
Model Risk Management Just Stopped Covering Your Riskiest Models
Regulators rescinded SR 11-7 in April and carved generative and agentic AI out of its replacement. The riskiest models banks run lost their rulebook, not their liability. Here is what still needs building, and why waiting on the promised guidance is the expensive option.
You shipped generative AI into the parts of the bank that actually matter. A copilot drafts the AML investigation narrative. A model summarizes a credit file and flags the exceptions a human would otherwise have to find by hand. An agent triages fraud alerts before anyone on the team sees them. The board asked for AI and you delivered it, and the upside was never imaginary: McKinsey puts the prize at $200-340 billion a year for banking. Nobody wanted to be the institution that sat this one out.
Then, in April, the framework you would have used to govern all of it quietly disappeared.
On 17 April 2026 the Federal Reserve, OCC and FDIC rescinded SR 11-7, the model risk management guidance that had told banks how to build, validate and monitor their models since 2011, and replaced it with a shorter, principles-based version: the Fed's SR 26-2, the OCC's Bulletin 2026-13. Buried inside is a line that deserved its own headline. Generative and agentic AI are explicitly out of scope, on the grounds that they are "novel and rapidly evolving." A request for information is promised at some unspecified point down the road.
Read that as the person who actually owns the deployment. The single riskiest category of model you run, the non-deterministic one that can hallucinate a figure into a credit memo or take an action nobody reviewed, is the one your regulator just declined to write rules for. That is not relief. It is a vacuum, and you are standing in the middle of it.
The exemption is from a document, not from accountability
The tempting read is "good news, less paperwork." Resist it. The carve-out removes a rulebook, not the liability the rulebook existed to manage.
Nothing about the underlying exposure changed. A generative model that touches a lending decision still answers to the Equal Credit Opportunity Act and fair-lending scrutiny. One that talks to a customer still answers to UDAAP and the CFPB. Anything adjacent to a securities communication still answers to FINRA, whose 2026 oversight priorities put AI governance, recordkeeping and assigning human accountability near the top of the list. Safety-and-soundness examination did not pause for the rewrite. And the moment a generative system feeds a traditional, in-scope credit or capital model, its errors become that validated model's errors too.
So the examiner in your next review is not going to be reassured by a footnote. The question they ask, show me how you know this model is sound, how you watch it, and what you do the moment it is wrong, is the same question SR 11-7 always asked. The agencies just stopped telling you the answer for the models where the answer is hardest to give.
We've made this argument before about where the real work in an AI deployment lives: the model is the easy 15%, and the operating layer is the hard 85%. Model risk management is that 85% in a banking accent, development you can defend, validation with the standing to fail a release, monitoring that catches drift before a customer does, governance that names an owner instead of a committee. April took the instructions away and left the obligation sitting exactly where it was.
Most banks can't find their own kill switch
Here is how wide that gap already runs. A mid-2026 survey of banking risk professionals asked where AI governance was weakest, and the two answers that dominated were the same two things: a reliable way to shut a misbehaving model down, and a reliable way to report its failure once it happened. Model governance and validation also outranked fairness and explainability as the top barrier to scaling AI further. Read plainly, most of the institutions in that survey cannot reverse a bad deploy or get ahead of a regulator finding out about one.
That is not a problem you write your way out of with another committee. A kill switch is a deploy you can reverse on demand. Failure reporting is an audit trail that already captured what the model did, when, and on whose behalf, the same kind of trail our regulated observability stack post argued you need, pointed at a model instead of a microservice. Those are engineering artifacts. No governance language conjures them after the incident has already happened.
What carries over, and what generative AI breaks
The SR 11-7 pillars survive the rescission because they were never really banking-specific rules. They were good engineering practice wearing a supervisory letter. Conceptual soundness. Independent validation with effective challenge, meaning a reviewer who actually has the standing to fail the model. Ongoing monitoring. Governance with a named owner. The revised guidance keeps that spirit, and your generative systems need it whether the footnote names them or not.
What breaks is the old validation playbook, and this is the part teams underestimate most. A traditional credit model has a documented specification you can trace any test back to. A generative model has weights, a prompt, and a distribution of behaviors nobody authored line by line. It is non-deterministic, so the same input can produce different outputs on different days. It usually depends on a third-party foundation model you do not control and cannot fully inspect. And an agentic version does not just recommend, it acts, which means a single bad classification can cascade across linked systems before anyone notices it happened.
