On Sept. 16, 2026, the Conference of State Bank Supervisors published its AI Supervisory Framework, examination materials for state examiners assessing AI at state-chartered banks and state-licensed nonbank financial institutions. The CSBS AI Supervisory Framework is examiner-facing, but CSBS released it in full, so the institutions they visit can read it first. Three state agencies acted within weeks.

Mississippi told money service businesses 12 days later to prepare for AI discussions at their next scheduled examination, starting in 2027. Alabama told the CEOs of its state-chartered banks and trust companies it will incorporate the same elements into scheduled exams starting in 2027. And Connecticut cites, in an amended Nov. 10 reissue, the CSBS and OCC examiner frameworks in one footnote.

None of this creates a new requirement, and every document says so in its own words. What the documents establish is a calendar and a question list. If you run on a state charter or license, 2027 is when the questions arrive.

What the framework is, and where it reaches

CSBS approved the framework in August 2026 through its State Supervisory Processes Committee and Non Depository Supervisory Committee. Its framework page describes a discretionary tool for state examiners “to identify and understand AI at financial institutions, assess associated risks, and determine when a deeper review may be appropriate using existing supervisory resources,” and grounds it in NIST’s AI Risk Management Framework, the Cyber Risk Institute’s Financial Services AI Risk Management Framework, and the Treasury’s AI Lexicon.

The package has five parts: a Core Examiner Guide, an Examiner Work Program, Nonbank AI Supplements, an optional AI Use Case Risk-Tiering Worksheet, and a Source Support Document. The Core Examiner Guide is the working core: “initial scoping questions, a document request list, and procedures for governance and oversight, AI inventory and use cases, and generative AI and emerging use.”

CSBS President and CEO Brandon Milhorn described the framework as “a principles-based approach to governance intended to help financial institutions explore and implement AI with additional confidence,” and the press release invites use: “Financial institutions can use the framework to assess their own AI programs, establish sound AI governance and risk management, and prepare for examinations.”

The reach matters more than the charter list implies. CSBS’s own figure: “State regulators supervise 79% of all U.S. banks and a variety of non-depository financial services.” Cooley itemizes the nonbank lane: “money transmitters, consumer and mortgage lenders and servicers, and debt collectors.” Fintech and crypto operations typically live there, on transmitter licenses these agencies issue rather than on bank charters.

Adoption is explicitly per-state: “Each state agency will determine the extent to which this framework is incorporated into their supervisory programs.” 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.”

Three regulators, three documents, the same disclaimer

Mississippi’s memo, signed by Commissioner Rhoshunda G. Kelly, says the framework “is principles-based and does not create new reporting obligations or supervisory requirements” (Sept. 28, 2026). Both state memos carry the same clarifications: the framework “does not impose new requirements beyond existing law, regulation, or guidance” and “does not function as a mandatory checklist.” Connecticut’s guidance agrees: “These frameworks are supervisory resources and do not establish additional requirements for Regulated Entities” (p.5).

Three supervisors repeating nearly the same disclaimer is the posture talking: expectations surface through examination conversations, not rulemaking. You will be asked to explain how you thought about AI risk.

Steptoe’s walkthrough of the Core Examiner Guide (Oct. 8, 2026) notes that an initial “does not use AI” answer may not end the inquiry: examiners are told to consider checking the vendor inventory, software inventory, approved tools, and recent product changes. If nothing is inventoried, the honest scoping answer is that nobody knows what is there.

The three actions, and the arithmetic behind them

Connecticut is the outlier in depth: the others are memoranda introducing the CSBS framework, while Connecticut’s Department of Banking, under Commissioner Jorge Perez, issued standalone regulatory guidance with governance expectations, written-plan components, and examination posture.

The dates: Connecticut’s original guidance six days after the release (Sept. 22, 2026), Mississippi’s memo 12 days after (Sept. 28), Alabama’s 49 days after (Nov. 4), Connecticut’s amended reissue 55 days after (Nov. 10). All three actions landed within eight weeks of the framework going public. Guardrails: adoption is per-state by design, and these are the only three state actions verified as of this writing. What the dates support: Mississippi’s and Alabama’s 2027 exam incorporations, and Connecticut’s governance reviews during its examinations.

Connecticut’s written plan is the working blueprint

Connecticut’s guidance is the closest thing to a build spec, because its components overlap with what the CSBS materials have examiners looking for everywhere. Calibrate scope first; it decides what counts. “AI System” means “an automated system or tool that uses AI and materially influences business operations, risk management, or consumer-facing decisions of a Regulated Entity,” and it includes AI functionality “embedded within a broader software product or service, including third-party or software-as-a-service tools” (footnote 1, p.1). Tools that use AI only incidentally are out; “a model that infers credit risk from historical data is Artificial Intelligence; a spreadsheet that applies fixed debt-income thresholds is not” (footnote 2, p.1).

