A strong funding request can die on a single question. A question does not have to touch the merits to stop a case. It only must sound prudent and land on someone who cannot answer it.
One question does that more reliably than any other right now. Someone asks what the examiner will think about AI, and a request you built around a real growth target stops moving. You brought the numbers. You brought the goal the institution already set. The question still holds the room, because raising a regulator sounds responsible, and waving one off does not.
“Your examiner has been reading models this way far longer than AI has been the word for it.”
The answer is older than AI
The question feels like a wall. It has a settled answer. An examiner treats an AI model the way they treat every other model your institution runs, by looking at how you govern it. The rulebook did not go blank when AI arrived, and it did not need a new chapter to cover it.
Model risk is old ground for an examiner. Banks work under interagency model risk guidance the OCC, Federal Reserve, and FDIC refreshed in 2026, which scales what it asks to how much a model can affect. A large, consequential model draws a hard look. A small one does not.
Credit unions have no separate AI rule at all. The NCUA has said its existing standards are technology-neutral and apply to AI the same as anything else, and its examiners reach AI through the frameworks they already use, vendor management, fair lending, and the same operational risk review they run on every system.
Neither is a new discipline invented for AI. Both are ground your examiner has worked for years. You do not recite either one in a meeting. You need to know the answer exists, and that your examiner has been reading models this way far longer than AI has been the word for it.
What an examiner asks
What that governance asks for is plain, and you can picture it without knowing a single regulation. An examiner wants to see that you know what your models do, that someone checks they work as intended, and that someone owns each one if it drifts off course.
Do you know what the tool does? Can you explain how it reaches what it shows? Does someone watch how it performs over time? If an outside vendor supplies the tool, can you show that you checked that vendor and still watch what the tool does in your own shop? Those are the questions an examiner works through, stated in plain terms.
What the tool decides
Examiners weigh a model by what it decides. The heaviest attention goes to models that decide who gets a loan, who gets an account, and what each one pays, because those decisions carry the most consequence for the person on the other side. A model that sets a rate or denies an application can harm someone, so an examiner presses on whether it is fair, whether anyone can explain it, and whether it has drifted.
Put the tool you want funded through that same test. It reads how a visitor is behaving and shows them content that fits. It decides nothing about who gets a loan, an account, or a rate. When it reads a visitor wrong, that person sees a product that does not apply to them and scrolls past. Its logic is written down and can be explained on request, which answers the question examiners press hardest, the one about models nobody in the building can account for. A tool that decides nothing about credit or price and can show its reasoning sits at the bottom of what an examiner weighs.
Answer it before anyone asks
You can get ahead of the objection. A budget meeting raises a concern cold, nobody in the room is holding the answer, and the request stalls while everyone looks around. The same concern, already worked through with the people who own it, arrives with its answer attached and passes without friction. The answer is the same either way. What changes is the timing.
You get there early. You bring your own risk and compliance people in while the idea is still forming, and you let their questions shape the request before you carry it upward. Your risk officer will ask real things. Where does the member data go? Who can see it? Does an outside vendor touch it, and can you turn it off? Can you explain why a member saw what they saw? None of this is new to them. They ask the same set of any system that touches member data, and AI does not change the list. The people who govern every other system built on member data know how to ask and answer these questions, because that work is already theirs. When you bring them in at the start, their answers travel with the request instead of arriving late as objections.
Bringing them in early asks nothing technical of you. You do not learn the guidance or argue the regulation. You walk down the hall early, ask your risk officer what would worry them, and fold the answers into what you bring forward. When the question surfaces later, you already hold the answer. You know the request is a fair one to raise. You know your own risk and compliance team governs this the way it governs everything built on member data. You know the tool decides nothing about who gets credit or what they pay. Those answers came from the people who own the risk, so they hold when someone leans on them.
A risk question that arrives unanswered stalls a good case. Answered before it lands, the same question lets the case through. You turn one into the other early, in a conversation you can start this week with the risk and compliance people down the hall, whose work already includes exactly this. Do that, and the examiner question stops being the thing that stalls your request. The growth target you built the case around goes back in front of the committee, carried by a case that answered the objection before anyone raised it.
