-breaker spec · override design note · named approver and review date · model card · fairness assessment · independent validation report · escalation test record · rollback runbook.

These turn trust from an aspiration into an operating discipline - and they make accountability visible. Someone proposes the autonomy level. Someone reviews the evidence. Someone approves. Someone monitors. Someone decides whether the conditions still justify what was granted. Without named ownership, trust stays a vague organisational ambition instead of a product responsibility.

Re-earn Trust as Your Ambition Grows

An autonomy decision made at launch shouldn’t stand forever. You expand into a new jurisdiction. The model starts using extra data. The product moves from recommending an action to executing it. The customer base grows beyond the population you validated on. Each change alters the risk profile - so autonomy decisions need a review date.

A Trust & Autonomy Canvas completed at launch and revisited six months later shows whether governance kept pace: Has the system gained independence? Has the maximum financial impact grown? Are more cases being overridden - and are those overrides concentrated in one segment? Do the explanations still reflect how the product works? Does the rollback still match the live system? Was the model validated for the population now using it?

Trust isn’t permanently earned by a clean launch. It’s maintained as the product evolves.

The Real Edge is Controlled Autonomy

The strongest AI-powered financial products won’t be the ones that automate the most. They’ll be the ones that know where autonomy creates customer value and where it introduces unacceptable risk - separating low-risk, reversible decisions from consequential financial actions, and defining when the system may proceed, when it must confirm, and when it must escalate.

They’ll explain individual decisions in language customers understand. They’ll make intervention practical. They’ll treat fairness, auditability, monitoring, and rollback as product capabilities, not compliance paperwork. And they won’t treat trust as something marketing adds after the product is built.

Go Deeper

Building AI-Powered Financial Products

Everything above - how much autonomy an AI system should have, how financial decisions should be explained, where human oversight belongs, and how compliance should shape the whole product lifecycle - is just one slice of building trustworthy AI-powered financial products.

In Building AI-Powered Financial Products, Abhijit Dey, Srinivasan Shanmuganathan, and David Roldán Martínez go far deeper. The book shows how product leaders, FinTech professionals, and financial-services teams can apply generative AI, AI agents, large language models, and proven product frameworks to build trustworthy, scalable products across payments, lending, digital banking, insurance, and wealth management.

The focus isn’t bolting AI features onto an existing product. It’s making better product decisions around customer value, data foundations, model behaviour, autonomy, explainability, governance, compliance, monetisation, and scale - with practical frameworks, templates, real-world lessons, and structured playbooks you can use to move from an AI opportunity to a product that survives real customer, regulatory, and commercial constraints.

Get the book (Exclusive discount for a limited-time)

Packt Website

Amazon

Best,
Swati S
Editor-in-Chief
Finance×AI Lab

Rishi lays out a practical, grounded vision: how finance professionals can encode their own expertise into “skills” and “context layers,” combine them into an organisational brain, and wrap the whole thing in an agentic harness so AI behaves reliably instead of confidently guessing.

Along the way he tackles the question everyone is quietly asking - if agents do the work, what’s left for me?, with a refreshingly optimistic (and historically grounded) answer.

If you work in finance and feel like you’re drowning in AI hype, this one is for you.

Meet Rishi Sapra

Rishi is a chartered accountant, Microsoft MVP, and founder of Learn Data Insights. His path is unusually well-suited to the finance-meets-AI moment: after studying philosophy and economics at the London School of Economics, he spent his first decade across the big four and major banks - product control at Deloitte and HSBC, Excel/VBA finance automation at Barclays, and financial modelling at KPMG.

When Power BI launched, he saw where things were heading, left KPMG for a BI consultancy, and eventually landed at Avanade (the Microsoft–Accenture joint venture), where he works today. As a Microsoft MVP he gets early access to products and a direct line to the teams building them. He’s also deeply embedded in the finance community, sitting on the ICAEW’s data analytics advisory committee.

His core belief threads through the whole conversation: “The biggest problems in technology aren’t technology problems, they’re human problems.”

About Finance×AI Lab

Finance×AI Lab is a community and platform built to bring practitioners together at the intersection of finance and AI. The Voice of Impact podcast is where working professionals share, in depth, how things actually operate inside the finance function - what’s changing, what’s working, and what’s just noise. This episode is a strong example of that mission: less hype, more hands-on wisdom.

What This Episode Covers

1. Meet people where they are - and start with Excel

Rishi opens with a problem every trainer knows: the finance community is anything but homogeneous. In the same room you’ll find data scientists who know more than he does, sitting next to people who live entirely in Excel. His answer is to start with the one thing everyone shares.

At AccountX (the UK’s largest accountancy conference) and at a recent Data Lens event in Newcastle, he demonstrated building custom agents and skills directly inside Excel - no leaving the application.

His warning is that tools like Power BI and Fabric started simple but have grown genuinely complex; a three-day course rarely survives contact with messy SAP or Oracle ERP data. Community, and a shared starting point, is how people bridge that gap without being permanently dependent on IT.

2. Self-service AI: stop being a consumer, start being an operator

This is the heart of the episode. Out of the box, Gen AI is brilliant but generic - its answers are “plausible, confident-sounding, and not personalised to you, your team, or your business.”

Rishi’s fix is a mindset shift: treat AI less like a search box and more like a team - coaches, collaborators, consultants, sparring partners. You bring what you uniquely know (how the business makes money, what controls make the numbers trustworthy); AI fills the technical gaps (data cleaning, DAX, engineering, visualisation) that take years to learn.

He calls this self-service AI - the honest successor to the self-service BI promise that so often collapsed back into centralised bottlenecks.

3. Skills, context layers, and the “organisational brain”

How do you actually make AI stop hallucinating? By giving it what you’d give a new human employee: a role, defined responsibilities, guidance on applying their skills, and access to the right data and systems.

Rishi walks through a concrete demo - downloading a free skill, feeding in a workbook and a three-minute transcript of an analyst explaining a process, and having AI generate an implementation plan, then a reusable skill that reproduces the exact calculation and journal sheets on demand. He describes it beautifully as “almost exactly like a VBA macro, but written in natural language rather than code” - deterministic where it should be, probabilistic where reasoning is needed.

Stack enough of these together and you get an organisational brain: a governed knowledge base of data sources, schemas, logic, governance policies and access permissions that agents sit on top of.

He points to Andrej Karpathy’s single-file “maintain a wiki” agent as the simplest possible starting point, and to Microsoft’s Foundry IQ as a purpose-built grounding layer for agents.

4. The agentic harness - and the future of finance jobs

More context isn’t automatically better. Rishi shares a cautionary tale: after loading tens of thousands of lines of Microsoft Fabric skills, his single instruction “never deploy directly” got overruled - because Gen AI is constantly weighting context and treating skills as merely advisory. The fix is an agentic harness: gating skills with defined inputs and outputs, chaining them into structured workflows (almost like a legal system you have to follow), and using LLM-as-a-judge to evaluate outputs against your standards for format, accuracy and PII safety.

Then comes the big question: are our jobs at risk? Rishi’s answer draws on history. When spreadsheets arrived in the 1980s, bookkeeping roles declined - but accountants, auditors and financial advisers grew. The roles displaced by technology tend to be replaced by higher-value ones: supervising agents, interpreting outputs, and making judgment-driven strategic decisions in the real world that AI can’t reach.

His closing warning to leaders and to younger professionals is pointed: don’t hand over your thinking. If AI is going to do the work, who trains it, and who validates the output? Foundational knowledge - of what “good” looks like - is what lets you check the machine. Learn through community, experiment relentlessly, fail fast, and stay honest about what you don’t know.

The Takeaway

Rishi’s throughline is empowerment. The most exciting thing about this moment, he argues, isn’t the tools themselves - it’s that finance professionals can finally get the knowledge in their heads and spreadsheets into AI and have it work with them, not just be another vendor’s black box. As he puts it: “It needs to come from me, because I’m the one with the knowledge.”

🎧 Watch the full conversation above - and if you’re a finance professional navigating AI, this is a genuinely practical place to start.

Connect with Rishi on Learn Data Insights, and join the Finance×AI Lab community to keep the conversation going.

Last issue we asked whose data your agent runs on. Three weeks later the market answered a harder question, and it had nothing to do with data.

For nineteen days, the single most capable model on the market simply vanished. Not because it broke. Because of a licensing clause and an export directive. Enterprises that had wired Fable 5 into production discovered they had built a single point of failure they did not control, and then, on 1 July, it came back, on changed terms.

That is the story of the fortnight. Agents that can act are only as reliable as the access underneath them, and this month more than one institution learned that access is not a settled utility. It is a supply chain. The moat is still the data. The fragility is the model.

Let’s get into it.

