Selling AI to US Banks: The New Third-Party Risk Proposal and the CSBS AI Exam Framework, and the Evidence Pack Vendors Need

Two things happened in US banking supervision in September 2026 that matter to any startup selling AI software into banks, credit unions or fintech partners. On 11 September the OCC, the Federal Reserve, the FDIC and, for the first time, the NCUA jointly proposed new third-party risk management guidance that would rescind and replace the 2023 interagency guidance and its supplements. Five days later, on 16 September, the Conference of State Bank Supervisors released an Artificial Intelligence Supervisory Framework for state examiners. Neither creates a new legal duty for you as a vendor, but together they shape the questions your bank customers will put in front of you during procurement and annual review.

Start with the federal proposal. Law firm summaries describe it as a move from process to principles: banks would scale third-party oversight to the likely size and probability of harm from each relationship, instead of treating every vendor as high risk by default. Examiners would be expected to defer to reasonable bank judgments, and the proposal says plainly that the guidance itself is not enforceable. Banks still keep ultimate responsibility for sound risk management. The comment period closes in mid November, with sources citing both 10 and 16 November depending on how they count from Federal Register publication, so check the notice itself if you plan to comment. It is a proposal, not final guidance, and the 2023 guidance stays in place until the agencies finalise.

For a vendor, the practical effect of a risk-tiered model is that your tier decides how much work you face. A bank that classifies your AI tool as low impact may accept a SOC 2 report and a short questionnaire. A bank that puts you in a critical or high impact tier, because your product touches customer data, credit decisions, fraud flags or customer communications, will want far more. The fastest way to land in the lighter tier is to make your data flows, model providers and failure modes easy to understand. Vague answers push reviewers toward the conservative choice.

The CSBS framework is where AI gets specific. It is a discretionary tool that each state regulator can choose to use, and it covers state-chartered banks and state-licensed nonbanks, which together make up most US insured institutions by count. Commentators note that it explicitly reaches generative and agentic AI, the category that federal model risk guidance placed out of scope earlier this year. The examiner workflow, as practitioners describe it, runs from scoping to governance records, then an AI inventory, use case documentation, vendor evidence and control testing. That middle step, vendor evidence, is where your product enters the exam.

Here is the evidence pack we would prepare now if banks are on your pipeline. First, an AI system description per product feature: what the model does, which provider runs it (OpenAI, Anthropic, Google or your own weights), where inference happens, and whether customer data is retained or used for training. Second, a subprocessor list that names every model and voice provider, with the region and the contract terms on retention. Third, a change record for model swaps and version retirements, because examiners will ask how the bank learns that the model behind your feature changed. Fourth, your human oversight design: which actions an agent can take without approval, and what happens when a guardrail fails. Fifth, logs that let the bank reconstruct what the AI did for a given customer on a given day.

Map that pack onto the certifications you already hold rather than building it from scratch. A SOC 2 Type II report covers change management, logical access and vendor management, and many auditors now test how you vet model providers as subprocessors. ISO 27001 gives you supplier relationship and secure development controls. ISO 42001 is the closest fit for what the CSBS framework asks about, because its AI system inventory, impact assessment and operational controls line up with the examiner steps almost one to one. If you already run Vanta, Drata, Secureframe, Sprinto or Thoropass, most of these platforms now ship ISO 42001 control sets, so the incremental work is writing the AI-specific documents and attaching them to controls you already monitor.

Two traps to avoid. Do not overstate what a SOC 2 report says about AI; unless your auditor scoped AI-specific controls, the report is silent on model behaviour, and bank reviewers increasingly know that. And do not treat the federal proposal as a loosening that lets you skip work. Principles-based guidance with examiner deference means the bank must be able to justify its tiering decision, and the easiest justification is a vendor who answered the AI questions clearly and on paper. Our compliance readiness checker and the ISO 42001 and SOC 2 guides on this site cover the control mapping in more detail.

A short plan for the next month. Read the federal proposal summary and decide whether your company or an industry group should comment before the mid November deadline. Build or update your AI system inventory so every feature has an owner, provider, data flow and risk note. Assemble the five part evidence pack above and store it where sales and security can both reach it. Then ask your two largest bank or credit union customers which tier they place you in and what they expect at the next review. That conversation is cheaper now than in the middle of a renewal, and it is the clearest signal of how much of this work you actually need.