What a Platform Cannot Absorb

The large AI platforms will absorb more of a regulated institution's work than its strategy assumes, and less than its fear assumes. Regulatory expertise will not protect it, and neither will the data it merely holds. Two things will: data it has turned into an asset, and standing, the trust it has earned with its supervisor and its clients and the policies its own history has written.

An AI strategy should therefore do two things. Concede on purpose everything that rests on neither, and spend what the concession frees up on both.

Most of your work is not yours

Drafting, summarizing, scheduling, searching, answering routine questions: a bank's needs here are a retailer's needs. Nothing about that work is specific to the institution, so nothing protects it. A second tier is specific to the industry but not to the institution: every bank screens transactions and builds client files in much the same way.

The sharpest objection I was given on this came from a senior AI leader at a global asset manager. Platform distribution beats domain expertise, he argued. The platforms already sit inside the office suite, the customer system and the cloud, and they will put AI into each of them until the work is simply done there. On this half he is right, and I concede it entirely.

An institution that builds its own AI for commodity work has chosen to fight where it holds no advantage. Goldman Sachs showed what that costs in another decade. It took one of the strongest franchises in finance into consumer lending, a business decided by price and scale, and in January 2023 disclosed more than three billion dollars of pre-tax losses in under three years on the unit that housed its card and installment lending.

One moat leaks, one has to be built, one has to be earned

Institutions usually name three things that protect them: what they know about regulation, the data they hold, and their standing. They behave differently.

Partha Roarke, my co-founder at Fourth Order Labs, asked the question that took the first one apart for me: can generations of Swiss banking practice simply be learned by a model? Most of it can. What a rule says, and how a screening is run or a client file is built, is material a model already recites faster than a compliance officer can, and in my experience most of it correctly. Regulatory knowledge of that kind is a syllabus.

Data is a different case. Holding it protects nobody: every bank has data, and I have argued that it is necessary and not sufficient. It becomes an asset when it is treated as one, with its lineage known, its quality managed, and an ontology layer on top that says what a product, a client and an industry mean across the institution. You can run AI on a single system without any of that. It works, and it returns much less than AI on an integrated architecture.

Few institutions have that layer, and the reason is how they pay for it. At a global systemically important bank, every project had to fund its own data into the curated layer, and was charged a premium on top for the infrastructure underneath. So the first AI project carried the cost of an asset that every later project would use, and its return on investment case did not close. Any finance function would recognize the error: a shared asset charged in full to its first user. An institution that puts no value on data will never find a business case for the layer others could consume from, and that stifles innovation.

This is why I have said an institution must own its orchestration and its data. A platform cannot build that layer for you, because it encodes the institution's own definitions. It is the first thing a platform cannot absorb, and it has to be built on purpose.

Standing does not transfer

Standing is a relationship with a counterparty, and that is why it does not encode.

With clients it looks like this. A private banker advising a family on a cross-border structure draws on twenty years of watching how that family behaves when markets fall, and on the fact that in 2008 the banker was the one who called. A system can log that call. What the family concluded from it sits with the banker and the bank, and no log transfers it.

With the supervisor it is less visible and worth more. A regulated institution carries its history in its policies. One bank I worked in had them in the thousands. Most were written after something had happened, an incident or a regulatory finding, to make sure it did not happen again. So the set is specific to the institution. A bank that never had a certain risk appetite never had those clients, never had that finding, and never wrote that policy. How much the regulator has intervened decides how specific the rules become.

Any AI system we deployed had to operate inside that. No external platform knew those policies existed. Inside the bank they were all written down, and finding them was not the difficulty. The difficulty fell on the individual. Whoever had to write the requirements for a system, build it, or audit against it had to work out which policies applied and what they meant for the case in front of them. That is interpretation, and a great deal of it.

The reaction to an incident was rarely only a policy, either. It was usually a new system as well, with a data architecture built to solve that one problem. Do that for years and the same word means different things in different places. A product was defined one way in the front office and another in the back office. Industries were classified under more than one standard, so a risk measured by industry depended on whose definition was in use. The regulation was the same as every other bank's. What the institution had made of it was its own. I call the result data spaghetti, and it is why getting the data was the hardest part of putting an AI system to work.

