Skip to main content
The First Picture of an AI Examination Came From the States

The First Picture of an AI Examination Came From the States

Agentic Assets

Agentic Assets Research Team

Agentic Assets Research

September 24, 2026

6 min read

ai-governancemodel-riskregulationcre-lendingexaminationsai-agents

Through most of 2026 the question "what will an examiner actually ask about our AI" had no published answer. Institutions deploying research and underwriting agents were inferring expectations from exam chatter and from frameworks written for other purposes. In September that changed, and not from where most lenders were watching.

On September 16, 2026 the Conference of State Bank Supervisors released its AI Supervisory Framework, a resource for state examiners assessing the use and risks of AI at state-chartered banks and state-licensed nonbank financial institutions. The detail that matters for anyone building an AI control program is that CSBS released it publicly, and said so explicitly: the framework "helps provide clarity to regulated financial institutions on the general approach, types of questions, and the information that a state examiner may request regarding the institution's AI-based products, services, and tools."

That is a published view of the examination, offered to the examined.

What it is, and what it is not

CSBS describes the framework as a discretionary tool that helps examiners identify and understand AI at an institution, assess the associated risks, and decide when a deeper review is warranted using existing supervisory resources. It is designed to scale to each institution's size, complexity, risk profile, and use of AI. CSBS President and CEO Brandon Milhorn described it as "a principles-based approach to governance intended to help financial institutions explore and implement AI with additional confidence."

It is built on existing risk management work rather than new doctrine, drawing on the National Institute of Standards and Technology's AI Risk Management Framework, the Cyber Risk Institute's Financial Services AI Risk Management Framework, and the United States Department of the Treasury's AI Lexicon.

Two limits are worth stating plainly. It is discretionary, and each state agency decides how far to fold it into its own supervisory program, so adoption will vary. And its direct reach is state-chartered banks and state-licensed nonbanks, which means a nationally chartered lender is not examined against it.

Why a federally supervised lender should still read it

Because the federal picture remains deliberately incomplete. When the banking agencies issued revised model risk guidance on April 17, 2026, the Office of the Comptroller of the Currency release stated that generative AI and agentic AI models "are novel and rapidly evolving" and that "they are not within the scope of this guidance." The agencies said they plan to issue a request for information addressing model risk management and bank use of AI.

So the institutions with the largest AI programs are the ones with the least written guidance aimed at them. A framework built for state examiners, grounded in NIST and Treasury material, is the closest available articulation of the questions. Reading it as a preview costs nothing. Waiting for the request for information costs optionality.

The commercial real estate version of the same gap

Parts of the CRE market are drawing the same conclusion from the private side. In a September opinion column in Commercial Observer, Arunabh Dastidar, co-founder and chief executive of the real estate investment platform Leni, argued that lenders and limited partners increasingly need to show which version of a file produced a figure such as net operating income, and predicted an "auditability spread" emerging over twelve to eighteen months, with verifiable systems clearing credit review faster than opaque ones. That is a vendor's forecast rather than evidence, and it should be read as one. The underlying operational point stands on its own: provenance is becoming a condition of speed, not a compliance afterthought.

What to do with this before the federal guidance lands

  • Build the AI inventory now. Every framework in circulation starts there, and it is the slowest artifact to assemble retroactively.
  • Tier by use case rather than by model. The same model behind a marketing draft and behind a credit recommendation carries different risk.
  • Make provenance recoverable. For any figure an agent produces, be able to name the source document and its version.
  • Record where a human decided. Not that a human was present, but what they changed.

None of this requires knowing what the request for information will say. All of it will be needed whatever it says.

Share this article