An explainer for community banks and credit unions — what “owned intelligence” is, where it helps, and how to evaluate AI and OI.
In development · shown by invitation
The artificial intelligence most people have heard of — Claude, ChatGPT, Gemini and the rest — works by prediction. It was built by having a program read an enormous amount of text and learn 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, then the next, until an answer has formed. This requires vast amounts of computing resources, which is why it happens somewhere else.
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. This is the one thing worth understanding before anything else:
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. The machine runs in a distant company’s data center, and whatever is typed into it travels there. The arrangement is a tenancy: it works while you pay, while you are connected, and while that company chooses to offer it. Any of it can change without notice, because none of it belongs to you.
For a great many uses, that is perfectly fine. For work that a bank examiner will read, or that touches a member’s or borrower’s private information, it is a poor fit.
The same underlying technology can be put to work in a completely different shape. Three changes, and they are the whole idea.
The machine is a box on your own network, behind your own locks. Member and borrower information never leaves the premises. There is no distant company on the other end of anything.
It is handed your documents and the rules that govern them, and its job is to find, compare, and quote — not to recall. Every figure it reports points back to the page it came from, so anything can be checked against the paper in the folder.
What it produces is a draft. Where a decision belongs to a person, the machine stops and waits. Each morning a named person reviews the night’s work item by item and signs, or doesn’t.
A chatbot is a brilliant, confident consultant who has read everything, often cites no source, keeps none of your files, and rarely says “I don’t know.” Owned Intelligence is a very fast clerk who works only from your own filing cabinet, shows you the page behind every number, and leaves the whole stack on your desk for signature.
Both are useful. Only one of them can be checked — and in a supervised institution, that is critical.
| Rented | Owned | |
|---|---|---|
| Where it lives | A distant company’s data center. Your questions travel there. | A box in your building. Nothing travels anywhere. |
| How it answers | By prediction, from patterns. Nothing is looked up. | 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. |
| What it costs | A meter that never stops — every answer generated again from scratch. | Bought once, plus upkeep. Cost does not grow with use. |
| Who sees your information | The company running it, under terms it can revise. | No one. There is no connection. |
| If the company disappears | The service stops. | The box keeps working. So do the paper files, as always. |
| Who is accountable | Unclear — no one signed anything. | A named person at the institution, per item, on the record. |
Most of an institution’s back office is one motion repeated: fetch the file, compare it to the rule, note what a person needs to see. That motion is what this machine is genuinely good at, and the same work is currently done by people whose time is worth more elsewhere.
Memos drafted from spread financials; your own written lending policy applied as the machine’s rulebook, with every exception flagged against the line it strains; files checked complete before closing; annual reviews and covenant monitoring assembled on schedule.
Alert workups drafted with their supporting records; member and customer reviews refreshed on time; regulatory reports assembled and checked field by field before a person files them.
Rule changes laid against your current policies with the differences marked; vendor review files assembled; examination preparation answered out of records the work produced simply by existing.
A scanner gives an institution its first machine-readable answer to “what is actually in our files?” The paper stays the system of record. The machine keeps an index of it and cites folder and page for everything, so the paper always wins any dispute.
Named in advance, before any demonstration: credit decisions — approving, pricing, declining. Suspicious-activity determinations. Fair-lending judgments. Adverse action. Anything where the wrong word is an enforcement matter rather than a correction. The machine drafts; it does not decide.
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.
A credit department is several functions sharing one book of business — origination, servicing, portfolio management, workouts, policy, strategy. Each learns something the others need, and many don’t have a practical way to pass it along.
Nobody withholds anything. The knowledge simply never travels, because retrieving it costs more than anyone has: the answer sits in a folder somewhere, reconstructing it takes a week nobody can spare, so the question stops being asked and each function runs on its own memory.
Today What was assumed at underwriting is often not systematically compared with what happened. The projection lives in the memo; the reality arrives years later in statements nobody sets beside it.
With a shared record Every annual review arrives with the original assumption printed next to current performance, from the same file, cited to the page.
Today A hundred individual observations never add up to a portfolio view until someone builds a spreadsheet.
With a shared record The same record that watches one loan answers questions about all of them — concentrations, expiring items, covenant drift — continuously, with the folders to pull listed.
Today Policy is revised on judgment and recollection. The pattern of exceptions granted over a year — the clearest comparison of policy against execution — is rarely assembled.
With a shared record Each exception is recorded with its reason at the moment it is granted. A year later that record is the agenda for the policy review, already written.
None of this requires anyone to file a report, attend a standing meeting, or maintain a new system. Each function’s ordinary work deposits into the same record as a byproduct of getting the work done.
These are stated before anything is sold, because they are the parts that would be easiest to quietly abandon later, and because an institution should be able to hold a vendor to something in writing.
On your premises, under your policies. No member or borrower record transits an outside company, trains anyone’s system, or sits in a place you did not choose.
Regulations carry effective dates; so does the machine. Every piece of work cites the edition it ran against. A change is a new, dated configuration that you review and adopt — never pushed to you overnight.
Human review is not a checkbox added at the end. Queues are sized to what attention can actually cover, and approval is one item at a time, so that rubber-stamping is harder than reading.
Checking happens against the published rule and the original document — by plain machinery, or by a person. Nothing is ever asked to grade its own work.
The audit trail is not a feature added for the examination. It is what the method produces by existing.
Your files are not our laboratory. Development happens on invented data.
A defined set of documents, for the length of one pass. It holds no credential to your core, your email, or any customer system, and it has no address to send anything to. When the pass ends, it keeps the report and nothing else.
The report is a file on your premises, carried by your own systems under your own policies. The machine does not transmit it. There is no share link, no public address, and no copy anywhere else.
The moment a person emails or exports a report, it passes from the machine’s control into yours. Every export is logged with who, what, and when — so that sending is a deliberate act rather than an accident.
If you take nothing else from this page, take these. They are short, they are answerable in a sentence, and the answers tell you nearly everything. Ask them of every vendor who brings artificial intelligence to your institution.
If the answer involves anyone else’s computers, ask what happens to it there, who can read it, and under what terms — and get the answer in writing before the pilot, not after.
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 what the reviewer actually sees. If approving a hundred items takes one click, the review is nominal.
If the price rises with use, you have rented a service rather than bought a tool — which may be fine, as long as everyone knows which one it is.
A small vendor is not a disqualification, but the honest ones have an answer that does not require them to survive.
smallbanc is in development. When there is something worth seeing, it will be demonstrated on invented files, so that the institution risks nothing while judging the work — a morning’s queue you can read yourself, on a piece of work of your choosing.
Write to us