The FS AI RMF is the framework your examiners are reading.

What happened. In February, Treasury — working through the AIEOG with the FBIIC and FSSCC — released two companion documents for the financial sector: a shared AI Lexicon and the Financial Services AI Risk Management Framework (FS AI RMF). Neither is binding. Then in April, the banking agencies issued SR 26-2, the revised model risk guidance that superseded SR 11-7 — also non-binding, and explicitly leaving generative AI outside its formal scope while the agencies gather input.

Why it matters to you. "Non-binding" doesn't mean "ignorable." It means the agencies have told you what good looks like before telling you what's required. When your examiner asks how you govern AI use, the vocabulary they bring to that conversation increasingly comes from the Lexicon — terms like AI use case inventory now have an official, shared definition — and the structure they expect comes from the FS AI RMF's risk-management lifecycle. An institution that can map its existing practices onto that structure walks into the exam speaking the examiner's language. One that can't is answering freestyle.

What to do with it. Three practical moves, none of which require a new system:

1. Adopt the vocabulary. Rename your internal AI tracking to match the Lexicon's terms, starting with a formal AI use case inventory. Free alignment with how your regulator talks.
2. Cross-walk, don't rebuild. Take your existing model-risk and vendor controls and map each FS AI RMF element onto something you already do. The gaps that fall out of that exercise are your real to-do list.
3. Mind the generative-AI seam. SR 26-2 excludes generative AI from its formal scope — which means your generative-AI use needs its own documented treatment rather than inheriting model-risk coverage by default.

The institutions that treat non-binding frameworks as early exam prep consistently have calmer exams than the ones who wait for a mandate.