The framework published by state bank supervisors last week is voluntary, principles-based and short. It also asks one thing that almost no bank or licensed fintech can yet answer: where AI sits inside the products you bought.
On 16 September the Conference of State Bank Supervisors published an artificial intelligence supervisory framework for state examiners. It comes as five documents: a core examiner guide, a work programme setting out procedures, a supplement for nonbank financial institutions, a risk-tiering worksheet and a list of source documents examiners may request. It covers state-chartered banks and, through the supplement, state-licensed nonbanks: the lenders, money transmitters and payment firms that make up much of the fintech sector.
It is guidance, not rulemaking, and the CSBS describes it as principles-based. That makes it easy to file and move on. It would be a mistake to, for a structural reason: it is currently the most specific statement any US supervisor has made about what a reasonable approach to AI looks like at a regulated financial firm.
Filling a gap the federal agencies left open
In April the Federal Reserve, OCC and FDIC updated their model risk guidance and explicitly carved generative and agentic AI out of it, on the basis that those models are novel and rapidly evolving. The state framework moves into that space. For a state-licensed fintech with no federal prudential supervisor, it is effectively the only examination standard for AI that exists.
The framework is organised around eight questions an examiner works through:
Figure 1
Examiner question | What it actually tests | Where firms tend to be |
|---|---|---|
1. Does the institution use AI? | Whether anyone has defined what counts | Answered, often narrowly |
2. Has it identified where? | An inventory, not a policy | Partial; internal builds only |
3. How does AI touch customers or shape decisions? | Mapping to outcomes, not systems | Rarely documented at use level |
4. Does AI come from vendors or outside platforms? | Third-party sourcing of known AI | Usually known for named AI products |
5. Has it looked for AI embedded in vendor products? | Discovery of AI nobody bought as AI | Mostly not started |
6. Does it use generative AI? | Policy and permitted use | Policies exist; usage data thin |
7. Does it classify AI uses by risk? | Tiering against the worksheet | Depends on whether 2 and 5 are done |
8. Does sensitive information pass through AI systems? | Data flow, including to vendors | Hard to answer without 5 |
Question five is the hinge. Three of the later questions cannot be answered properly until the embedded-AI inventory exists. The "where firms tend to be" column is FinancyHub's editorial assessment, not survey data.
Why the fifth question is the hard one
Most AI inventories at financial firms were built by asking business lines what AI they had bought or built. That produces a list of things purchased as AI: the chatbot, the document-extraction tool, the underwriting model a data science team trained. It misses the AI that arrived inside something else.
The fraud module in the card processor's platform. The transaction-monitoring system whose alert scoring moved from rules to a model two releases ago. The KYC vendor whose document verification now uses a model for liveness checks. The core banking provider that added a generative assistant to its back-office screens and turned it on by default. None of these went through an AI procurement process, because none of them were procured as AI. Many of them changed through a SaaS release note that nobody in risk read.
An AI inventory built by asking what you bought as AI will miss most of the AI you are running.
For a bank, this becomes a third-party risk question: vendors will be asked to disclose where models sit in their products and how changes are communicated. For a fintech, it runs in both directions. If you sell into state-chartered banks, you are the embedded AI in someone else's inventory, and the questionnaires are coming. If you hold state licences, the nonbank supplement means you will answer the same questions about your own stack.
Tiers turn on the use, not the model
The risk-tiering worksheet sorts uses into three tiers. Tier 1 covers internal use with human-reviewed outputs, limited consumer impact, limited data sensitivity and low potential harm from errors. Tier 2 covers consumer-facing or decision-support roles with moderate data sensitivity or exception-based human oversight. Tier 3 covers direct consumer outcomes, sensitive personal data, limited human review, significant operational reliance or material potential harm.
Two things follow. Sophistication does not set the tier: a large language model drafting internal policy summaries is Tier 1, while a comparatively simple scoring model that automatically reduces credit lines is Tier 3. And human review is a deciding variable, which means the design of that review needs to be documented. "A human checks it" and "a human checks exceptions the model flags" land in different tiers. Most firms have not written down which one they do, use by use.
Five things worth doing before the next examination cycle
Rebuild the inventory from the vendor list.Start from every system that touches customer data or decisions, and ask each vendor whether models are used in the functions you rely on.
Ask for change notification.Model changes delivered through routine releases are the main way embedded AI enters unnoticed. Request advance notice of material model changes in contract terms.
Document the human-review design per use.Full review, sample review or exception-only review. It is the variable most likely to move a use between tiers.
Tier conservatively where facts are missing.An undocumented use with direct customer impact is easier to defend as Tier 3 than to argue down after the fact.
If you are a vendor, prepare the disclosure now.A short, maintained statement of where models sit in your product, and how clients are told when they change, will be requested repeatedly.
Voluntary until something goes wrong
The framework creates no new obligation. What it does is give examiners a shared vocabulary and a starting list of documents, and supervisory guidance has a way of becoming the benchmark for reasonableness once a customer complaint, an outage or an enforcement matter puts a specific AI use under review. When that happens, the question will not be whether the firm followed the framework. It will be whether the firm knew the model was there.
Sources and notes. Conference of State Bank Supervisors, artificial intelligence supervisory framework for state examiners, released 16 September 2026, comprising a core examiner guide, a work programme, a nonbank supplement, a risk-tiering worksheet and a source document list; described by CSBS as discretionary, principles-based guidance. Tier definitions and the eight examiner questions are summarised from trade and legal press coverage (including Banking Dive and the Consumer Finance Monitor blog). The April 2026 interagency model risk guidance's exclusion of generative and agentic AI is as reported in that coverage. Details are drawn from secondary coverage rather than the CSBS documents themselves — verify tier wording and nonbank supplement scope against the published framework before relying on it. Individual state adoption and examination practice will vary. FinancyHub has no commercial relationship with any organisation named. Journalism, not legal, compliance or procurement advice. Corrections welcome.
