Article

How SR 26-2 Moves AI Exam Questions Into Your Loan File

Avatar photo
Stephanie Rolewicz
Nov 16, 2026
Illustration of a performance gauge with analytics, financial growth, and document automation icons representing AI-powered construction lending workflows.

Picture the second day of an exam, when the examiner pulls a construction draw from the sample. An AI agent reviewed it and flagged a missing lien waiver, someone cleared the flag, and the draw funded.

The examiner has three questions: 

  • Who told the agent to check for lien waivers? 
  • Who cleared the flag, and why? 
  • And who decided that person had the authority to do it?

Those are credit administration questions as much as model validation questions, and they get answered by the people who run your loan operations, the people who own credit policy, the people who audit both, and the people who chose the vendor. That is the shift lenders should plan for.

The Rule Changed, and the Accountability Stayed

On April 17, 2026, the OCC, the Federal Reserve, and the FDIC retired SR 11-7 and replaced it with new model risk guidance, issued as OCC Bulletin 2026-13, SR 26-2, and FDIC FIL-15-2026. The new guidance is nonbinding and says it does not set enforceable standards. If your policies still reference SR 11-7, those references are worth a review. A footnote in the guidance says generative and agentic AI “are not within the scope of this guidance.”

Read quickly, that sounds like a pass, but the rest of the footnote says otherwise: “Nonetheless, a banking organization’s risk management and governance practices should guide the determination of appropriate governance and controls for any tools, processes, or systems not covered in this document.”

The agencies handed the decision on AI oversight to each institution, without a checklist. Federal Reserve Vice Chair for Supervision Michelle Bowman made the intent clear: the agencies expect “other risk-management and governance practices to support adoption of generative and agentic AI.” Put simply, regulators are supportive of lenders using AI, and they expect the disciplines lenders already have in place to govern it.

Recent signals show where the AI questions are heading. In October, Reuters reported that the Fed, OCC, and FDIC are asking banks about their AI agents. They want to know what limits are in place, whether agent decisions can be audited, and whether people stay involved in key decisions. Those questions focused on cybersecurity, but they apply just as well to lending.

In November, the OCC published a supervisory framework to help its examiners evaluate banks’ use of AI, which the OCC now calls “super intelligence.” The framework is a resource for examiners and says it does not set expectations for banks. Still, it tells examiners that a bank’s AI uses can be classified by “potential consequences, transfer of agency, and the degree of human oversight.” Those line up closely with the questions below.

Our read is that examiners will ask about AI in parts of the exam they already run.

If the agent does credit work, it gets examined like credit work. The OCC’s Comptroller’s Handbook already tells examiners to review “construction loan administration including draw and disbursement practices.” An agent reviewing draws is doing exactly that.

If you bought it from a vendor, you still own it. The agencies’ 2023 third-party guidance says using a vendor “does not diminish its responsibility” to meet requirements. The vendor may run the agent, but the responsibility stays with you.

If the decision is yours, expect to show it. The regulators left AI governance to each bank, so examiners can ask what you decided. Be ready with a list of your AI tools, what each one is allowed to do, and who approved it.

What the Examiner Is Really Asking

Take the AI out of it and these are questions lenders have answered for decades. They come down to three things: 

  • Authority: what was the agent allowed to do, and who said so? 
  • Evidence: what did it actually do on this loan, against which rule, using which documents? 
  • Oversight: who checked its work, and who approved changes to how it works?

The difference is scale: an agent can touch more draws in a week than a loan administrator touches in a month. Every one of those draws is a file an examiner might pull, and every one needs to stand on its own.

What a Good Answer Looks Like

Take the lien waiver draw from the opening. In one version of the file, the examiner can see the policy the agent checked, the flag it raised, the pay application and waiver log it relied on, the approver’s comment when the waiver arrived, and the approval limit that covered the draw. The file answers the questions without anyone relying on memory.

In the other version, the flag appeared and then disappeared. The status says “resolved,” with no comment. The rule the agent followed lives in the vendor’s settings, and nothing on the draw shows who had the authority to clear it. Someone now has to explain all of that from memory, months later, in front of an examiner. That is the draw that is hard to defend.

None of this is a new standard. The OCC handbook already lists segregation of duties, lien searches before disbursement, and dual approval of loan disbursements as effective controls. With an agent in the workflow, a good draw file looks the same, and the human decisions in it matter more.

Who Owns the Answer Now

Responsibility for these answers sits with four groups. Each one should know its part before the exam letter arrives.

Chief credit officers own authority. Credit policy should name which AI tools touch lending decisions, what each one can do without a person, and what dollar limits or approval rules apply. Those permissions need a documented approval. If the authority questions above would be hard to answer today, that is the place to start.