Inside that frame, the Written AI Governance Plan carries seven components (section D.2):

  1. Lifecycle, from acquisition or development through testing, deployment, monitoring, material changes, and retirement, with drift, change-management, and deactivation controls.
  2. Inventory, risk assessment, and management, with cybersecurity and data governance, tiered by materiality, data sensitivity, and consumer impact.
  3. Compliance mechanisms for applicable law, monitoring for unfair or discriminatory outcomes, consistent with the Equal Credit Opportunity Act and the Fair Housing Act.
  4. Human oversight of AI-supported decisions, with escalation protocols, accountability, and “the ability to intervene, override, or take corrective action when warranted” (p.5).
  5. Training and competency for board members, senior management, and relevant personnel, proportionate to role.
  6. Third-party risk management for vendor-provided or vendor-supported AI: due diligence, contractual controls, audits, ongoing monitoring.
  7. Documentation “sufficient to support internal governance, supervisory review, and examination readiness” (p.5).

The plan may sit inside the entity’s existing risk management plan, and the department “plans to review AI Systems, AI Governance Frameworks, Written AI Governance Plans, and related policies and procedures during supervisory examinations” (p.5). The Sept. 22 original headed this section “Required Components of the Written AI Governance Plan”; the Nov. 10 revision retitled it “Components of the Written AI Governance Plan,” dropping “Required” while keeping plan reviews on the exam calendar. Supervision by expectation, not obligation.

The reissue also moved the explainability bar. Institutions should be able to “explain in a clear and concise manner whether and how AI Systems materially affect consumers or decisions concerning them” (p.5). The Sept. 22 version asked only whether; the added how shifts the question from knowing that a system matters to narrating its mechanism.

Footnote 7 is where the wiring went in: it cites the CSBS Artificial Intelligence Supervisory Framework and the OCC’s Supervisory Framework on Bank Use of Super Intelligence (Nov. 4, 2026, Bulletin 2026-54). Law-firm coverage of the guidance predates Nov. 10, so this went unremarked; we walked through the OCC framework when it landed. The two documents share a design philosophy: existing risk categories, proportionality, and examiner judgment. Footnote 6 cites the revised interagency model risk guidance (SR 26-2, FIL-15-2026, OCC Bulletin 2026-13), notes “Generative and agentic AI are outside the scope of the revised guidance,” and keeps governance practices in charge of controls for those systems.

For anyone worried about overbuilding, one sentence does the work: “Regulated Entities should address these risks within their existing information security programs rather than through a separate or parallel cybersecurity regime” (p.2).

What examiners will ask

The CSBS component PDFs sit behind download links, so examiners’ exact wording is not quotable here; between the framework descriptions and law-firm walkthroughs, the shape is legible. Steptoe’s summary of the document request list reads like a readiness checklist: AI policies; board reporting; AI-system and use-case inventories; risk assessments and tiering documentation; vendor products with embedded AI; generative-AI policies and approved-tool lists; vendor oversight materials; testing and validation records; consumer-facing AI outputs; and data-use contracts. Plenty of institutions could assemble that folder in a month; many have not.

The sharpest new ground is agentic. Cooley’s reading of the Examiner Work Program describes evaluation procedures for agentic AI systems, covering “permission boundaries, human checkpoints, logging, reversibility and capability to halt actions when needed.” Cooley describes the tiering worksheet, which “assigns each AI use case a risk tier from 1 (lowest) to 3 (highest),” with controls escalating by tier from basic inventories toward independent model validation. The accountability question around agentic AI now has an examination answer forming: where the boundaries are, and who can stop the system. We tracked the accountability side of agentic AI earlier this month; the expectations converge.

Building into programs you already run

The mistake to avoid is a parallel AI regime beside the risk program; Connecticut names that mistake in writing. The inverse failure is overbuild: a thick document nobody maintains gives an examiner more to ask about, not less.

The sequence that survives an exam runs inventory, tiering, controls, documentation, and the inventory is where these programs stall: vendor features arrive silently inside products already owned. An inventory-first build beats a policy-first build: the policy is accurate only after the inventory says what exists.

Third-party work is the same muscle rebuilt this year under the interagency third-party risk guidance: knowing which vendors embed AI, what data they process, whether it trains their models, and what the contract says about data use, model changes, incident disclosure, and audit access.

Ownership is harder for nonbanks. Venable’s read on the guidance (Oct. 14, 2026) says “Nondepository licensees likely face the steepest climb,” locating the problem an operator would name: easier to validate a model than to answer “who owns the Framework or is responsible for exercising human oversight over a particular AI system.”

A readiness list for the 2027 cycle

The three documents converge on a short list, every item traceable to the framework, the memos, or the guidance:

  1. Scope the AI you actually have, including vendor-embedded AI; a scoping answer with no document behind it is the first gap.
  2. Inventory and tier the use cases: purpose, owner, risk, consumer impact, written rationale per tier.
  3. Refresh vendor diligence for AI: data use and training rights, model change notice, incident disclosure, audit access.
  4. Define the human oversight path: who reviews, who can intervene, and where the record lives.
  5. Keep the document-request folder standing: policies, reporting, testing records, tiering rationales, vendor AI lists.
  6. Write the sentence on whether and how each material system affects consumers; if you cannot write it, an examiner will.
  7. Brief the board: oversight belongs there in all three documents.

Treat the calendar as the deadline. Mississippi and Alabama incorporate framework elements into scheduled exams starting in 2027, Connecticut reviews governance documents during its examinations, and the framework will be updated periodically. No new rule arrives in this lane. The questions now have published answers, and 2027 decides only when they get asked.