The Editor

The Brief- Things Worth Knowing

Fable 5 and Mythos 5 are Back - But the Terms Changed

The US Department of Commerce lifted export controls on Claude Fable 5 and Mythos 5 on 30 June; Fable 5 returned globally on 1 July across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. The 19-day suspension had caught every user, domestic and foreign, because Anthropic couldn’t verify nationality in real time and shut everything off rather than risk breach.

What actually changed / conditions it’s back on:

• Trigger was a jailbreak, not model quality. Amazon researchers found a way to prompt Fable 5 into identifying software vulnerabilities and, in one case, producing exploit code. Anthropic’s review concluded less-capable models (Opus 4.8, GPT-5.5, Kimi K2.7, and others) could do the same, and it retrained a safety classifier that now blocks the reported behaviour in >99% of cases.

• Fable 5 vs Mythos 5: Same underlying model; Fable 5 ships with the strongest safeguards Anthropic has deployed (high-risk sessions route to Opus 4.8 on ~5% of sessions).

  Mythos 5 carries the full offensive-cyber capability stack and is only restored to ~100 vetted US organisations defending critical infrastructure (approval granted 26 June, via a letter from Commerce Secretary Lutnick to co-founder Tom Brown, who led negotiations).

• New commercial condition: For Pro, Max, Team and select Enterprise plans, Fable 5 was included for up to 50% of weekly usage limits through 7 July, after which it moves to usage credits. AWS, Google Cloud and Microsoft Foundry re-enablement was still pending at restore.

• The governance quid pro quo: Anthropic agreed to proactively detect/address security risks, work with the government on standards for future models, and report malicious activity - no export licence required.

Use Case: Why it matters to a finance function?

Treat model access as a supply-chain risk, not a settled utility. Any bank, insurer or asset manager that auto-routed workflows to “latest Anthropic model” learned that removing Fable didn’t degrade gracefully.

The practical move: multi-vendor fallback by design (the HSBC Google-plus-Mistral pattern), model-access clauses in vendor contracts, and a documented degradation path in your model-risk framework. For anyone doing second-line model validation, this is a live example of vendor/model-availability risk belonging in the operational-risk section of your validation report, not just accuracy and stability.

Taktile Raises $110M (Goldman-led) to Make Banks “agent-first”

Taktile closed a $110M round led by Goldman Sachs Alternatives. Its platform deploys AI agents purpose-built for regulated institutions, automating complex operational workflows while keeping regulatory guardrails.

Use Case

Autonomous decisioning in commercial lending, insurance claims and business underwriting - a small-business loan that took weeks of manual underwriting compressed toward minutes; claims that took months evaluated in hours. CEO Wehmeyer’s framing: the advantage is decision-time compression, not headcount cost savings.

Why it matters?

This is capital validating the “agent-first interface” thesis - conversational AI as the front door for account opening, loan applications and servicing. The Goldman lead is the signal to read.

UPCOMING EVENT (FREE REGISTRATION)

Happening next week: The workflows in finance and reshaping finance skills for next 3-5 years, live.

On 17 July we are hosting a free panel on the finance skills that actually matter in an agentic era.

One more thing. For a limited time, every subscriber* who registers through our newsletter gets a complimentary copy of the $30 Packt title, The Profitable AI Advantage, on us. Register for free, walk away with the book, join the discussion.

Register Here

Quantifind Raises $200M to Industrialise Financial-crime Intelligence

Quantifind announced a $200M growth investment led by Summit Partners; existing backers include Citi Ventures, S&P Global and Deloitte. Money targets AI-native risk intelligence and governed agentic middleware for financial-crime and national-security operations.

Use Case

AML/financial-crime detection - the same category FIS+Anthropic’s Financial Crimes Agent targets (compressing investigations from hours to minutes by assembling evidence across core systems and scoring against typologies).

Why it matters?

Pair this with the UK Finance fraud numbers from our last issue - investment fraud up 40% YoY, APP fraud two-thirds online. Capital is flowing into the defensive intelligence layer, not just the productivity layer.

The Workflow

The Toolkit

Weekly Coverage - The Broader Signal

Wealth Management Crosses From Assistant to Workflow Partner

BetaNXT launched Val, an agentic platform, plus its InsightX ecosystem - rules-based intelligence across documents/data/workflows with human oversight.

Use Case: High-volume operational processes (document validation, business-rule checks) without added risk.

Signal: The industry's framing has shifted from "does AI matter" (settled) to "is it producing measurable value or still stuck in pilots."

Emirates NBD + Techstars Back Agentic-finance Startups

Use case: Enterprise-grade AI in compliance, WealthTech, SME banking and capital markets, tested on cloud-native banking infra across 9M customers.

Signal: Incumbents building startup pipelines to source agentic capability rather than build it all in-house.

Why Reserving Won't Be Automated- Insurance/Actuarial Angle

WTW's Magdalena Ramada made the counter-point worth quoting: Generative AI is ill-suited to risk prediction (same prompt, 50 different answers → explainability fails under scrutiny); traditional ML and GLMs remain the right tools, and complex actuarial functions such as reserving cannot be fully automated.

Use case: scoped agents on low-risk, high-volume steps (FNOL triage, document handling), humans on consequential decisions.

Sixfold Ships a Straight-through AI Underwriter

Use case: An agent that learns a carrier’s book/appetite and can take submissions through to quote- and bind-ready materials, each carrier’s data walled off. Customers span $270B GWP (Zurich, Generali, Guardian, Axis, New York Life).

Signal: Underwriting timelines collapsing from days to minutes on standard risks.

Everything in this issue points the same way. The debate about whether AI can act is over. The question now is quieter and sharper: on whose terms does it act, and what happens the day that access changes without warning. The institutions that treat that as a supply-chain question, not a technology one, are the ones who will not be caught flat-footed by the next nineteen-day blackout.

And that is exactly why we do this. We bring you the opportunities to adapt & upskill, and the discussions worth being part of, so that when the important decisions land on your desk, you can make them easily.

See you at the panel on 17 July.

Best,
Swati S
Editor-in-Chief
Finance×AI Lab

Share Finance×AI Lab

*Offer exclusive to subscribers of Finance×AI Lab.

We recently hosted a live session on AI in Capital Markets, and the conversation didn't stop when the slides ended.

This episode captures the Q&A that followed, where the audience pushed past the headline numbers into the practical questions: what is moving from pilot to production, where agentic systems are already on the desk, and how model governance is straining to keep up.

The discussion stayed candid and exciting, with more on what practitioners actually need to watch.

If you work anywhere near markets, risk, or model validation, this is the part of the event worth replaying.

Thanks,

Swati Sharma
Editor-in-Chief
Finance×AI Lab

The conversation in most finance circles still treats AI as a tool decision: which platform, which vendor, which price point. After a year inside the work, my view is different. The tool is the easy part. The skills shift it forces is the hard part, and the longer you live with it, the clearer that becomes.

Four shifts in particular stand out-

From Data Preparation to Data Interpretation

The hours that used to disappear into reconciling, formatting and chasing have shrunk dramatically. But the saved time does not stay free. It gets reallocated, almost overnight, to harder questions. What does this number mean for next quarter? Why is the trend bending? What should we do about it? Finance professionals who built a career on accuracy now have to build a second career on interpretation. That is a different expertise, and senior finance professionals are expected to thrive in this arena.

From Reporting to Storytelling

Boards do not want a pack of numbers anymore. They want the story behind the numbers, and they want it now. AI can produce a variance analysis in seconds. But translating that variance into a narrative someone can act on still depends on a human who understands the business well enough to write it. Storytelling, with structure and persuasion, is becoming a core finance skill, not a soft one.

Skill shift here is from producing to persuasion.

From Process Owner to Process Designer and AI Supervisor

This is the shift I underestimated most. Implementing AI in AP, AR or reporting is not a set-and-forget exercise. It requires deliberate human design. You decide where the AI fits, where the judgement still sits, where the controls live, and who signs off when the output is wrong. The role moves from owning a process to designing the system in which AI operates and supervising it daily.

Skills need to be rewired to question better and understanding the Output.

Which brings me to the most uncomfortable point.

From Spreadsheet Fluency to Prompt and AI-governance Fluency

AI makes mistakes. Not occasionally. Regularly. Sometimes confidently. There is a real gap between what I expect the output to be and what the model produces, and that gap closes in only two ways. First, by writing better prompts. Second, by checking the output thoroughly before it touches a report that we are producing. Both of those are learned skills. They are not intuitive, and they are not optional.

The encouraging part is that the gap does narrow with time. As I have fed back to the AI tools both the qualitative and the quantitative responses I want, the outputs have improved. The model learns the house style. AI use can be highly personalised. The work gets better the more you train it. But that is also a reminder that the model is only as good as the hours you invest in it.