The obvious reply is that a policy is a document, and a document can be handed to a platform. The documents can. The interpretation is harder to hand over. Thousands of policies written in reaction to separate events are a record more than a rulebook, and deciding which of them bind a given process, and how, is judgment exercised by the people who write the requirements and the people who audit them.

Part of the answer is the oldest rule in outsourcing: the institution stays fully responsible for what its provider does, in Switzerland under FINMA Circular 2018/3 for any significant function, and in the European Union under the Digital Operational Resilience Act for technology services. But that rule follows the institution into most of what it buys, so it is a floor and not the argument. The argument is that in the core the thing being applied is the institution's own position, formed by its own history with its supervisor. A platform can execute that position. It was not there when the position was formed, and it cannot defend it.

Consider a client-facing assistant that answers from the published rule and not from the institution's own, stricter policy. Every answer is defensible on the text and inconsistent with what the institution committed to after its last finding. The first time one of them surfaces, the institution is explaining why its own system did not know its own position. That is the start of a finding, and findings are how standing is lost.

Accounting has a rule that fits. Goodwill a firm generates internally cannot be booked. It appears on no statement, and in a regulated institution it is often the most valuable thing there is. Standing is internally generated goodwill with a supervisor. It can be bought only with the institution, and even then the supervisor takes a fresh look at the new owner. It accrues through conduct, and it can be impaired. Certified trust is what an institution extends to its own agents. Standing is what a supervisor extends to the institution.

Concede on purpose, and spend the difference

Stop building AI for commodity work, and consume it. There are exceptions, data residency first among them, but consumption is the default and construction needs a reason. Conceding the work is not conceding the exit: what you consume, you should still be able to leave.

Then move budget and scarce people to the data layer and to the processes that run on standing, and measure those deployments differently. Productivity is the floor. A system in the core should leave the institution easier to trust than it found it. In practice that means evidence a supervisor could read: who decided, under which policy, on what date. That evidence is what you build. The model you can rent.

Why this matters most in Europe

Europe will not win on infrastructure, and I do not see it winning on models. European models exist, but they do not have the adoption of the American and Chinese ones. I have made that case at length, as a problem of capital and distribution and of a continent that cannot act as one.

What Europe has is the third layer: institutions with their own data and their own history. That is a moat only for those who build it into one. A European bank that turns its data into an asset and builds specific agents on it is competing where it can win. One that waits for a European model is competing where it cannot.

Changing layers is possible, and the price is visible. Oracle built its business on databases and enterprise software. In 2026 it is cutting thousands of jobs while it borrows tens of billions of dollars to become an AI infrastructure provider. It may be right that it has no choice, and whether the bet pays is not yet known. A regulated institution does not have to make a bet like that to win at its own layer.

The boundary moves, and you can watch it

I may be wrong about where the line sits. A former country head at one of the platforms told me the layers will converge into a handful of full-stack providers, and on infrastructure and models I think he is right. For infrastructure, supervisors already let an institution rely in part on a provider's own certifications and pooled audits, and since 2025 the European oversight regime for critical technology providers puts the largest ones under direct oversight.

The signal is narrower than that. When a supervisor accepts a platform's evidence about a judgment, in place of the institution's, the core is shrinking.

The test you can run this quarter

Take the list of AI initiatives your institution is funding and sort each one into three columns.

Commodity. Any firm in any industry would need the same thing.

Industry. Any firm in your industry would need the same thing.

Yours. It depends on something only you have: your own data layer, a policy your own history wrote, or trust you earned with a client.

Then look at where the budget and the best people sit.

If most of the spend is in the first column, you are building what you will be sold.

If the building is in the second column, buy it from someone who serves your whole industry, and keep the layer that carries your own policies.

If the third column is empty, your AI strategy has no moat in it, whatever it says about differentiation.

And if the third column is funded but produces no evidence of who decided, under which policy and when, you are impairing standing, not building it.

The position I hold

Capability can be rented, and increasingly it will be. The data layer has to be built, and it will not be until someone gives data a value. Standing is earned, and every AI system in the core either adds to it or impairs it. Governance is what tells the two apart. That is what I mean when I say governance enables trust, and trust enables speed: an institution that can show how its policies were applied can move in the core without waiting to be asked.

Concede the commodity work on purpose. Fund the data layer as the shared asset it is. Then govern standing as what it is: often the largest asset the institution holds, and the one that appears on no statement, yours or a platform's.