In a single week this May, Anthropic launched ten pre-built agents for financial services covering credit memos, underwriting reviews, KYC files, pitchbooks and earnings analysis, announced a data partnership with Moody’s, and unveiled a $1.5 billion joint venture with Blackstone, Apollo, Goldman Sachs and others to embed Claude across mid-market companies. The same week, FIS shipped a financial-crimes agent built on Claude, with BMO among the first deployments, and OpenAI announced a CFO-suite partnership with PwC, followed days later by an agent platform with Fiserv aimed at thousands of banks. Goldman Sachs, JPMorganChase and Citi are named production customers. Financial services is now Anthropic’s second-largest business.
The question in front of every bank board is no longer whether to adopt frontier AI. Adoption is settled. The open question is value capture: when intelligence this powerful is applied to banking, who keeps the durable value? The uncomfortable answer, on current trajectory, is mostly not the banks.
I. The value capture problem
Banking is the judgment business, but be precise about what that means. The franchise assets set the size of the game: the license, the deposit base, the balance sheet, the client relationships. Judgment decides who wins it. Two banks can hold similar books at similar margins and post very different returns through a cycle, and the difference is credit costs, deal selection, and the blow-ups avoided. That difference is the bank’s real intellectual property. Not the kind you can register, but operating know-how: the credit officer who can tell which covenant breach is noise and which is the first crack, the syndicate desk that knows how to build a book when the market is nervous, the workout veteran who has seen three cycles and knows when to extend and when to enforce. None of this is written down anywhere useful. It surfaces only as decisions. Approvals, exceptions, overrides, escalations, walk-aways.
Decisions are exactly what a model learns from. That is the whole design. And here is the tension nobody has priced: AI creates value by making your private know-how usable at scale, and in the same motion makes that know-how less scarce. If two hundred lenders screen borrowers through the same model, the best underwriter loses the spread between its process and the market’s. It gets worse. When everyone prices credit with the same intelligence, the loans you win at the thinnest spread are disproportionately the ones the shared model has mispriced. Lenders have a name for that. It is the winner’s curse, and pooled underwriting judgment institutionalizes it. Meanwhile the exception discipline that kept your charge-offs below the industry’s becomes the benchmark. The fraud typology your team caught two years early becomes a feature in a vendor’s next release, sold to everyone including the banks that missed it.
The value created does not disappear. It splits three ways. Customers keep most of it, as better service and thinner pricing, because when every competitor has the same capability, competition hands the surplus to the client. The model vendor keeps the learning, which compounds. The bank keeps a front-loaded productivity gain that erodes as rivals adopt, plus a growing dependency on the vendor’s next upgrade.
Notice who wins and loses from pooling. A bank with a mediocre credit culture is a net beneficiary of a shared model: it uploads noise and downloads the industry’s blended judgment. A top-quartile institution runs the trade in reverse. It contributes above-average judgment and receives back the mean. No serious trading firm would let rivals train on its best PM’s kill criteria. Your general counsel will object that the contracts prohibit anything so crude, and the contracts do. The transfer happens anyway, through channels no training clause touches.
The deepest version of the problem is this: data rights are not learning rights. Your lawyers know how to negotiate retention, confidentiality and training opt-outs, and by and large they have won those clauses. But the asset leaking out is not rows of data. It is the derived layer: which workflows banks route to models, where the ambiguity sits, what a good escalation looks like, how a strong credit memo differs from a weak one. Once a vendor understands the shape of a decision, it does not need your data to reproduce it. It can hire domain experts, including your former MDs, to write training tasks that exercise the same judgment. Anthropic’s ten financial agents were co-developed with a handful of institutions and are now sold to everyone those institutions compete with. Co-developed with a few, distributed to the industry: that is the learning flywheel operating in plain sight. Alex Karp has warned that enterprises are leaking their IP into the labs; Satya Nadella frames the answer as sovereignty over your own intelligence. Different vocabulary, same observation. The scarce asset has moved from the model to the context and know-how the model learns from.
One objection deserves an answer before it is raised: no model approves a loan today. Correct. Fair-lending rules, model risk frameworks and common sense keep a human on the final call, and the current tools draft and assemble rather than decide. But the drafting layer is where judgment rehearses in the open, and the vendors are candid about the direction of travel. Goldman’s CIO describes the third and most valuable wave of adoption as AI applied to risk and investment decisions themselves. Treat today’s assistant as the training ground for tomorrow’s decision-maker, because that is how its makers treat it. And be honest about the horizon, because capital allocation depends on it: in regulated jurisdictions, models will not be pricing loans for years. But the two clocks in this problem run at different speeds. Adoption gains you forgo can be bought back later at market price. The learning that leaks in the meantime compounds and cannot be repurchased at any price. The distant clock governs when models decide; the near clock governs who owns what they will have learned by then. Sort now, pace the spend.
