The 60-cent dollar: where a Chicago bank's cost actually sits

The pitch deck always opens in the front office. A chat assistant for customers, a smarter app, a faster call center. Meanwhile the cost that decides the quarter is two floors down, where a team is reconciling transactions by hand, chasing a missing document for an onboarding file, and routing exceptions through an inbox. None of that is glamorous. All of it is where the money is.

The spend is already moving, just not always to the right room.

Find your own 60-cent dollar: one back-office workflow, one working session

Everyone is buying the front office. The cost is in the back.

Adoption is no longer the question. By 2025, 77% of banks had launched or soft-launched a generative AI application, up from 61% two years earlier. The interesting part is where the value actually lands: 56% of use cases target internal efficiency rather than new revenue, and the pioneers are using AI to cut reconciliation backlogs by roughly half and to drop the cost of verifying a commercial client by about 40%.

That tracks with where the drag lives. Regional banks still run a median efficiency ratio near 60%, which means roughly 60 cents of cost for every dollar of revenue, and the largest sources of that cost are manual draw administration, fragmented document collection, and multi-step approval workflows. The good news is that those are exactly the rule-based, high-volume tasks that software handles well: automating them can cut processing costs by a quarter to a half in the targeted function.

Where the time actually goes

When we map a bank or fintech operating week, the same back-office rooms show up every time, and they are all workflow, not intelligence. Reconciliation, where records from two systems have to agree and a human resolves every break. Onboarding and KYC verification, where a file waits on one document and one approval. Exception handling, where anything that does not fit the happy path lands in a queue and ages. Reporting, where the same numbers get assembled by hand every month. Each is repetitive, each is measurable, and each is a place where AI assists a person or clears a queue rather than replacing judgment. That is the difference between a back-office AI project that ships and a front-office demo that impresses a board and changes nothing.

What Chicago banks and fintechs get wrong

Three mistakes show up over and over, and none of them are about the technology.

They start customer-facing instead of back-office. A support assistant is visible and easy to approve, so it goes first. But it sits on top of the same slow process, and now the slow process has a friendly voice. The back-office workflow is less visible and far more valuable, because it moves the efficiency ratio, not just the satisfaction score.

They buy the model before mapping the process. A reconciliation is not slow because nobody bought AI. It is slow because the match rules are undocumented, the exception path is tribal knowledge, and two systems disagree by design. Automate that as-is and you have paid to run the mess faster. Map it first, and the automation becomes obvious and small.

They treat governance as paperwork for later. This is the one that separates finance from every other industry. A restaurant can pilot a tool on a Tuesday. A bank cannot, because the model has to survive an exam. The model risk guidance from the Federal Reserve and OCC, revised in 2026 on top of the long-standing SR 11-7 framework, still sets the bar: documented governance, independent validation, and ongoing monitoring for any model that drives a decision. AI does not get a pass on that, and vendor models do not either. If you cannot show your work, the capability does not matter.

And the Chicago fintechs that partner with a chartered bank should read that last point twice. When you ride a bank's charter, you inherit the bank's exam. The governance is not optional because you are a startup. It is the price of the rails.

In finance the winner is not whoever has the best model. It is whoever can put a governed one into production and prove it to an examiner.

Where a Chicago bank or fintech should actually start

The sequence is boring and it works. Pick one high-volume, rule-based back-office workflow where the cost is real and the decision is not a judgment call: reconciliation, KYC verification, or exception routing are the usual first three. Map how it actually moves, break by break, handoff by handoff, not how the procedure manual says it moves. Decide what to automate, what to assist, and what to simply route and measure. Then build the model risk documentation as you build the system, not as a scramble before the exam.

Done that way, AI stops being a line item you defend to the board and becomes a governed improvement to a process you can already see. Done the other way, you have a chatbot, a slower quarter than you hoped, and a validation file you have to write after the fact.

If we had one week with a Chicago community bank or a fintech on a bank partnership, we would not train anything. We would map the reconciliation and the onboarding file, show you where the hours and the exceptions are leaking, and sketch the governance packet the examiner will ask for. The automation comes after that, and it comes cheaper, because by then you can see exactly what you are automating and prove exactly how it behaves.