When the Bank's AI Fails, Who Is Blamed?

On 17 April 2026, the Federal Reserve, the OCC and the FDIC replaced the model risk guidance that tells a bank how to test and track the models it relies on, and that banks had followed since 2011. The replacement is SR 26-2. In a footnote about scope, the agencies put it plainly: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance."

Read that twice. And the rule about how a bank checks a model stopped covering the models banks are buying.

Two reactions follow from that sentence. The first is relief, because a validation process built for credit scorecards does not fit a system that answers differently every time you ask it. The second is the one I never see written down. If the rule stopped saying who owns this, who does?

What the rule said before April

But the 2011 guidance had a quality nobody praised and everybody leaned on. It put a name on the model, and it turns out that was the useful part.

A model had an owner. Someone inside the bank was responsible for the thing, a validation team with its own reporting line argued with that owner, and documentation and monitoring carried on after launch. So you can call that bureaucracy. I have called it worse, in meetings, while waiting for a validation question to land somewhere other than my desk. And it did one thing that mattered at the end of the argument: it made a specific person answerable before the model ever ran.

That person could not answer from memory either. The guidance asked for the theory, the data, the testing, the assumptions, and the limit that failed the last time.

The sentence they deleted

But SR 26-2 did more than carve AI out of scope. It removed the sections that assigned the people. Sullivan & Cromwell compared the two texts in April and found the revised letter drops the detailed duties of the board and senior management, the model owner, compliance and internal audit, and keeps one general line in their place. Their summary of that line: model risk management "benefits from clear roles and responsibilities with well-defined accountability."

Roles, yes, and accountability, yes, with nothing in the letter that produces either one.

And the agencies explain the reasoning in public. Michelle Bowman, the Fed's vice chair for supervision, said on 1 May that supervisors had expanded the earlier guidance "beyond its original purpose to apply it in unintended ways," and that AI needs "a different approach." Then she said what comes next: "We expect other risk-management and governance practices to support adoption of generative and agentic AI."

Other practices. Yours, most likely.

I want to be fair to that decision, because the cheap version of this piece says the supervisors ran away. Because there was no agreed method for validating an AI agent when they wrote the letter. A bank can produce four hundred pages about a system that nobody has a test for, and those pages certify nothing while making everyone in the room feel covered. False assurance costs more than a missing owner does. Nobody goes looking for the thing they believe is handled.

And the agencies left the door open. The OCC bulletin says they plan a request for information on model risk and AI "in the near future." Planned, not issued. Banks are in production now.

The questions came back

But on 17 September the Conference of State Bank Supervisors published guidance for examiners: eight questions to ask about AI, three tiers of risk, five documents a bank should be able to produce. It is discretionary, and it is the first concrete description of what a supervisor will want to see. It arrives without an answer key, because the answer differs in every bank.

So the federal rulebook moved away from the thing, the state examiners moved toward it, and in between sits the person who has to answer both in a room next year.

The result I could not prove

I have stood on the wrong side of this, which is why I care how the record gets built.

And years ago, in an agency, we ran a campaign for a hotel. Business rose during the campaign, and that much was true. What we could not say, and said anyway, was how much of the rise belonged to us. Other things were moving that month, because a hotel has a season, and a price, and a sales team of its own. So we sort of worked backwards from a number that had already moved, and by the end we had a result with figures in it, every figure real, and the causal claim doing decoration.

And nobody lied. That is the part people get wrong about these situations. There was no villain in the room, only an empty space where the evidence should have been, and a claim will grow into any space you leave that wide.

But this happened before AI, which matters less than it sounds. Attribution was primitive then, so over-claiming was easy and mostly unpunished. It taught me the thing I now design around. When the proof gets assembled afterwards, you are defending a result. Measuring it stopped being possible the moment the campaign ended.

Who the blame finds

The same thing happens after an AI incident, with a credit decision or a closed account on the other side of it.

And the bank stays answerable, which nothing in April changed. Fair lending law still covers an AI lending decision, the anti-money-laundering rules still cover an AI alert, and consumer protection still covers what a chatbot tells a customer. The institution carries all of it, and it should.

But the individual does not. The letter that used to name a model owner no longer names one for this, so the name gets chosen after the fact, by whoever is holding the file. And with no record, the file ends up pointing at whoever is closest to the work.

Blame lands lower than the people who made the decision. The analyst who ran it, or the team lead who said yes on a Tuesday, while the committee that approved the vendor stays in the clear.

I have watched a room agree on a name in about four minutes. And I have watched the person who did the work go quiet.

Before the first decision

I have not run a bank's model risk review, so I will not hand you a compliance answer. What I know is the vendor side of being asked to prove a result, and the design decisions that make proof possible or impossible.

So the name is the last step, and it only means something when three things sit underneath it.

1. 1A person in the risk register before anything is built, with the limit of their authority written next to their name, and the kinds of decision that go past them.
2. 2The evidence recorded while the work happens: what went in, which model produced the output, how confident the system was, whether it escalated, and who accepted it.
3. 3An output that arrives with its evidence attached.

But the second one is pretty much the whole job. A record made during a decision and a record made after it look identical on a slide. A dashboard tells you what happened. A record tells you why anyone should believe it. And they behave nothing alike in a room where somebody is asking questions.

This is what I build for. TikSense measures each creator's ability to convert sales for your specific product, across commerce data, content fit, audience intent, and visual selling quality. Every analysis gives you a verdict and the outreach copy to send. It is built on The Seven Dimensions of Enterprise AI. And the verdict travels with the inputs behind it, so a seller can check my work instead of trusting it, which is the only version of a decision engine worth shipping.

So put the three together and the structure in The Agentic Authority Escalation Model has somewhere to write. The tier above the agent makes the call, and the record shows that the call happened, what it saw, and who accepted it. Without the record, the tier above is one more opinion in the chain.

And the questions will be asked either way. You get to choose whether the answer already exists.

This lives wherever you already are, Substack, X, or LinkedIn, drawn from building TikSense.