Traditional Finance skills becomes a secondary expertise.

Where SMEs Are Getting Stuck

After speaking with a lot of small and mid-sized business owners who feel challenged by the question of where to begin. The barrier is rarely access. Most of the AP, AR and reporting software they already use is now AI-enabled, often without them realising it. The real barriers are cost and confidence. Using the full capability of these platforms at scale can be expensive, and very few SME finance teams have the in-house knowledge to deploy them well.

My Practical Advice is to Build the Muscle Internally Before You Spend

A starting point: protect ten percent of every finance employee’s weekly hours for two things. First, learning to write better prompts. Second, learning how AI agents work, what they can do, and where they break. That is roughly four hours a week per person. In a year you will have a finance team that understands the technology from the inside out, rather than one that depends on a vendor demo to understand what they are paying for.

Start with cheap models. Pick a narrow problem, like supplier statement reconciliation or variance commentary. Get one small win. Then expand. The companies I see succeeding with AI in finance are not the ones with the biggest budget. They are the ones with the most patient training plan.

The future finance team will be smaller, more senior, more strategic and more accountable. None of those qualities arrive with the software. They arrive with the training. AI does not replace professional judgement. It gives you more time to use it, but only if you have invested in the skills to deploy it well.

Author Box

About the Contributor

Akm Boby (FCCA) is a senior finance executive with over 15 years of experience leading finance functions across engineering consulting, construction, technology, and consulting.

He currently serves as Finance Director of BPP Tech Group and CDC Group, where he oversees multi-entity financial management, strategic planning, and board level reporting, and is a Non-Executive Director of the ESG Foundation.

A growing focus of his work is the hands-on implementation of AI and automation inside finance departments, building tools that take the manual load off accounts payable, receivable, payroll, and management reporting so teams can spend more time on judgment and analysis.

He has completed the Senior Management Programme at Cambridge and is pursuing an Executive Diploma in Strategy and Innovation at Oxford, alongside significant experience in fundraising and capital structuring for growing organisations.

Subscribe now

Finance×AI Lab Note:

This piece by Akm Boby is part of a wider conversation we’re hosting on how AI and automation are reshaping finance work in practice.

Boby will be joining Chengai Ruredzo, FCCA, and Philip Maher, FCCA, in our upcoming panel discussion, AI and Automation for Finance: Reshaping Finance Skills for the Next 3–5 Years, where we’ll explore what is actually working inside finance teams, what still needs caution, and which skills may become more important next.

July 17, 2026 | 10 AM EDT | Free registration
Register Here

Six months ago, the hard question in financial AI was whether an agent could be trusted to act - to move money, place a trade, file a report - without a human pressing the button each time. We spent the better part of a year debating autonomy.

That debate is mostly settled now. Agents act. They’re acting in production, at institutions you’d recognise, in workflows that used to be untouchable.

So the question moved. And it moved somewhere more interesting - and more uncomfortable.

Because an agent that can act is only as good as what it knows. Point a brilliant model at thin, generic, or badly-governed data and you get confident nonsense at machine speed. Point that same model at deep, proprietary, well-grounded data and you get something that starts to look like genuine institutional judgement. The model, it turns out, was never the moat. The intelligence beneath it is.

That realisation is quietly rearranging the entire industry. Everyone is now racing to own the layer the agents run on - the data, the analytics, the grounded reasoning that makes an AI output something you’d actually stake a decision on. It’s a quieter fight than the one about autonomy. It doesn’t make for dramatic demos. But it’s the one that will decide who wins.

And it has a sharp edge most people haven’t fully realised yet: if intelligence is the asset, then control over access to that intelligence is power. The institution that grounds its agents in something nobody else has is building a moat. The institution that builds everything on an intelligence layer it doesn’t own is building on someone else’s terms - and this month, more than one of them found out exactly what that means.

Let’s get into it.

- The Editor

THE BRIEF - Four Things Worth Knowing

HSBC put a $100 million price tag on every AI project. That filter matters more than the deal

HSBC and Google Cloud announced a multi-year partnership, targeting more than 200 new AI use cases over two years - built on Gemini models, the Gemini Enterprise Agent Platform, and an existing base of 600+ HSBC applications already running on Google Cloud. The three opening focus areas are familiar: wealth management support, financial crime risk management, and frontline tools for relationship managers.

But the headline number isn’t the 200 use cases. It’s the filter HSBC is applying to them. Only initiatives the bank estimates could each return more than $100 million - in direct revenue or efficiency - get prioritised for delivery. That single sentence is the most useful thing in the announcement.

Here’s why it’s worth your attention regardless of whether you’ll ever sign a hyperscaler deal. Most institutions are still struggling with AI prioritisation as a qualitative exercise - which projects feel promising, which sponsors are loudest, which use cases demo well. HSBC has replaced that with a value gate: articulate the return before you invest, or it doesn’t make the list. It’s the difference between an AI strategy and an AI wish list. The financial-crime detail makes it concrete - HSBC monitors nearly a billion transactions a month and expects to intervene twice as fast under the new architecture, which is a measurable target, not an aspiration.

The takeaway: Steal the filter, not the vendor. Before your next AI initiative, ask the HSBC question - what’s the quantified return, and over what horizon? If the project can’t survive that question, it shouldn’t survive your roadmap.

Tradeweb and LTX made the same bet 24 hours apart: the data is the moat, not the model

Within a day of each other, two credit-trading platforms shipped AI that tells you exactly where fixed-income markets are heading - and they did it from opposite ends of the same idea.

Tradeweb launched TARA, a conversational assistant embedded in its institutional platform for US credit. Traders ask, in plain language, about trading activity, market flows, execution performance, liquidity and pricing - and TARA answers from Tradeweb’s proprietary historical and intraday data, plus its Ai-Price pricing engine. T. Rowe Price is an early adopter. The point isn’t the chat interface. It’s that the answers are grounded in data no competitor can replicate.

LTX, backed by Broadridge, launched agentic capabilities in BondGPT - the more aggressive move. Traders now build agents that monitor live market conditions and take action: generate alerts, create trade tickets, select dealers, launch RFQs, and accept prices to auto-execute. All under trader-defined parameters, with human-in-the-loop approvals, explainability before every action, and full auditability. Goldman, J.P. Morgan, Morgan Stanley, BofA and TD recently joined as integrated liquidity providers.

Put them side by side and you can see the whole direction of credit-market AI. Tradeweb proves the grounding thesis: the model is a commodity, the proprietary data is the defensible asset. LTX proves the autonomy thesis: AI is crossing from “answer my question” to “execute my trade.” Both are true at once, and both run on data the platform owns.

The takeaway: If you’re evaluating any AI trading or analytics tool, the sharpest question isn’t “how good is the model?” - models converge fast. It’s “what data is it grounded in, and can a competitor get the same?” That’s where durable advantage actually lives. And on the autonomy side, watch LTX’s own framing: auto-execution is live, but they expect traders to monitor before they trust. The binding constraint on agentic execution isn’t the technology. It’s how fast a desk is willing to let go of the button.

Moody’s turned 115 years of credit analysis into something any AI platform can call

Moody’s launched its first set of AI “skills” - and this one rewards a close read, because it’s the cleanest expression of where data providers are heading.

A skill is a packaged instruction kit that encodes Moody’s analytical frameworks and connects an AI agent to its proprietary ratings, research and risk intelligence. The first five cover real analyst workflows: earnings-call summaries, peer analysis, public information books, rating pitches, and sector analysis. They launch on Microsoft 365 Copilot Cowork first, but here’s the crucial part - they’re built on the open SKILL.md standard, which means the same skill runs on any compatible platform: Microsoft, OpenAI, Google, Amazon, Anthropic.

Think about what that actually does. Moody’s isn’t selling you a model, and it isn’t building a walled-garden app. It’s packaging its methodology and its proprietary data into a portable asset that travels to wherever you already work - and because the standard is open, that institutional knowledge isn’t locked to any one AI vendor. The strategic repositioning underneath is the real story: data providers are moving from “we have the data” to “we have the grounded reasoning that makes anyone’s AI output valid, explainable, and auditable.” Moody’s has already signalled credit analysis, due diligence, and insurance underwriting are next.

The takeaway: This is a preview of how specialised financial expertise gets distributed in an agentic world - not as a subscription you log into, but as skills that plug into your existing AI stack. If you’re a data or analytics provider, the question is whether you’re building skills or still building apps. If you’re a buyer, the question is whether your AI tooling can consume grounded skills from providers like Moody’s - because that’s about to become the difference between an agent that guesses and an agent that cites.

