By Published On: November 16, 2026

September gave bank management teams two developments worth looking at together.

On September 16, the Conference of State Bank Supervisors published its Artificial Intelligence Supervisory Framework. It gives state examiners a structured way to identify AI use, assess related risks and determine when a deeper review may be useful.

Meanwhile, federal agencies were proposing changes to third-party risk management guidance. Those proposals arrived only a few months after the Fed, OCC and FDIC revised their model risk guidance in April.

It is a complicated picture, particularly for a bank with multiple technology providers, fintech relationships and AI tools turning up in places nobody originally expected.

That complication is precisely why management needs a clear view of the institution’s actual exposure.

 

An examiner resource you can use yourself

The CSBS framework is a discretionary supervisory tool. Each state agency decides how far to incorporate it into its program. It is publicly available and financial institutions can use it to examine their own AI governance.

I like that approach. It gives a bank a practical way to anticipate the kinds of questions an examiner might ask without waiting for another round of formal guidance.

The framework addresses governance, risk management and emerging applications, including generative and agentic AI. The April federal model risk guidance takes a different approach. It focuses on certain quantitative models and places generative and agentic AI outside its stated scope.

Those documents serve different purposes. Treating either one as a complete substitute for institution-wide AI oversight would leave gaps.

 

Follow the use case across its whole chain

Here is where the banking reality becomes interesting.

Suppose an institution introduces an AI-enabled customer service tool through a vendor. The tool draws on account data from the core, searches an external knowledge base and recommends an action for a service employee. Three or four parties may touch the workflow. The customer still sees one bank.

Management should be able to explain which data the tool accesses, what its permissible actions are and who owns the outcome. The bank also needs to understand how the vendor updates the model and how it will recognize a change that affects the risk profile.

That information rarely sits in one department. Operations knows the workflow. Technology understands the connections. Compliance knows the regulatory obligations. Vendor management has the contracts. Someone has to bring those views together.

 

Prepare with an operational inventory

I would start with a compact working inventory rather than another long policy document.

  1. List material AI deployments, including tools embedded in existing vendor services. Identify the business owner and the systems they touch.
  2. Document which decisions the AI may influence or take. Include human checkpoints, monitoring and escalation paths.
  3. Record applicable third-party dependencies and the contractual rights the bank has to information, audit evidence and changes.
  4. Map the material use cases to the supervisory frameworks relevant to the institution’s charter and activities.
  5. Name one person who can reconcile the issues and take unresolved items to senior management.

The final item is where many good efforts become difficult. A committee can provide oversight. Day-to-day accountability still belongs to an identifiable person.

The coming federal guidance may alter some documentation and vendor-management expectations. The bank can make progress on its underlying operating controls today.

Examiners will have questions as AI spreads through banking workflows. A bank that understands its own systems and can explain its decisions will be in a much stronger position than one carrying a policy binder with no operating evidence behind it.

That is the work I would put on the calendar before the next examination cycle.

 

Covered in Don’t Blink, my weekly Banking Intelligence Service brief: Issues #33 through Issue #35.

Share This Story, Choose Your Platform!

Subscribe to Newsletter