AI FOR COMMUNITY AND REGIONAL BANKS

From Whether to How: What Banks Are Asking About AI Now

Bank leaders have mostly settled whether to use AI. The questions now are about data, ownership and proof.

Clarista · November 17, 2026 · 6 min read

TL;DR

  • The question has changed. Most banks now ask how to use AI safely. Examiners ask the same question in routine exams.
  • Data comes first. Policies, procedures and other documents shape AI answers more than the core system does. Old or conflicting documents lead to wrong answers.
  • AI often arrives through vendors. Many banks first meet AI inside products they already pay for. Responsibility for it stays with the bank.
  • Proof comes from the systems. A sign-off shows what the bank intended. Access records and logs show what actually happened.

A couple of years ago, most bank conversations about AI started with "should we?" Today they start with "how?" At the NJBankers 2026 AI in Banking conference, the discussion moved past tools. It centered on data, ownership and oversight.

This post sums up the questions bankers are asking now, and how a community or regional bank can begin to answer them.

Why Has the Question Moved From Whether to How?

Three things changed at once.

AI is already in the building. It comes built into email, office software, core platforms and vendor products. Staff use it whether or not the bank has a plan.

Examiners are asking. Reuters reported in July 2026 that the OCC and the Federal Reserve now discuss AI in routine bank exams. Examiners ask how human review works, who can stop a system, how vendors are overseen, and whether AI tools reach data they should not use.

Guidance leaves the design to the bank. The revised model risk guidance issued in April 2026, SR 26-2, states that generative and agentic AI models are outside its scope. The agencies plan a request for information on AI. Until then, each bank applies its broader risk and governance practices in its own way.

Put together, the bank has to decide how AI fits its own controls, and be ready to show its work.

Which Questions Are Banks Asking Now?

The new questions are more specific, and each one points to something the bank can produce.

Earlier questionQuestion nowWhat it takes to answer
Should we use AI?Which use comes first, and who owns it?One named owner, a baseline and two or three measures
Which tool should we buy?What data will it read, and is that data current?A list of sources, each with an owner and a review date
Is the vendor's AI safe?What can the vendor change without telling us?Contract terms on notice, data use and exit
Did we approve it?What does it actually do each day?Records of access, outputs and human review
Will AI replace staff?Which tasks stay with people?A clear line between preparing work and making decisions

Why Does the Data Matter More Than the Model?

Most of a bank's knowledge sits in documents, not in the core system. Policies, procedures, training guides, credit memos and emails far outweigh the account and loan records.

When an AI assistant answers a staff question, it draws on those documents. If an old policy is still on the shared drive, the assistant can quote it. If two departments wrote different procedures for the same task, the answer depends on which one it finds. Many wrong answers trace back to the source material rather than the model.

This changes where to start. Waiting for perfect data stalls the project. A better approach is to use AI to find the problems. Point it at one set of documents, such as deposit operations procedures, and ask it to flag conflicts, outdated references and gaps. Fix those, then move to the next set. Each pass leaves the bank's documents in better shape, with or without AI.

Four questions to answer for every AI use 1. Data What does it read? Sources, owners and review dates 2. Owner Who answers for it? One person at the bank who can stop it 3. Access What can it reach? Approved scope, checked against actual access 4. Evidence How do we show it? Logs of outputs and the reviews people made
The same four questions apply to AI the bank builds and AI that comes inside a vendor's product. The last one is where most reviews fall short.

Who Owns AI That Arrives Inside Vendor Products?

Many community banks meet AI first inside products they already use. A core provider adds a summary feature. A fraud tool adds a new scoring model. A document vendor adds a chat assistant. Some of these arrive with a release note rather than a conversation.

The 2023 interagency guidance on third-party relationships is clear that using a third party does not reduce the bank's responsibility for the activity. For AI, that leads to three questions at every renewal:

Collecting documents only goes so far. A SOC 2 report or a signed attestation describes the vendor's controls. It does not show the bank what happened to its own customers' data. The stronger test is whether the bank can rebuild a customer's journey and the related transactions from its own records.

It also helps to keep one register for vendor AI and in-house AI. The board then sees the full picture in one place.

How Should Governance Keep Up With AI?

Most AI governance today is top-down. A committee reviews a use, a risk owner signs off, and the policy is updated. That work matters. But a sign-off shows intent. Behavior shows up only in the systems: what data was read, who had access, what was produced and what a person changed.

Work on AI and data governance in regulated firms points to five practices.

  1. Govern the AI and its data together. An AI tool carries the risk of the data it can reach. Each register entry should name the data sources, their quality checks and who can see them.
  2. Compare approved scope with actual access. Treat an AI agent like a new hire, with a defined role and limited permissions. Then check the access records to see what it actually touched.
  3. Make evidence a by-product of daily work. When reviews, approvals and exceptions are recorded in the systems people already use, the exam file builds itself.
  4. Size controls to risk. An internal tool that drafts policy answers needs lighter review than a model that informs credit decisions. Keep most early projects in the lower tier.
  5. Map many frameworks to one set of controls. State rules, federal guidance and frameworks such as the NIST AI Risk Management Framework overlap heavily. One well-run control can answer several of them.

Where Should a Bank Start?

A practical first 90 days looks like this:

  1. Pick one internal use with a clear pain point. Good candidates are staff policy questions or annual loan review packages.
  2. Name one business owner who is accountable and can stop it.
  3. Record the baseline today: hours spent, backlog or error rate.
  4. List the data it will read and check those sources first.
  5. Keep a record of every output a person reviews and what they changed.

At day 90, the bank has a measured result, a record an examiner can follow, and a template for the next use.

Why Is Trust the Community Bank's Advantage?

Customers stay with a community bank because they trust their banker. In J.D. Power's 2026 U.S. Retail Banking Satisfaction Study, trust ranked as the most important of seven satisfaction factors. The same study found that 20% of customers moved money away from their primary bank in the prior three months, up from 17% a year earlier.

That makes the human relationship worth protecting. AI works best behind the banker: preparing the file, pulling the history, flagging what changed. The conversation with the client, and the decision, stay with people.

Frequently Asked Questions

Does a community bank need a separate AI policy?

Most banks can extend the policies they already have for model risk, third parties, information security and data governance. What they usually add is an AI register, a short intake form for new uses, and risk tiers that decide how much review each use gets.

Is generative AI covered by the 2026 model risk guidance?

SR 26-2 states that generative and agentic AI models are outside its scope, and the agencies plan a request for information on AI. Examiners still ask about AI in routine exams. Banks apply their broader risk and governance practices, sized to the risk of each use.

How can a bank find the AI already inside its vendor products?

Ask each critical vendor in writing which features use AI, what bank data those features read, whether they are on by default, and how changes are announced. Record the answers in the same register the bank uses for its own AI.

What is a good first AI use case?

An internal, document-heavy task where a person reviews every output. Examples are answering staff policy questions, preparing annual loan review packages, and assembling exam request lists.

A Question for NJBankers Members

Which of these questions is your bank working through right now: data, ownership, vendor AI or evidence? Share your experience with your peers through NJBankers. We will build future posts around what members are facing.

Next in this series (December 1): Start With the Data You Have. How to use AI to find outdated, conflicting and missing documents, one source at a time.

See how a bank can answer the data, ownership and evidence questions on its own systems.

Book a demo