So "validate the model" has to become something concrete. An evaluation harness that measures the behaviors you actually care about, accuracy, fairness across the subgroups a fair-lending examiner will slice by, refusal on out-of-scope prompts, factual grounding, run continuously rather than once at launch. We made this same argument about safety-critical systems generally: validation is a continuous discipline, not a one-time gate. For a model whose behavior drifts with every foundation-model update and every shift in your customer base, a check you ran in June is a historical document by September.
Everyone is writing the playbook except your primary regulator
If you were planning to wait for the promised RFI, notice who else is not waiting. On 25 May 2026 IOSCO published its Supervisory Toolkit for AI Use in Capital Markets, a full-lifecycle framework that, unlike SR 26-2, explicitly covers generative and emerging agentic techniques. The Financial Stability Board has its own work on AI adoption and vulnerabilities in the financial sector, with a consultation that has been open for comment for months. FINRA has already told its own members what it expects.
The pattern repeats across every regulated industry we work in. The binding, prescriptive standard shows up late. The accountability shows up now. The international bodies are converging on exactly the controls SR 26-2 left out, lifecycle governance, validation, monitoring, human accountability, which tells you roughly where your domestic examiner lands the moment that RFI turns into guidance. You cannot buy that evidence the quarter before an exam. Either it has been accruing all along, or you spend the weeks before the examiner arrives interrogating your own commit history and the memories of engineers who have since changed teams, hoping what comes back adds up to a validation record.
What to do this week
You do not need the rulebook rewritten to find out whether your AI would survive the exam that is coming regardless.
- Put every generative and agentic system under model-risk discipline whether the footnote exempts it or not. For each one, write down the intended use and its boundaries, point to the validation evidence, name the monitoring, and name the human who owns the outcome when it is wrong. The exemption covers the guidance, not the risk, and this inventory is a half-day most teams have never spent.
- Find your kill switch, or admit out loud that you do not have one. Can you shut a misbehaving model down, route around it, and report the failure before a regulator asks? If the honest answer is no, that is finding number one, and it is an architecture decision rather than a compliance memo.
- Make validation continuous and independent. Stand up the eval harness that answers whether a version is still sound on data it has not seen, across the subgroups that matter, automatically. Then give the person running it the actual standing to fail a release, not just the title.
All three are things an engineering team can start on Monday, and none of them wait on a single line of new guidance.
Building for guidance that hasn't been written
That is the work we do first. In a First Build, about a month against your real data, we build the generative system and the evidence it will eventually be judged on at the same time: an eval harness that reruns on every version and every foundation-model update, monitoring pointed at the failures that would actually cost you, a shutdown path someone has tested rather than diagrammed, and a one-page record per model naming its use, its limits and the human who answers for it. Built while the system is still small, that is a week of engineering; reconstructed after a supervisory finding lands, it is a quarter of it, done in front of an audience. Here is how we partner.
The agencies have said they are not ready to describe these models. Nobody has told your examiner, your fair-lending counsel or a future plaintiff to wait for them, and each of the three will ask the same thing: what did the model do, how did you know it was sound, and how fast could you stop it. If you could not answer that tomorrow for the generative systems already running, that gap is the conversation worth having. Scope a First Build and bring the model that just lost its rulebook.
Related Insights

Insight
Why Most AI Business Cases Are Built Backwards
A pattern is emerging across healthcare, automotive, and IoT companies. It often starts with a straightforward question from leadership: what is our AI strategy, and what kind of returns should we expect?
Read More
Insight
Your 'Human in the Loop' Is a Rubber Stamp
Most "human in the loop" reviews are a person clicking confirm with no realistic way to catch the model's mistakes. Override rate, not a sign-off policy, tells you whether oversight is real, and regulators are already checking for it. Agentic systems make the measurement more urgent, not less.
Read More
Insight
Trust That Software Works: Verification in the Age of Agents
Agentic tools don't fix delivery pipelines, they flood them with unverified code. We call the gap verification debt. Here's how to close it: separate verification from validation, match assurance to blast radius, and know exactly which agent loops can safely run dark.
Read MoreFacing a problem like this one?
Tell us what the business needs to win and who owns it on your side. If it is a fit, a First Build starts within weeks. If it is not, we will tell you why.