II. The three routes banks are actually taking
Adoption is not monolithic. In practice we see three architectures, and each answers the value capture question differently.
Route one: frontier models behind an in-house gateway. This is the flagship pattern at the largest institutions. JPMorganChase built LLM Suite entirely in-house as a model-agnostic portal to OpenAI and Anthropic models, rolled it out to roughly 250,000 employees, refreshes it every eight weeks, and keeps deepening its connections into the bank’s own data and systems. Goldman’s GS AI Assistant went firmwide in 2025 on the same logic: one governed interface, multiple external models including open-weight options, everything inside the firm’s audit perimeter. Standard Chartered runs SC GPT across dozens of markets. The wrapper buys real things: switching power over vendors, controlled data flows, and a place to accumulate the bank’s own prompts, evaluations and integrations. JPMorgan said the quiet part early, when its data chief explained that the bank’s data is a differentiator and would not be used to train external models.
Be honest about what the wrapper does not buy. It protects the data; it does not stop the deeper transfer. Every co-development session, every roadmap conversation, every public deployment, and the sheer pattern of what regulated institutions ask these models to do teaches the labs the job logic of banking. And the gateway does not repeal the arithmetic of sameness. If the top fifty banks all route the same decisions through the same two or three frontier models, convergence arrives anyway. It just arrives with better governance documentation. The standard rebuttal is that the model is a commodity and the bank’s context is the edge: same engine, proprietary fuel. That argument holds while AI remains a drafting layer sitting on top of your documents. It weakens as agents take over the workflow itself, because the more of the decision path the vendor’s agent executes, the more of the differentiating context lives in the vendor’s scaffolding rather than in your files.
Route two: buying the vendor’s packaged judgment. HSBC signed a multi-year partnership with Google expected to enable more than two hundred use cases on Gemini’s agent platform. FIS is embedding Claude into the system of record for financial crime; Fiserv’s agentOS with OpenAI aims to put agents into thousands of institutions; both major labs now run their own deployment and consulting arms to accelerate exactly this. It is the fastest route to capability and, by construction, the fastest route to the baseline. A credit-memo agent available to every client of your core provider cannot be a moat for any of them. For a sub-scale bank this trade is often still correct: if your internal judgment on a workflow is below the industry average, renting the industry average is an upgrade. That math is honest, and it cuts both ways. The banks it flatters least are the ones with the most judgment to lose.
Route three: open-weight models inside the perimeter. The third pattern runs Mistral-, Qwen- or DeepSeek-class models on infrastructure the bank controls, whether on-premise or in dedicated tenancy. The capability gap to the frontier has narrowed to months on many workloads, at a fraction of the token cost, and Chinese open-weight models have been running above 30 percent of US token traffic on developer routing platforms in recent months. No bank compliance team is deploying DeepSeek tomorrow, and the realistic in-perimeter menu is Western weights and dedicated tenancies. The number matters anyway, as a price signal: it caps what anyone can charge for commodity cognition. Here, sovereignty is architectural rather than contractual: no egress, no vendor telemetry, and the learning loop itself, the fine-tunes, evaluation sets and task libraries, is owned outright. The costs are equally real: a lag on the hardest reasoning, and an engineering burden that perhaps a dozen banks worldwide can carry well. Almost nobody runs this as the whole answer. The sophisticated institutions run it as the protected tier of a portfolio.
So, is the wall truly a wall? The fair summary: enterprise agreements with zero retention and no-training clauses genuinely protect your documents and your customers’ data, and dedicated deployments protect them further. What no current contract protects is the learning layer, the map of where banking’s hard decisions live and what good judgment on them looks like. That layer is exactly what the labs are now spending billions to acquire, through co-development, through expert-built training tasks, and through first-party products that sit inside your workflows. Anyone telling you the data clauses settle the matter is answering the wrong question.
III. Our recommendation, and where this goes
The mistake is treating this as one procurement decision. It is a sorting exercise, workflow by workflow, governed by a single test: if every competitor learned exactly how we make this decision, would we still win?