JPMorgan cut off Claude in Hong Kong- The reason wasn’t performance, and that’s the lesson

Every story above is about building on AI. This one is about what happens when the ground underneath shifts without warning.

JPMorgan stopped its Hong Kong staff from accessing Anthropic’s Claude models. The trigger wasn’t model quality or cost. It was contractual - the wording of Anthropic’s licensing terms, which exclude usage across Greater China including Hong Kong. JPMorgan simply removed Claude from the internal list of approved models for that region. Goldman Sachs made the identical move in April. The backdrop: days earlier, on June 13, Anthropic had suspended its frontier Mythos and Fable models worldwide for all foreign nationals after a US export-control directive.

For a practitioner, this is the most important story in the issue, even though it ships no product. Everything else here assumes the intelligence layer is stable - that the model you build your agent on today is the model you’ll have tomorrow. This is the reminder that it might not be. When your AI capability depends on a US frontier-model provider, access can be revoked by geography, by a licensing clause, or by an export directive - none of which you control, and any of which can land overnight. There’s a quiet irony worth noting: several of the banks restricting Claude regionally are reportedly lined up to lead Anthropic’s IPO. They’re underwriting the company and ring-fencing it at the same time.

The takeaway: Treat model access as a supply-chain risk, not a settled utility. The practical move is the one HSBC is already making - multi-vendor by design (Google and Mistral), so no single provider’s licensing terms or a government’s export policy can take a critical capability offline. If your AI roadmap has a single point of failure at the model layer, this month is your warning to build the hedge before you need it.

The Workflow

How an AI Agent Executes a Bond Trade - From Market Signal to Settled Trade

Corporate bond markets are fragmented and opaque. A trader watching for a specific condition - a bond hitting a target spread, a liquidity window opening - has to monitor continuously, then manually move through discovery, dealer selection, RFQ, and execution. By the time that chain completes, the opportunity is often gone.

An agentic trading workflow compresses that chain. But it’s worth being precise about how it differs from the agentic workflows we’ve covered before. An AML agent assembles evidence and hands it to a human to judge. A trading agent can be authorised to act. That single difference is why the entire architecture here is built around bounded authority and explainability - because when an agent can spend money, the guardrails aren’t a feature, they’re the product.

Walk through the diagram and the design logic becomes clear. The mandate comes first, and it’s human - the trader defines, in plain language, what condition to watch, what to do when it fires, and the hard limits on size and scope. Everything the agent does downstream is constrained by what’s set at this step. Then the agent takes over the part humans are worst at: continuous monitoring across a fragmented market, all day, without fatigue.

The pivot point is step three. When a condition fires, explainability is mandatory before any action. For a low-authority mandate, the agent stops and hands to the human. For a higher-authority one, it proceeds - but only after showing its reasoning, so the trader can see why it’s about to act, not just that it acted. From there the agent handles dealer selection and the RFQ - and this is where the proprietary-data moat from The Brief shows up in practice: intelligent counterparty selection only works if there’s deep, reachable liquidity in the network to act against.

Step five is the consequential one. If the returned price meets the trader’s parameters, the agent can accept it and auto-execute. This is the genuine “analysis to action” moment - live on the platform today, and the capability institutions will adopt most cautiously. Every step, finally, lands in a full audit trail: what fired, what the agent did, what price it took, against what limits. That’s what post-trade compliance and model-risk teams examine, and it’s built in rather than bolted on.

The problem that hasn’t been solved is trust velocity. The architecture works and auto-execution is live, but the binding constraint isn’t technical - it’s behavioural. How quickly will a buy-side desk move from monitoring the agent to trusting it to trade unsupervised? There’s also no public, third-party validation yet of latency, execution quality, or audit artifacts versus a human workflow. Until that exists, the highest-autonomy mode stays slow to adopt - and that caution is the correct instinct, not a failure of the technology.

The takeaway: If you’re assessing an agentic execution tool, don’t start with the model. Start by mapping exactly where the human authority boundary sits - step 1 (mandate), step 3 (explainability gate), and step 5 (execution). The quality of a trading agent isn’t in how much it can do autonomously; it’s in how precisely you can bound what it does, and how clearly it shows its work before it acts. A tool that makes those boundaries crisp is safer than a more capable one that blurs them.

The Toolkit

Weekly Coverage - The Broader Signal

Lloyds - The Readiness Number

Lloyds’ sentiment survey is the clean industry-pulse datapoint: 93% say AI/ML will have the biggest impact on UK financial services over the next five years, 91% expect AI investment to rise in the next 12 months, 77% call emerging-tech investment a growth priority. Backing it with action - 300 new tech hires for AI (per the Guardian), and £50m of AI value delivered in 2025 with £100m+ expected in 2026.

The signal: intent and budget are no longer the bottleneck. Grounding and execution are.

Broadridge + Anthropic, Project Glasswing - AI spend moves into defence

Broadridge joined Anthropic’s Project Glasswing, applying the unreleased Claude Mythos Preview model to harden cyber defences across critical financial software.

Use case: defensive security on infrastructure that underpins $15 trillion in average daily trading.

Why it matters is because it shows AI budget in finance flowing somewhere other than productivity and revenue - into cyber resilience.

UK Finance Fraud Report - the intelligence layer cuts both ways

The Annual Fraud Report is the sobering counter-signal to every build story this month: £1.28 billion stolen through payment fraud in 2025 (up 4%); APP fraud at £576.4m (up 19%); investment fraud the single biggest category at £221.5m, up 40% year-over-year. Two-thirds of APP fraud starts online. Criminals are industrialising with deepfakes, cloned voices, and synthetic identities at scale.

The same grounding-in-data that powers a trading desk powers a scam. The report’s core argument lands on the platforms where the fraud originates: banks can’t be the only line of defence.

Mastercard Agent Pay for Machines - the payments edition of the governance problem

Mastercard launched infrastructure for machine-to-machine payments: credentialing, programmatic spend limits, and guaranteed multi-rail settlement across cards, accounts and stablecoins, down to fractions-of-a-cent microtransactions. 30+ partners including Stripe, Adyen, Coinbase, Ripple, Solana Foundation.

The thread to this issue: as agents act, the payment rail needs its own trust layer - credentialed identity, enforced limits, full auditability. Same governance challenge, money-movement edition.

Experian Agent Operating System - the grounding problem, stated as a survey

Experian launched its Agent Operating System (June 2, Money20/20 Europe) within the Ascend Platform - a governed layer bringing Experian, client, and partner agents together for data, decisioning and control across the lending lifecycle (ServiceNow as first integration partner).

The datapoint that fits the theme: 48% of organisations say integrating data into AI workflows is the hard part; a third cite poor data lineage; a third cite siloed data. That is the intelligence-layer problem, in the industry’s own words - and it’s why this issue is named for it.

One question Before You Go

Everything in this issue points the same direction: the skill that mattered most five years ago - knowing how to do the analysis, is quietly being replaced by knowing how to direct, ground, and check an agent that does it for you.

One click. No follow-up unless you want one. We’re trying to understand where our readers actually are.

Everything here is built around one question: does it make adapting easier for you? The case studies, the worked examples, the expert conversations - that’s the whole point. I just wrapped one last week, and there’s an exciting one to announce soon.

See you soon!

Have a good rest of the week everybody.

Thanks,
Swati Sharma
Editor-in-Chief
Finance×AI Lab

We explore how AI is transforming finance, from adoption and practical use cases to governance, compliance, risk, and responsible innovation.

Supreet also shares insights from her book, The AI Optimization Playbook, offering a deeper look at how organizations can move beyond AI experimentation and build systems that are effective, scalable, and trusted.

This month, several major answers arrived simultaneously. Fiserv named its answer agentOS. Citi named its Arc. FIS embedded Anthropic engineers to co-build the architecture from the inside. Anthropic itself shipped ten production-ready templates with the data connectors already attached. And Blackstone, Goldman Sachs, and Hellman & Friedman backed a $1.5 billion firm whose entire purpose is to install this infrastructure inside companies that can’t build it themselves.

Every one of these moves is solving the same problem. Not whether AI agents work -that question is largely settled.

The problem is the layer underneath: governance, auditability, human oversight, and the ability to stop an agent mid-task when something goes wrong.

In a week where the IMF warned that AI is turning cyberattacks into potential systemic financial risks, that last capability isn’t a feature. It’s the thing regulators will ask about first.

The vocabulary has shifted. Agent is no longer the product; the platform is the product. And every institution that can’t build its own is now buying someone else’s.

- The Editor

The Brief - Things Worth Knowing

Fiserv launched agentOS - the first governed agent marketplace built natively for banking. The governance story is the lead, not the AI story.

What Happened

Fiserv launched agentOS on May 14 - an agentic AI operating system running on AWS Bedrock AgentCore, built natively across its core, payments, issuer processing, and servicing platforms.

