An explainer for community banks and credit unions — what “owned intelligence” is, where it helps, and how to evaluate anyone selling you AI.
Engagements begin 2027 · shown by invitation
Every institution is being offered artificial intelligence right now. Almost all of it arrives the same way — as a service, from somewhere else, priced by the month. There is a second method, and the difference matters more in a supervised institution than almost anywhere else.
Claude, ChatGPT, Gemini and the rest work by prediction. A program read an enormous amount of text and learned the patterns in it. When you ask it something, it does not look anything up — it produces one word, then guesses the most plausible next word, until an answer has formed.
There is no library inside. Nothing is consulted. That is why these machines are fluent, fast, and frequently right — and also why they can be smoothly, confidently wrong.
For a machine built this way, inventing an answer feels exactly the same as knowing one. It cannot tell the difference, and neither can the reader, because both arrive in the same confident voice.
It also lives somewhere else, and whatever is typed into it travels there. For a great many uses this works well. For work a bank examiner will read, or that touches a member’s private information, it is not a great fit.
The same underlying technology, in a different configuration. Three changes, and they are the whole idea.
Installed on your premises in the first week of the work, holding no password to any of your systems and no way to reach the outside. Your files never leave. You may buy the machine at any time.
It works from your documents and the rules that govern them. Its job is to find, compare and quote — not to recall. Every figure points back to the page it came from.
What it produces is a draft. Where a decision belongs to a person, it stops and waits. Your officer reviews the work and signs, or doesn’t.
Put plainly: AI is a brilliant, confident consultant who has read everything, cites no source, and rarely says “I don’t know.” OI is a very fast assistant who works strictly from your documents, shows you the page behind every number, and leaves the report on your desk for your signature.
Both are useful. Only one of them can be verified. And verifiable is not verified — the citation makes checking take seconds, but a person still checks.
| AI — as usually sold | OI — owned | |
|---|---|---|
| Where it runs | A distant company’s data center. Your questions travel there. | A machine in your building, with no line out. |
| How it answers | By prediction, from patterns. | By finding and quoting your own documents and rules. |
| When it doesn’t know | It often produces something plausible anyway, in the same confident voice. | It stops, and says so, because it was built to. |
| Who sees your files | The company running it, under terms it can revise. | No one outside your building. We hold nothing. |
| Who is accountable | Unclear — no one signed anything. | A named person at your institution, on the record. |
Regulation does not scale down. A credit union with nine employees answers to substantially the same rulebook as one with nine hundred. The work does not shrink with the balance sheet — it lands on fewer desks, belonging to people who are also doing three other jobs.
That is the gap. Not a lack of technology. A lack of hours, in a shop that cannot solve it by hiring a compliance department. Every engagement starts by finding where those hours actually go, which is rarely where anyone expects. So far it has been in four places.
1040s, K-1s, Schedule E, entity returns. The forms are identical at every institution in the country, so a machine can read them once and read them the same way every time. What is not identical — your add-backs, your treatment of undistributed K-1 income, your vacancy assumptions — is your method, written down and signed by you, then applied the same way to every file.
Every reporting requirement, financial covenant, insurance and tax obligation in every agreement — read out, dated, and matched against what is actually in the file. Borrowing base certificates monthly, compliance certificates quarterly, statements annually. Most institutions cannot answer this today, and it is where credit-administration findings come from.
Related entities, pass-through structures, cross-guarantees, and debt held at other institutions — assembled from documents your covenants already require and your files already hold. Collected by most. Read as a set by almost none.
Loans written at the 2021–22 trough are repricing now. Coverage recomputed at reset, loan by loan and across the portfolio, under your own signed method — before the examiner asks.
It runs on the documents you already have, and hands the result to a person.
Most of what is sold in this field arrives as a system your staff must learn, operate and answer for from the first day. That is the wrong order. The work comes first; the machinery follows.
A week of looking rather than a proposal. We sit with the people doing the work and find where the hours go. Then we take on the single job that gives back the most for the least effort — one job, done properly, rather than four done partway.
A machine is installed in your building in the first week and nothing else ever runs on it. Onto it goes your setup: what your agreements require, how you calculate things, what a complete file looks like. We work on site, in the same room as your people. Between visits there is no standing connection — only a session someone there opens and closes.
Your people learn what the machine does, what it refuses to do, and how to check its work in seconds rather than minutes. An institution that finishes this can evaluate any AI vendor on earth, including us.
Which is not an event, because the machine never moved. You may take title whenever you want it — at the end, after a year, or on the first day. We stay for the annual work, or we don’t. Either way you keep the machine, the record, and everything on it.
We hold no client files, ever. A room containing many institutions’ credit files would be a target that did not previously exist. We decline to build one.
Named in advance, before any demonstration: credit decisions — approving, pricing, declining. Risk ratings and grade migration. Whether a covenant breach is an Event of Default. Suspicious-activity determinations. Fair-lending judgments. Adverse action. Anything where the wrong word is an enforcement matter rather than a correction.
The work may state that a covenant computes as not satisfied under your signed method, with citations. Calling that a default is a credit decision and stays with your credit officer. There is no score, no grade, and no recommendation — not as a setting we chose, but because the machine has no way to produce one.
A tool that claims to do all of this is worth questioning. A stated ceiling, in writing before you buy, is what your examiners will look for.
In writing, before anything is sold — because these are the parts that would be easiest to quietly abandon later.
Not for the engagement, not for the annual work, not ever. Nothing is typed into a rented service, transits a third party, or trains anyone’s system. We hold no copy anywhere, and there is no standing connection into your machine.
Which rules apply, which editions, which of your policies, and how coverage is calculated — assembled together and signed by your institution before the first pass runs. Nothing goes onto the machine without your signature.
Regulations carry effective dates; so does your calculation policy. Every piece of work cites the editions it ran against, and a change is a new dated version you review and adopt — never applied retroactively, never pushed to you overnight.
Checking happens against the published rule and the original document, by plain machinery or by a person. Nothing grades its own work, and every number in a ratio is arithmetic rather than a prediction.
We own that the process was followed, that every item is cited to its source, and that the record proves it. What the rules require of your institution is yours to decide, with your counsel.
Your files are not our laboratory. Development happens on invented data.
The machine is already in your building. The setup, the record and the citations stay with you and stand on their own.
Ask the vendor who brings artificial intelligence to your financial institution:
If the answer involves anyone else’s computers, ask who can read them, how long they stay, and under what terms — in writing, before the work starts.
If the answer is anything other than “it stops and says so,” ask what it does instead, and how you would know.
Every figure should point somewhere a person can check in seconds. A number without a source cannot be verified.
Ask to see the calculation policy in writing, with a date and a signature. If coverage comes out at 0.86 rather than 0.94, the answer should be a method you adopted — not a model’s judgment.
A small vendor is not a disqualification, but the honest ones have an answer that does not require them to survive. Ours is in Part Five.
smallbanc is taking a small number of engagements beginning in 2027, one region at a time.
Every engagement is a field test. The record the work produces is built to stand in front of an examiner.
Write to us