Run honestly, the test returns different answers for different banks. A retail franchise whose moat is deposit cost and distribution will find true judgment work in a few corners of the firm; a wholesale, markets or advisory house will find it everywhere. That is not a weakness of the test. It is the point. Your exposure to this whole problem is proportional to how much of your economics is judgment rather than franchise, and flattering yourself about that ratio is precisely how the wrong workflows end up on the shared feed.
Where the answer is yes, pool aggressively. KYC assembly, document extraction, reconciliation, translation, routine drafting, code assistance: use the best shared model or packaged agent, negotiate hard on price, and take the entire efficiency gain. Defending proprietary mediocrity is expensive vanity, and most of a bank’s cost base sits in work where the answer is yes.
Where the answer is no, the judgment is the franchise, and it stays off the shared feed. Credit exceptions, structuring, syndicate pricing, workout calls, the pattern recognition of your best bankers. Here, be clear-eyed about what “keep it home” means at your scale, because the protected tier is defined by who owns the learning loop, not by who owns the GPUs. For the handful of institutions with JPMorgan-class engineering depth, it can mean open-weight models inside the perimeter with everything built in-house. For everyone else, a self-hosted frontier stack is a nine-figure science project that will lag while competitors ship, and the honest version is narrower: dedicated deployments with contractually owned fine-tunes, evaluation sets and task libraries built and owned in-house even when the model itself is rented, and every human correction, override and escalation captured as your training signal rather than exported as someone else’s. That last discipline costs almost nothing and is where most of the value sits. Treat those artifacts as balance-sheet-grade IP. Inventory them. Hand them to the same model risk discipline that governs your credit models, because functionally that is what they are becoming.
At the boundary between the tiers, extend the gateway’s mandate from data protection to learning protection. The LLM Suite architecture is the right chassis; the upgrade is to make it the place where the firm’s accumulated judgment compounds, not merely the place where data leakage is prevented. And in every vendor negotiation, put learning rights on the table next to data rights: what does the vendor learn from serving us, who owns the derived artifacts, and what may be productized for the market we compete in. Expect supervisors to force the issue regardless: AI concentration is already being discussed as a system-level risk rather than a technology line item, and the banks that arrive with a defensible sorting framework will have the easier examination.
On the future, three observations.
- First, intelligence is becoming a purchasable input; Goldman’s CIO captured the shift by noting that firms can now buy intelligence instead of infrastructure. Anything purchasable is available to your competitor at the same price, so bought intelligence can lower your costs forever and be your moat never.
- Second, the commodity layer of banking cognition will deflate quickly, because open-weight competition is already dragging token prices down, and the vendors know it. That is precisely why they are climbing the stack, from models to tools to workflows to decisions, where the pricing power lives. The labs’ pre-built banking agents and in-house consulting arms are that climb in progress.
- Third, in the world this produces, the durable bank assets are the license, the balance sheet, the client franchise, the proprietary transaction exhaust no one else sees, and whatever decision-making judgment you kept proprietary. Everything else converges.
Consumer platforms taught us how this movie ends. On TikTok and YouTube, the user believes he is the customer while his behavior is the raw material, and the learning extracted from it is the real product. Enterprise AI points that same machine at the most expensive know-how in your firm: every hesitation, override and second look from your best people, teaching a shared system how your bank thinks. At least the creators get paid.
If this reaches your board, attach three asks, or it becomes an interesting discussion that changes nothing.
- First, have management run the sorting test on the twenty most AI-exposed workflows and report which tier each belongs in, flagging every place the current architecture contradicts the answer.
- Second, have the general counsel report what the vendor agreements actually say about derived artifacts, co-development IP and productization; in most institutions the honest answer is nothing, and the board should hear that said aloud.
- Third, have model risk report whether the firm’s evaluation sets, override logs and escalation patterns are inventoried and governed as intellectual property. That third item is cheap and regret-free whatever you conclude about the rest.
Banks that manage this boundary well will enjoy the best of both worlds, riding deflating costs on commodity cognition while compounding proprietary learning where it decides outcomes. Banks that manage it badly will remain large, regulated, and profitable for a while, as distribution channels for someone else’s intelligence, with margins to match. The difference between the two is not the size of the AI budget. It is the discipline to ask, before any high-value workflow goes onto a shared model, the only question that matters: if everyone learned how we do this, would we still be better? Where yes, pool it and bank the gain. Where no, that judgment is the moat. Keep it home.