Six financial institutions co-developed it; two are live in beta today. Broad availability August 2026. The marketplace launches with four Fiserv-built agents - Commercial Loan Onboarding, Daily Operational Analysis and Reporting, Agentic Deposit Intelligence, and Agentic AML Triage Analysis, and nine third-party agent partners covering risk, compliance, disputes, and reconciliation. OpenAI is a strategic collaborator, with select first-party agents developed jointly.

Early proof points: First Interstate Bank compressed commercial loan onboarding; Boulder Dam Credit Union cut report generation from 10 minutes to seconds.

Why It Matters

Fiserv's pitch isn't "here is a powerful agent." It's rather "here is the only place where you can run agents, build agents, and deploy third-party agents - all under the same identity controls, policy framework, and audit trail." That's an OS claim, not a feature claim. And it's a credible one for any institution whose core already runs on Fiserv.

The proof points are real but modest - loan onboarding compression and faster reporting are production outcomes, not modelled projections. The marketplace model is the more interesting long-term bet: if third-party agents can be deployed, governed, and audited within the same architecture at scale, Fiserv becomes infrastructure rather than just a vendor.

Nine third-party agent partners at launch is a start. If that grows to 50+, agentOS becomes the App Store for banking AI. If it stays at single digits, it's a managed services extension. That trajectory is worth watching more closely than the launch announcement.

FIS embedded Anthropic engineers inside its walls to build a Financial Crimes AI Agent.

What Happened

FIS, the financial technology company, powering nearly 12% of the global economy - announced on May 4 that it is building a Financial Crimes AI Agent with Anthropic's forward-deployed engineers embedded inside FIS. The agent will compress AML alert and case investigations from days to minutes, automatically assembling evidence across a bank's core systems, evaluating activity against known typologies, and surfacing the highest-risk cases for investigator review.

BMO and Amalgamated Bank are in development; general availability planned for H2 2026. Client data stays within FIS-controlled infrastructure throughout. The roadmap beyond financial crime spans credit decisioning, deposit retention, customer onboarding, and fraud prevention.

Why It Matters

The embedded-engineer model is the detail worth sitting with.

Anthropic isn’t licensing the technology and sending a contract. Its engineers are physically inside FIS - co-designing the agent, learning the regulatory infrastructure, and building the evaluation frameworks FIS will use to develop future agents independently. That’s a knowledge transfer model, not a vendor relationship. It’s also how trust gets built in a regulated domain: not through a model card, but through co-design.

The AML framing is strategic - US institutions spend $35–40 billion annually on AML operations, and investigators spend the majority of that time assembling evidence before any analysis begins. The agent attacks the assembly step. The judgment step stays with the human.

Amalgamated Bank’s quote is the most credible signal in the announcement: “Our teams are contributing to the design of the agent, not just deploying the technology.” When a bank’s Financial Crimes Compliance team is embedded in the design process, the resulting agent is more likely to reflect real investigation workflows. When they’re not, it usually isn’t.

FIS serves as the system of record for transactions, payments, deposits, credit, and customer activity across thousands of financial institutions. If the Financial Crimes AI Agent works at BMO and Amalgamated Bank, it gets deployed across the FIS client base with relatively minimal friction. The blast radius of a successful deployment here is considerably larger than most comparable announcements.

Anthropic shipped 10 financial services agent templates with the data plumbing already attached.

What Happened

On May 5, at an invite-only financial services briefing in New York, attended by JPMorgan CEO Jamie Dimon among others, Anthropic launched ten ready-to-run agent templates covering the most labor-intensive workflows in banking, asset management, and insurance.

• Five are research and client coverage: pitch builder, meeting preparer, earnings reviewer, model builder, market researcher.

• Five are finance and operations: valuation reviewer, GL reconciliation, month-end close, statement auditor, KYC screening.

Each template packages skills, connectors, and subagents into a single reference architecture. Deploy as plugins in Claude Cowork or Claude Code, or as cookbooks for Claude Managed Agents. Alongside: eight new financial data connectors - Moody's MCP app (600M+ companies), D&B, Verisk, SS&C IntraLinks, IBISWorld, Guidepoint, Third Bridge, Financial Modeling Prep - plus expanded Microsoft 365 integration across Excel, PowerPoint, and Word. Claude Opus 4.7 was also announced, leading the Vals AI Finance Agent benchmark at 64.37%.

Why It Matters

The templates are the attention-grabbing announcement. The connectors are the more significant one. The biggest practical barrier to deploying financial AI agents hasn't been model quality - it's been the data plumbing. Connecting an agent to Moody's, D&B, and IntraLinks without building custom integrations removes a meaningful bottleneck for teams that want to prove out a use case before committing to platform infrastructure. These templates are starting points, not production deployments. But they now start considerably closer to the finish line.

A 64% benchmark score on financial agent tasks means something specific: roughly one-third of tasks still fail or require human correction. That's not a reason not to use them - it's a reason to be precise about which workflows they're deployed into and what the human review layer looks like. The institutions that figure out that boundary will outperform the ones treating templates as production solutions.

Blackstone, Goldman Sachs, and Anthropic formed a $1.5 billion AI services firm.

What Happened

On May 4, Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs announced the formation of a new AI-native enterprise services firm, backed by approximately $1.5 billion in committed capital, with additional backing from General Atlantic, Apollo, Leonard Green, GIC, and Sequoia. The firm is a standalone entity with Anthropic engineers embedded permanently inside it.

Its initial client base is the portfolio companies of the investment firms - giving it a built-in pipeline across hundreds of businesses. The model mirrors Palantir's forward-deployment architecture: implementation capability combined with ownership of the underlying model. Blackstone President Jon Gray named the bottleneck directly: "one of the most significant bottlenecks to enterprise AI adoption" is the scarcity of engineers who can implement frontier AI systems at speed.

Why It Matters

This is how foundation model capability reaches mid-market companies that don't have engineering teams to build their own Arc or FIS partnership. The agent is Anthropic's. The implementation is the new firm's. The enterprise gets the outcome without building the capability. It also puts Anthropic into direct competition with major consulting firms for AI transformation work - which Fortune noted explicitly. The more interesting question is whether permanently embedded engineering produces meaningfully different outcomes than advisory relationships. That's what this firm is designed to test.

Anthropic CFO Krishna Rao's framing is the most useful: "Enterprise demand for Claude is significantly outpacing any single delivery model." That sentence is the rationale for the entire structure. It's also a candid admission that frontier AI providers cannot reach enterprise at scale through direct sales alone. The implications extend beyond this deal.

The Workflow

US institutions spend $35–40 billion annually on AML operations. The UN estimates $2 trillion in illicit funds moves through the global financial system each year. Despite that investment, investigators spend the majority of their time not analysing, but assembling evidence across disconnected systems before any analysis can begin. That's the step the agent attacks.

The Architecture - Step by Step

The alert fires. A case opens and is assigned — this step is not automated, and the agent does not decide which alerts become cases. Alert-to-case escalation involves judgment about threshold calibration and resource allocation that institutions are not ready, and regulators are not prepared, to fully automate.

From case opening, the agent takes over the evidence assembly. It connects to the bank’s unified data environment — full transaction history, customer profile and onboarding documentation, prior alert and disposition history, correspondent bank records, and third-party watchlist results including OFAC, PEP, and adverse media. In the FIS/Anthropic architecture, this all happens within FIS-controlled infrastructure; no data leaves the governed environment. Every source is cited in the output. What previously took 2–4 hours now takes minutes.

The agent then evaluates the assembled activity against a library of known AML typologies - structuring, layering, smurfing, trade-based money laundering, round-tripping, and others. The output is an explanation, not a score: which patterns match, which don’t, and why. SAR narrative quality is expected to improve as better evidence assembly produces better narrative. Honest limitation: typology libraries reflect known patterns. Emerging methods - AI-generated transaction obfuscation, novel crypto-fiat layering - may not yet appear in them. Novel patterns still require human pattern recognition.

High-risk cases surface first. The investigator receives the complete, sourced evidence package alongside the typology-matched risk assessment and a draft SAR narrative where warranted. The human reviews, challenges, or confirms. Decision authority remains entirely with the investigator - including SAR filing decisions. This is not a design choice. It is a regulatory requirement. Every step of the agent’s work, including the investigator’s decision and any documented override, is logged in a full audit trail.

The Problem That Hasn’t Been Solved

The agent compresses evidence assembly. It doesn't fix alert volume. If the false positive rate remains unchanged, the efficiency gain is real but partial. The next frontier is better prioritisation before the case opens - smarter triage before the investigation begins. No system has fully solved that yet, and it is where the next generation of AML infrastructure will be built.

