Innorve Academy · AI Oversight ProgramAcademy edition · tool-neutralAs of 2026-09-28Research, not legal advice
NCUA has written no AI rule, and says so. What binds a credit union's AI is the law it already follows: its security program, consumer protection, fair lending, BSA, and a growing set of state laws. Set your profile and see which of these apply to you, what to do, and what evidence an examiner will expect.
00 · Read this first
NCUA's FAQ says it 'has not issued AI specific rules or regulation' and applies existing, technology-neutral rules. Every item here is an existing duty that AI triggers. source
Under 12 CFR 791 subpart D, examiners will not criticize a credit union for non-compliance with guidance such as an FAQ or Letter. Findings rest on law, regulation, or an unsafe or unsound practice (12 U.S.C. 1786); GAO reports NCUA has issued an AI-related document of resolution. source
NCUA's model guidance covers interest-rate-risk models; SR 26-2 / OCC 2026-13 is bank-only, unenforceable and excludes generative and agentic AI. Use it as a proportionate benchmark with a separate GenAI addendum. source
State laws bind where the credit union has members or staff in that state (federal-charter preemption is unresolved for some). GSE terms bind sellers and servicers by contract.
02 · By use case
Most credit unions think in projects, not statutes. Each card lists the binding duties specific to that use, on top of the baseline that applies to any AI, plus the security controls that matter most.
03 · Security
Each control names the binding duty it helps evidence, usually 12 CFR 748.0(b)(2) as examined through the Part 748 Appendix A guidelines, or says plainly that it is a benchmark. The last column is the honest limit: what a tool on the credit union's side cannot see, such as AI running inside a vendor's own cloud.
| ID | Control | Binding duty it evidences | Evidence | Tier | What a CU-side tool cannot see |
|---|
04 · Time
Changes since the first research pass on September 13, then dated events ahead.
Challenge to the 2026 Reg B amendments; briefing runs to Jan 12, 2027. source
AI in employment discipline; CIPA pen-register fix (wiretap claims stay open); chatbot disclosure plus a route to a human. Unsigned bills become law. source
Would narrow revocations by category, allow a designated revocation channel and widen the fraud-alert exemption. source
Deployer duties for automated employment-decision technology follow on Oct 1, 2027. source
Leaves disparate impact to the courts; mortgage AI exposure continues. source
Rules for SB 26-189 (automated decisions) and HB 26-1263 (chatbots). The revised draft promised for Sept 23 had not been posted as of Sept 28. source
91 FR 58536; first interagency version that expressly covers insured credit unions. source
Automated-decision notices, explanations and human review; chatbot duties; CPPA ADMT compliance date. source
Vacatur would restore a Reg B effects test for AI credit models. source
Opt-outs would carry across all robocalls and robotexts. source
PA 26-15 §§ 7–12; no financial-institution exemption. source
Proposed Dec 2025. If finalized, Appendix A content moves to a Letter to Credit Unions and control hooks should cite 748.0(b) directly; the binding 748.0 and 748.1 duties stay.
Rule enjoined; reconsideration NPRM at White House review since Aug 4, 2026.
Could produce the first AI-specific model-risk expectations.
Ordered reported 28-21 on Sept 16, 2026; would raise CFPB supervision to $30B.
Introduced Sept 2, 2026 and framed around AI cyber risk; opposed by credit union groups.
September 2026 target passed without publication.
No drafts released as of Sept 28.
Proposed Jan 14, 2026; comments closed Mar 16; not in the Aug 5 final batch. Would drop the FCU-specific effects test and five banned factors; the Fair Housing Act still applies.
05 · Credibility
Each of these circulates in vendor material, board decks or conference talks, or simply sounds plausible. Use them to test what you are told, and what you write.