Loan administrators own the evidence. They are the ones clearing flags and approving draws. Every override or resolution needs a reason recorded at the moment it happens. A flag marked resolved with no comment is hard to defend in an exam.

Internal audit owns the test. Sample loans the way an examiner will. Can you reconstruct what the agent did and who decided, from the record alone, without interviewing anyone? Were changes to the agent’s rules or permissions approved, and by whom? It is much better for audit to find a gap than for an examiner to find it.

CIOs and CTOs own the vendor file. We’d look for three things in each AI vendor contract: audit rights, data access, and notice when the agent’s rules or behavior change. The guidance doesn’t require those terms by name, but they make it easier for the bank to meet its responsibility. The OCC’s Spring 2026 Semiannual Risk Perspective names “lack of explainability” among the challenges banks face with AI. If your vendor can’t show you what the agent did and why, you can’t show your examiner either.

These groups don’t always work closely together, and now they should.

Why This Gets Harder at Scale

This is manageable on one loan and harder across a construction book of several hundred loans. An agent may review many draws a week. An examiner samples a few dozen, and a small team has to pull together the history on each one.

When the record lives in email threads and side notes, each sampled file takes real work to piece together. Approvals sit in inboxes, and two files may describe the same step in different ways. The team spends exam week reconstructing what happened.

When each rule, action, and decision is recorded on the loan itself, the team can pull the answers straight from the file. That is the standard we’d hold any AI in a lending workflow to, regardless of who builds it.

Credit Unions and Private Credit Lenders Face It Too

The guidance discussed here is written for banks. Credit unions answer to the NCUA, and private credit funds don’t have a prudential examiner, but both hear similar questions from their own examiners, limited partners, auditors, and the banks that lend to them. The OCC’s Spring 2026 report points to “growth in lending to private credit funds,” which puts a fund’s draw controls in front of all three. The same recordkeeping approach can help credit unions and private credit lenders answer every one of them.

The Bottom Line

The agencies kept the bar for AI in lending where it was, and the questions now land on the loan file and the people who manage it. Lenders who can answer those questions from the record will be in a strong position at their next exam.

How We Think About This at Built

We build our AI Draw Agent around one principle: your institution decides what the agent can do, and the agent works inside those limits. When it flags an issue, the flag stays on the record along with the person’s decision and comment. Approval limits and dual control are set by the lender, not the agent. If you are building out your vendor file, our security program is a good place to start, and we are happy to walk through a draw record with your team.

Examiner Questions About AI in Lending FAQs

What will bank examiners ask about AI agents in lending now that SR 26-2 has replaced SR 11-7?

No agency has published an AI exam checklist, so this answer is an inference from existing guidance. Expect questions on authority, evidence, and oversight, covering what the agent can do, what it did on a sampled loan, and who checked it. In construction loan administration, those questions follow the OCC procedures for draw and disbursement practices, and the draw record holds most of the answers.

What replaced SR 11-7?

On April 17, 2026, the OCC, the Federal Reserve, and the FDIC replaced SR 11-7, the 2011 model risk guidance. The replacement is interagency guidance issued as OCC Bulletin 2026-13, SR 26-2, and FDIC FIL-15-2026. The OCC also rescinded OCC Bulletin 2011-12, its issuance of SR 11-7.

Does SR 26-2 apply to agentic AI?

No. The guidance says generative and agentic AI are “not within the scope” and leaves their controls to each bank’s own risk management and governance. Examiners can still reach agentic AI through the three doors of governance, third-party risk management, and credit administration.

Does third-party risk management guidance apply to AI vendors?

Yes. The agencies’ 2023 third-party risk management guidance says using a third party “does not diminish its responsibility” to meet requirements. Most lending agents are vendor products, so the bank still owns the governance, the contract terms, and the audit rights for each AI tool it uses. The agencies issued a proposed replacement in September 2026, and the 2023 guidance applies until it is final.

Do credit unions and private credit lenders face the same questions?

The April 2026 guidance was issued by the OCC, the Federal Reserve, and the FDIC. Credit unions are examined by the NCUA, and the same three doors apply to any lender using AI in loan administration. Private credit funds hear the same authority, evidence, and oversight questions from LPs, auditors, and the banks that lend to them.

Written by Stephanie Rolewicz

Stephanie Rolewicz brings more than 20 years of experience in banking and financial services to her role as Principal Solutions Engineer at Built Technologies. She spent a decade at Union Bank, rising to Managing Director of commercial and consumer lending technology, before joining Built more than five years ago in a series of leadership roles. Today, she helps the company’s largest banking and lending clients modernize construction finance operations.

See What an Examiner Sees

Every flag, decision and approver comment stays on the draw record. We’ll walk your team through one.

Illustration of a performance gauge with analytics, financial growth, and document automation icons representing AI-powered construction lending workflows.