Auditability and ongoing performance evaluation are not the same thing. Every platform in this issue builds the trail. Almost none of them have systematic frameworks for using that trail to improve the agent over time. That gap is where the next wave of agent governance tooling will emerge — and where the regulatory questions will land in 2027.

What Happened This Month

Broadridge says agentic AI is live in production across 40+ clients.

Broadridge Financial Solutions announced on May 11 that its agentic AI capabilities are live in production across capital markets and wealth management workflows - with more than 40 clients since 2024, processing millions of operational transactions monthly across post-trade, account management, and client services.

Capabilities in production include automated trade fails management and break resolution, account opening and maintenance, real-time valuation exception handling, customer inquiry automation, and email workflow processing. New clients can achieve up to 30% Day 1 operational cost reduction across two deployment paths: full managed services where Broadridge runs operations end-to-end, or standalone deployment of Broadridge's agentic platform into the firm's own infrastructure.

OpenAI and PwC are building AI agents for the CFO office - with OpenAI’s own finance team as the test subject.

PwC and OpenAI announced on May 5 an expanded collaboration to help enterprises build AI-native finance functions. The structural detail worth noting: OpenAI's own finance team is acting as "customer zero" - using agents across investor relations, treasury, tax, reporting, corporate development, and contract review, then taking those learnings to enterprise clients through PwC's implementation capability.

Internal production metrics rather than projections: Codex processed 5x more contracts with the same headcount; IR-GPT managed 200+ investor interactions during a recent fundraise. PwC's role is the implementation muscle - the enterprise gets the output without building the engineering capability.

The IMF warned that AI is turning cyberattacks into a potential systemic risk for the global financial system.

In a blog post on May 7, the IMF warned that AI is making cyberattacks faster, cheaper, and more scalable, and that extreme cyber incidents could create funding strains, raise solvency concerns, and disrupt broader markets. The structural argument is the one worth sitting with: as financial institutions standardise on the same AI infrastructure providers, their attack surfaces become correlated. A vulnerability in a shared platform is a shared vulnerability.

The IMF cited Anthropic's Claude Mythos Preview as an illustration, a model found capable of identifying and exploiting vulnerabilities across major operating systems and browsers, even in the hands of non-experts. The IMF's framing is significant: this is not an IT security concern. It is a macro-financial stability concern, and one that supervision frameworks haven't caught up to.

The Tool-kit

Three Models for Deploying Governed AI Agents -Which Architecture Fits Your Institution?

Fiserv built a banking OS. Citi built its own. Anthropic shipped templates. These aren't competing products - they serve different positions on the deployment maturity curve.

The Resource Shelf

READ
“Every bank wants AI that acts, not just assists” - FIS CEO Stephanie Ferris

The full FIS primary announcement is worth reading directly. The language around governance architecture - data staying within FIS-controlled infrastructure, every agent decision traceable, human authority over all decisions - is the clearest articulation of what regulated AI deployment actually looks like in practice.

WATCH
Can AI be trusted to make financial decisions? Agentic AI pushes deeper into banking

Watch this discussion on whether financial institutions are actually ready for AI that acts, not just analyses - covering explainability gaps, liability, hallucinations, and what governance needs to look like before agents start getting closer to moving money.

SAVE
Fortune: Anthropic takes shot at consulting industry - the Blackstone-Goldman venture explained

The Fortune piece is the clearest account of why the Blackstone-Goldman-Anthropic venture matters beyond the funding announcement. The "permanently embedded engineers" model and its implications for consulting firm displacement are worth having in your reading pile before the next internal AI strategy conversation.

Before You Go

Every deployment in this issue is solving the same problem: how do you run AI agents at enterprise scale, inside a regulated institution, with governance a regulator can actually examine?

The IMF's warning is the other side of that coin - the same concentration that makes governed agent infrastructure powerful makes the attack surface correlated. A vulnerability in a shared platform is a shared vulnerability. Both things are true simultaneously. Building is the right move. Building carefully, with governance that extends beyond the audit trail to systematic performance evaluation, is what separates the institutions that will look good in 2027 from the ones that won't.

The next issue is already in the making. Same lens, different workflows; with the stories that are worth your time.

Also, great news!

We’re coming up with the first expert-led webinar on Finance×AI Lab. We’d love for you to be on the list.

See you then.

Swati Sharma
Editor-in-Chief
Finance×AI Lab

It is no longer a metaphor.

From Visa’s C-suite to the Bank of England’s stress tests, the industry is realizing that AI has moved from a “feature” to a foundation.

It’s now embedded inside the rails: Citi is putting it in client relationships, BlackRock is giving staff an “AI factory,” and the card networks are redesigning commerce for machines.

But as deployment hits escape velocity, governance is still on the launchpad. The gap between what we can build and what we can control is the story of this issue. And it is getting wider.

Welcome to this ‘trustworthy’ issue of the Finance×AI Lab, pun intended.

- The Editor

The Brief - Things Worth Knowing

Regulators are moving. Just not as fast as the deployments they’re trying to govern.

What Happened

On April 1, the Bank of England and PRA confirmed they are holding their technology-agnostic framework - no new AI-specific rules, no sandbox, approach under continuous review.

The harder news: the PRA has made AI adoption a named 2026 supervisory priority. Expect direct questions on governance, model risk, and oversight in your next regulatory dialogue. No new rulebook, but the scrutiny is now formal.

Three weeks later, the FCA announced eight firms - Barclays, Experian, Lloyds/Scottish Widows, UBS among them, for its second AI Live Testing cohort. Use cases span targeted investment support, agentic payments, AML, and KYC. Testing runs through end-2026; a Good and Poor Practice report follows, with a full evaluation in Q1 2027. Applications to the FCA’s broader Regulatory Sandbox and Innovation Pathways were up 49% year on year.

Why It Matters

The BoE and PRA are not writing rules but embedding AI governance into private supervisory conversations, firm by firm. That means there’s no public signal for everyone else to calibrate against. The FCA’s forthcoming Good and Poor Practice report will be the first substantive indication of where the lines are being drawn. That report is worth more than any guidance document that doesn’t exist yet.

Across Europe, the Dutch AFM warned twice in April that AI adoption in asset management is outpacing governance and could distort price formation. ESMA’s chair flagged AI as accelerating the speed and scale of cyber threats in financial markets. India’s Finance Minister Sitharaman told banks to harden cyber defences against AI-borne threats on April 25. The regulatory response is global, simultaneous , and consistently running behind the deployments it’s trying to govern.

The regulatory posture across every major jurisdiction right now is a variant of the same sentence: existing rules apply, we are watching, and we will tell you what good looks like after we've seen enough of it. That is not a compliance framework. That is a observation programme. The firms building AI governance now - ahead of any guidance - won't just be compliant. They'll be the ones the regulators call when they're writing the guidance.

Citi built an AI team member. BlackRock gave its employees an AI factory. The wealth management arms race just got serious.

What Happened

Citi has been laying track before announcing the train. In Q1, CitiScribe - AI-powered note-taking, rolled out to all North American wealth advisors.
Client 360 (unified advisor dashboard), AskWealth CIO (conversational access to CIO research), and Portfolio Intelligence (client-facing portfolio insights for Citi Private Bank) are all in pilot or limited rollout, with broader expansions planned.

Then on April 22, Citi announced Citi Sky - an always-on AI team member built on Google Cloud and Google DeepMind’s Gemini Enterprise Agent Platform, launching to Citigold clients this summer. Live audio, video, avatar interface. English and Spanish at launch. Andy Sieg called it the shift from interface to intelligence, from transactions to outcomes.

Separately, as reported by the Wall Street Journal, BlackRock rolled out RockAI - a no-code internal platform letting its 5,000 developers, and eventually non-technical staff, build and deploy custom AI agents across investment and operational workflows. Pre-built guardrails, natural language interface, agents live in minutes.

Why It Matters

The Citi sequencing is the detail worth sitting with. CitiScribe wasn’t just a documentation tool. A quarter of structured advisor-client conversation logs across every North American wealth advisor is an extraordinary foundation for any client-facing AI that comes next. Whether or not that’s the stated objective, the sequencing of CitiScribe in Q1 and Sky in summer is not accidental.

RockAI is the more underreported story. Giving non-technical employees the ability to build and deploy their own AI agents isn’t a productivity initiative - it’s a restructuring of how a $10 trillion asset manager uses human capital. BlackRock’s AI executive sponsor envisions AI as the default mode for most processes, with humans working in cross-functional squads to oversee the output. That’s not a vision statement. That’s a org chart in progress.

Visa and Mastercard just rewrote the rules for who - or what - can be a customer. The answer is: an AI agent.

