Be a careful fast follower: putting AI to work in lending

Every community bank and credit union leader we talk to is hearing the same two things. Your board has read the headlines and wants to know what you are doing about AI. Your examiner has read the guidance and wants to know how you will control it. Both are reasonable, and they pull in opposite directions.

The way through is to stop treating speed as the goal. For a regulated institution, being a careful fast follower is a legitimate strategy. The institutions that win are the ones that pick a process, prove the result, and can explain it to a board and an examiner, not the ones that demo first.

What boring AI looks like

The AI worth deploying at a community institution today does a small number of well understood jobs. It reads and classifies documents. It pulls fields out of them. It summarizes long files and call notes. It drafts routine correspondence. It flags exceptions for a person to look at. None of that is exotic, and all of it is measurable.

The safe versions share a few traits:

• A person reviews the output before anything customer facing or financially meaningful happens.
• It runs inside the systems and workflows you already have, instead of beside them.
• It comes from a vendor or builder with a track record in regulated work.
• If it is wrong, the damage is small and recoverable.
• You can turn it off tomorrow and go back to the old way.

A four question screen for any use case

Before you spend money on an idea, ask four questions.

1. Is the process already stable and written down? If the people doing the work cannot describe it, a tool will not fix it. See why the process comes first.
2. Is there a human check on the output? If the answer is no, the use case is not ready, however good the demo looked.
3. What is the worst case if it is wrong, and can we recover? A mistaken summary of a call is a very different problem from a mistaken credit decision.
4. Can we measure before and after? If you cannot say what the work costs today, you will not be able to say whether it worked.

Where to start, and where to wait

Good places to start are document intake, internal search across policies and procedures, call and case summaries, quality review sampling, and exception triage. These are high volume, repetitive and easy to measure, and a person stays in the loop. Our piece on document intake walks through one in detail.

Places to wait are fully automated credit decisions, unreviewed customer messaging, and anything you could not explain to an examiner in a few plain sentences. Waiting is not falling behind. It is sequencing. The foundations you build on the safe projects, which are clean data, written processes and a governance routine, are exactly what the harder projects will need later.

Why starting now still makes sense

The slow part of AI at a bank is rarely the model. It is the data, the process and the paperwork. Those take months regardless of what you deploy, and none of them is risky to start. An institution that spends this year tidying its data, mapping its processes and writing its model governance will be able to adopt tools quickly and safely when it chooses to. One that waits will start from zero.

A practical first step

Pick one process that hurts, such as getting documents into a loan file. Write down the steps as they actually happen today. Measure how long a typical item takes and how often it is reworked. Decide who will review the output and what they will look for. Write one page describing the risk and the controls. Then run it for 60 to 90 days beside the existing process before you change anything.

That is a small project, and it is the one that makes every later one cheaper and safer. If you want help scoping it, talk to us.