Sooner or later, a regional bank's first AI business case lands on the CFO's desk, and the numbers look wrong. The total runs heavier than the vendor quoted, heavier than the case studies, heavier than whatever a peer bank supposedly paid for the same thing. Two conclusions could present themselves, and both could be wrong. The first: AI doesn't pay at our size. The second: someone padded this. The truth is less dramatic and more useful. The business case is honest. It is simply paying for two things and admitting to one.
No bank prices its first loan to recover the cost of building its credit department. Yet that is exactly how most banks think about their first AI model.
A first use case is two purchases wearing one price tag. The first purchase is the use case itself: the license, the integration, the running cost of this one system. The second is the program, the bank's standing ability to own models at all. That ability has contents: an inventory of what models exist and who owns them, a validation standard and the capacity to execute it, a monitoring baseline, a working definition of what "model owner" means at this institution, and an agreed answer to what happens when a model is wrong.
The largest US banks made the second purchase years ago and have amortized it across hundreds of models since; their next model pays marginal cost. Your first model carries the entire program on its back. So when a regional benchmarks its first AI case against a big bank's economics or a vendor's reference deployment, it is not comparing model to model. It is comparing its founding cost to their marginal cost, and losing a comparison it was never in.
Bundling the two purchases misprices everything it touches, three ways.
It kills good use cases. Judged against marginal-cost benchmarks, the bundled case looks unjustifiable, and the bank concludes the technology doesn't pay at its scale. What didn't pay was the accounting.
It punishes caution hardest. The program cost is fixed whether the first use case is ambitious or modest, so the smaller the use case, the worse the bundled ratio. The arithmetic tells the prudent bank that its prudence doesn't pay. Much of the industry's pilot graveyard is this perversity at work: modest pilots carrying full program freight were never going to clear anyone's hurdle rate.
And it builds the program anyway, by accident. Approve the bundle and the capability still comes into being, shaped by whatever the first vendor's paperwork and the first team's improvisations happen to establish, owned by no one. The first use case doesn't just consume the program; it defines it. Every use case after inherits the accidents.
None of this argues for standing up a bureaucracy before the first model. A regional's rational opening position is to rent the program: co-sourced validation, outside counsel, vendor tooling, fixed salaries converted into per-model fees. That is the right answer at one or two models. It stops being the right answer somewhere, and the somewhere has a location: a model count at which the rented validator costs more than the hire, at which monitoring deserves a platform, at which governance deserves a standing owner. One use case is a tax on existing jobs; a portfolio is headcount. Program cost is a step function, not a slope, and scheduling the steps in advance is ordinary capacity planning, a discipline the CFO already runs everywhere else in the bank.
Nor does the program need a money-center's shape. It needs a skeleton sized to the portfolio the bank actually intends: a handful of decisions and documents, small enough to write down, deliberate enough that the first use case inherits them rather than invents them. Sizing it forces the most useful question in the whole exercise: what portfolio does this bank actually intend? A program for one model is overbuilt. If one model is the honest answer, that is a finding too.
So, three moves, all before the first approval, none requiring a hire.
Split the invoice. Two lines: program and use case. Judge the use case on its marginal economics. Judge the program as infrastructure, amortized over the intended portfolio, the way the bank already judges every other piece of infrastructure it owns.
Size the skeleton to intent, not imitation: the minimum standing capability the intended portfolio requires, not a scale model of what the largest banks built.
And schedule the conversions. Decide now, on paper, the model counts at which rented capacity becomes owned. Discovering that point after passing it is how renting quietly becomes the expensive option.
The cheapest decision in a bank's AI program is the one made before anything is bought. This is what that decision contains.
A bank doesn't price its first loan to recover the cost of the credit department; the loan pays for the loan, and the department is what makes lending a business rather than a bet. Price your first model the same way. One invoice for the model, one for the capability. Sign both, just sign them knowingly.