Visa’s B2AI report - 2,000 consumers, 512 business decision-makers, conducted with Morning Consult, put numbers on what the industry has been sensing. 53% of businesses would let AI agents negotiate directly with other AI agents. 77% are already using or piloting AI. 71% would optimise products specifically for AI agents. But consumers are more cautious: only 38% are comfortable with fully autonomous purchases, and 60% want approval on every transaction.
Trust in bank-backed AI: 36%. Independent AI agents: 28%.

That gap is everything.

On April 6, Mastercard announced the ASEAN rollout of authenticated agentic transactions - building on its first live Singapore transaction in early March with DBS and UOB, and Malaysia pilots with CIMB, RHB, and UOB - via Agent Pay: Agentic Tokens, Payment Passkeys, and Verifiable Intent co-developed with Google. Two days later, Visa launched Intelligent Commerce Connect - a single integration point for merchants, AI agents, and enablers, supporting multiple protocols across networks and token vaults.

Why It Matters

Commerce is being redesigned for machine intermediaries. Brands are no longer just marketing to people - they are preparing infrastructure for AI agents that will evaluate, select, and transact on people’s behalf. Consumers will accept it, but only when a trusted financial institution is in the loop. That one data point - bank-backed AI trusted at 36%, independent agents at 28%, is the reason both networks are moving this aggressively. The trust layer belongs to to the early builders.

The stat nobody is quoting: 40% of Americans have already made a purchase they wouldn't have considered without an AI agent's intervention. Commerce isn't just being automated- it's being expanded. AI agents aren't replacing human browsing. They're generating demand that didn't exist before the agent showed up. That changes the revenue model for every institution that touches commerce infrastructure. Not eventually. Now.

The Workflow

Both Visa and Mastercard have been talking about agentic payments for two years. This month, both went live. And the architectures they've built to make it work are more thoughtful than most of the coverage suggests - because the hard problem isn't getting an AI agent to initiate a transaction. That's trivially easy. The hard problem is making sure the transaction reflects what the human actually wanted, is protected against fraud, and creates an audit trail that a regulator can examine. That problem is significantly harder, and how each network has solved it tells you something important about where payment infrastructure is going.

The Architecture - Step by Step

Both frameworks start in the same place: explicit consumer authorization. The user sets the scope - purchase category, spend limit, time window, when the agent acts independently and when it asks first. Everything else is built on top of that consent layer.

From there they diverge. Mastercard’s Agent Pay issues a unique Agentic Token per AI agent - not a stored card number, not a static credential, but a token mapped to a specific authorized use case. Payment Passkeys authenticate the consumer’s identity at execution. Verifiable Intent, co-developed with Google, creates a tamper-resistant record of what was authorized - a shared source of truth across consumers, merchants, and issuers. In the first live deployment, an AI agent booked a ride to Changi Airport via DBS and UOB. The architecture held.

Visa’s Intelligent Commerce Connect takes a different approach. ICC is a single integration endpoint supporting multiple agent protocols simultaneously - TAP, MPP, ACP, UCP - across networks and token vaults, including non-Visa cards. The Trusted Agent Protocol, built with over ten partners, lets merchants verify an agent’s purchase intent before checkout proceeds. Consumer mandates are validated throughout.

The Problem That Hasn’t Been Solved

Both frameworks solve authentication and authorization. Neither solves liability. If an AI agent makes a purchase the consumer disputes - not fraud, but a misread mandate - who is responsible?
The consumer set the parameters. The agent acted within them. The outcome wasn’t what was intended.

That gap doesn’t have a clean answer yet in either architecture.

What Happened This Month

Oracle announced AI-driven financial crime investigation is coming to its compliance stack, and extended agentic AI into corporate banking.

On April 9, Oracle announced AI-driven investigation capabilities - via Lucinity’s technology - would be available within its financial-crime compliance portfolio within 12 months, targeting the most manual part of AML: alert investigation. On April 14, it extended its agentic AI platform into corporate banking - treasury, trade finance, credit, and lending. The logic is the same as JPMorgan’s AML system: the value isn’t detecting more, it’s removing the manual burden from the analysts acting on what’s detected.

European regulators flagged two different AI risks in April. Neither has a clean solution yet.

The Dutch AFM warned twice in April that AI adoption in asset management is outpacing governance and could distort price formation - the concern being correlated model behavior across firms amplifying market moves that no single institution intended. On April 24, ESMA’s chair told Reuters that AI is increasing both the speed and scale of cyber threats in financial markets.

OpenAI acquired Hiro Finance. Two personal finance acqui-hires in six months. The pattern is the strategy.

Hiro Finance, backed by Ribbit Capital and General Catalyst, shuts down April 20. Founder Ethan Bloch and the team join OpenAI. Product gone, capability absorbed. This is OpenAI’s second PFM acqui-hire in six months after Roi in October 2025.

What they’re buying isn’t a product. It’s the expertise to know exactly where general-purpose LLMs fail when money is on the line - specifically, Hiro’s accurate financial math and user-controlled verification layer.

The Tool-kit

AI Personal Finance Tools: Who Survives OpenAI’s Entry?

The Resource Shelf

READ
The FCA’s AI Live Testing cohort - what the use cases actually tell you

The choice of eight firms and their specific use cases is more revealing than the announcement. FStech’s analysis of who got in and why is the cleanest read on where the FCA thinks the highest-risk AI deployments currently sit.

WATCH
Visa’s B2AI framework - “Commerce is moving from market-to-human to market-to-machine”

The full B2AI report is worth reading primary. The 53% AI-to-AI negotiation stat and the consumer trust breakdown by institution type are the two numbers that will shape commerce infrastructure strategy for the next three years.

SAVE
TLT LLP - BoE and PRA AI plans: what firms need to know

The cleanest plain-English summary of what the April 1 BoE/PRA letter actually means for your supervisory relationship this year. Save before your next regulatory prep conversation.

Before You Go

Step back from the individual stories and look at what they have in common. Every deployment in this issue - Citi Sky, RockAI, Mastercard Agent Pay, Visa ICC, Oracle’s agentic compliance stack - shares one architectural feature: the human is still nominally in the loop, but the loop is getting longer.
The AI is doing more. The human is approving more abstractly. The governance frameworks are trying to keep up.

The institutions that will navigate this well are not the ones that move fastest. They are the ones that are building the infrastructure to make AI accountable at the same time as they are building the infrastructure to make it capable. Those two things are not in conflict. But most institutions are treating them as if they are - building the capability first and planning to bolt on the accountability later.

That approach will work fine, right up until it doesn’t.

The next issue is in the making; more next week when the stories are worth your time.

If something in this issue landed differently than expected - if you disagree, if you want us to go deeper, if there’s a story we should be covering - reply and tell us. Every response shapes what we cover next.

See you then.

Swati Sharma
Editor-in-Chief
Finance×AI Lab

The productivity gains from AI in finance are real. They are just not where anyone is looking.

Something broke the spell last month.

For the past two years, every finance conference, every LinkedIn post has carried the same subtext: AI is coming, and it's going to change everything. Most of us nodded along- half convinced, half waiting for the other shoe to drop.

This now, both shoes dropped at once.

Goldman Sachs moved AI agents into live production- not a pilot, not a proof of concept- handling trade accounting and compliance autonomously. And in the same breath, Goldman’s own research team published a note saying there’s still no measurable AI productivity gain at the economy-wide level.

Both of those things are true. And that tension is exactly what this newsletter exists to sit inside.

Because here’s what I think most of us have been quietly anxious about: we don’t know whether we’re early, late, or just watching expensive experimentation with someone else’s money. The noise is deafening and the signal is thin. Every vendor promises transformation. Every internal AI initiative seems to be either stuck in procurement or still in a sandbox three quarters later.

What’s actually changing, and I mean changing in production, with numbers attached- is narrower and more specific than the headlines suggest.

Let’s get into it.

The Brief - Things Worth Knowing

JPMorgan’s AI cut AML false positives by 95%. The lesson isn’t what you think.

Everyone heard the number. Almost nobody drew the right conclusion from it.

The instinct when you see 95% false positive reduction is to think: great, the AI is catching more fraud. But is that really what happened?

Absolutely not.

The AI isn’t catching more fraud - it’s removing the noise that was preventing humans from catching fraud in the first place. Compliance investigators were spending 80–90% of their time chasing alerts that turned out to be nothing. The AI didn’t replace them. It finally let them do their jobs.

That reframe matters enormously if you’re evaluating AI for any compliance workflow. The question then wouldn’t be “can it detect more?”, rather “can it remove enough noise that your best people stop wasting their time?” Those are completely different procurement conversations, different implementation architectures, and different ROI calculations.

The bank estimates $1.5 billion in total AI-related value; fraud prevention is the single largest contributor.

The EU AI Act's August 2026 deadline is 120 days away. Most financial firms are not ready.

Here’s the uncomfortable truth about where most financial institutions actually are right now: they’re waiting. Waiting for the December 2027 extension that a Digital Omnibus proposal floated in late 2025. Waiting for someone else to go first and absorb the regulatory friction. Waiting for their vendor to sort it out.

None of those are safe bets.

The European Commission has not confirmed the extension. The EBA has explicitly stated it will begin supervisory activities in 2026–2027 for AI in banking and payments. And there is a distinction buried in the Act that most compliance teams have not fully absorbed: if you deployed a third-party vendor’s AI tool, you are still the deployer. The vendor’s conformity assessment does not cover you. “We outsourced it” is not a defence.

As of August 2, 2026, any AI system used for credit scoring, loan approval, fraud detection, or AML risk profiling within EU markets is classified as high-risk under the EU AI Act. That means mandatory conformity assessments, documented human oversight, full auditability, and registration in the EU AI database - whether you built the system or bought it from a vendor. Non-compliance fines reach up to 6% of global annual turnover.

The institutions treating the extension as guaranteed are making the most expensive assumption in their compliance roadmap right now. The cost of preparing early is a fraction of the cost of being wrong.

25 million Americans have no usable credit score. AI is fixing that - but it has a compliance pothole.

The problem has always been clear. FICO needs a credit history to generate a score. If you’ve never had a mortgage, a credit card, or an installment loan, the model has nothing to work with. You’re not a bad risk. You’re an invisible one.

And for decades, the answer was: declined, or priced punitively.

What’s changed is that open banking data - 12 months of transaction history, income patterns, spending behaviour, rent payments - turns out to be a richer signal than a FICO score ever was.

• Upstart uses over 1,600 variables versus the 15–20 in conventional scoring. The outcomes: 27% more approvals, 16% lower default rates.

• Zest AI automates 60–80% of lending decisions with a 20% reduction in charge-offs. The architecture works.

The CFPB now requires model-specific, applicant-level explanations for every adverse credit decision. Not “the model scored you low.” Specifically why, in terms the applicant can act on and a regulator can audit. At scale. “The algorithm decided” is explicitly non-compliant.

Every one of the leading platforms - Upstart, Zest AI, Scienaptic - has this gap. None of them have fully solved explainability at the volume and specificity regulators now require.

The Workflow

How AI Is Underwriting the Borrowers That FICO Forgot

Let me be direct about something before we get into the architecture.

The “25 million unscored Americans” framing gets used a lot in fintech pitch decks. It’s real, but it undersells the actual problem. The deeper issue is that FICO was never designed to assess financial behaviour - it was designed to assess credit behaviour. Those are not the same thing. Someone who pays rent on time every month for ten years, manages a household budget carefully, and has never missed a utility payment is invisible to FICO. Not because they’re a bad borrower. Because they never needed to borrow.

Open banking data fixes the input problem. LLMs fix the interpretation problem. And that combination is now in production - not in pilots, not in whitepapers.

Here’s how it actually works:

What this means really?

The technology works. That’s no longer the hard part.

The hard part is explaining the decision- specifically, why a particular applicant was declined, in terms that satisfy three different regulatory frameworks simultaneously: CFPB, EU AI Act, and ECOA. At scale. For every single decision.

Nobody has fully built that yet. Whichever team does will own the most important piece of lending infrastructure of the next decade.

Looking Back

9fin raised $170M. The reason why tells you everything about where AI in finance actually creates value.

There’s a version of this story that gets told as a funding announcement. That’s not the interesting version.

The interesting version is this: $145 trillion in global debt capital markets is still being run on PDFs, emails, and data rooms. Credit analysts, some of the most expensive professionals in finance, are spending hours manually parsing legal documents, covenant packages, and bond prospectuses. Not because they want to. Because no one had built the infrastructure to do it any other way.

9fin built that infrastructure. Their AI ingests the documents, structures the data, and lets credit professionals identify opportunities, analyse risk, and monitor markets in one place. The company has delivered 100% ARR growth for multiple consecutive years. Canada Pension Plan Investment Board was a client before it became an investor in this round- which is the most honest endorsement a product can receive.

The $170M Series C at a $1.3 billion valuation, led by HarbourVest, isn’t a bet on AI in the abstract. It’s a bet on a specific thesis that this newsletter will return to repeatedly: proprietary data built first, AI layer built second. The platforms getting funded at scale right now are the ones that understood that order of operations from the start.

Every credit team still running on PDFs and emails is essentially subsidising their competitors’ advantage. The infrastructure gap in debt markets is closing faster than most participants realise.

Goldman Sachs just made agentic AI real. Not as a pilot. As production infrastructure.

For the past two years, “agentic AI” has been the most overpromised term in enterprise technology. Demos at conferences. Pilots that never shipped. Announcements that quietly became nothing.

Goldman Sachs just changed that conversation.

The bank has moved AI agents into live production across trade accounting, client onboarding, and compliance workflows- built on Anthropic’s Claude, with Anthropic engineers embedded inside Goldman’s technology teams for six months to co-develop the system. These are not chatbots. They are not copilots. They handle transaction reconciliation, trade break resolution, and client vetting autonomously, with human oversight reserved for exceptions.

The reason this matters beyond Goldman is what it signals about the threshold that’s just been crossed. Trade accounting, KYC processing, client onboarding- these workflows have resisted automation for decades precisely because they require reading large volumes of unstructured data against strict regulatory frameworks and making judgment calls. If that’s now in production at one of the most risk-conscious institutions in global finance, the door is open for every institution operating below Goldman’s risk tolerance.

66% of banks say AI is a top strategic priority. The number that actually matters is 44%.

Every industry survey on AI right now leads with the optimistic headline. This one is more useful if you read it from the bottom up.

Yes, 66% of banking professionals say AI is at least a high strategic priority. Yes, most institutions increased AI spending by 10% or more last year. National banks are leaning in hard, 63% say it’s their number one or top-few priority.

But credit unions? 44% rate AI as moderate to low priority. And that gap is the most consequential number in the entire survey.

Because what that gap actually represents is a competitive moat being constructed in real time. The institutions treating AI as a moderate priority in 2026 are not standing still- they’re falling behind institutions that are compounding their advantage every quarter. By the time the 44% decide it’s urgent, the 63% will have 18 months of production data, trained models, and institutional muscle memory that can’t be bought off the shelf.

There’s one more thing buried in this survey worth flagging. The top justifications for AI investment are productivity improvements and workflow automation- not revenue generation, not customer experience. That tells you exactly where the ROI conversation is happening right now: internal efficiency, not growth.

The institutions that start measuring revenue impact in 2026 will have a fundamentally different story to tell their boards in 2027.

The Tool-kit

Three Tools, Same Problem: Which AI Underwriting Platform Is Actually Right for You?

Upstart, Zest AI, and Scienaptic all claim to solve the same problem- better credit decisions for borrowers that traditional models fail. They are not the same product. Here is what actually differentiates them.

Bottom Line

• If you want speed and are comfortable with a partner-model approach: Upstart.

• If you want to own the model and have the engineering capacity: Zest AI.

• If you are a traditional institution modernising incrementally without ripping out your stack: Scienaptic.
  

  None of them have fully solved explainability at the volume and specificity that regulators now require. That is the gap in every one of these platforms right now.

The Resource Shelf

READ
More than a third of consumers are now using Claude and ChatGPT for investment guidance

McKinsey’s Vlad Golyk puts hard numbers on what advisors have been sensing for months. The question it raises, ‘what exactly are you charging for?’ is one the industry isn't ready to answer.

WATCH
BlackRock’s Larry Fink warns AI may intensify wealth inequality

Fink’s annual shareholder letter doesn’t usually double as an ethics brief. This year it does. His argument is simple and worth sitting with: AI will generate enormous value, and that value will flow disproportionately to those who already own financial assets - at an even larger scale than before. The argument is simple, uncomfortable, and worth hearing directly from the source.

SAVE
Goldman Sachs: No economy-wide AI productivity link- but 30% gains in 2 specific use cases

Save this one for the next time someone in your organisation asks you to justify an AI investment, or questions why the last one didn't move the needle. Goldman's own data shows that economy-wide, AI productivity gains are still unmeasurable. Don’t forget to read about the two use cases which are the story.

That's everything for this issue. Take a moment with what you've read- not the individual stories, but the pattern underneath them. The institutions winning right now didn't buy a platform and wait. They picked a workflow, scoped it tightly, and measured honestly. That's a replicable playbook, and it's available to anyone willing to do the unglamorous work.

I'm already pulling together the stories for next issue. See you then.