The Federal AI Rules Require a System, Not a Product
A verified plain reading of the live federal AI-use policy: which obligations attach to a use case, which scope lines change them, and why procurement leaves the compliance duty with the agency.
It is safe to say yes to a federal AI use case only if the sponsor can show the decision system behind it: a named chief AI officer chain, a current use-case record, a written high-impact determination, and, where that determination lands high-impact, the documented evidence M-25-21 requires plus a rehearsed path to safely discontinue. A vendor can supply evidence toward that system. No purchase, badge, or human-in-the-loop label satisfies it.
The obligation ladder
Each covered agency keeps a chief AI officer. CFO Act agencies convene a governance board, and compliance plans post publicly through 2036.
Agencies other than DoD and the Intelligence Community inventory AI use cases at least annually, report to OMB, and post a public version.
High-impact status turns on whether output becomes a principal basis for a consequential decision. A human reviewer does not settle it.
High-impact uses carry documented testing, impact assessment, independent review, monitoring, trained operators, appeal where appropriate, and a safe stop.
Owner, briefing, proof
Owner
An agency-named determination owner, risk acceptor, and chief AI officer chain. The duty never transfers to a supplier, whatever the contract says.
Briefing
A one-page read per use case: scope status, must versus should, the high-impact result, and the evidence gap that blocks sign-off.
Proof
A living record holding the testing, impact assessment, independent review, monitoring, and the rehearsed stop path.
What leaders should take from it
EO 14148, not EO 14179, revoked the 2023 AI executive order; EO 14179 directed the rewrite, and M-25-21 expressly replaced M-24-10. M-26-04 adds contract requirements for procured large language models. Advice built on the superseded memo misstates the obligation.
Covered agencies keep a chief AI officer; most inventory their AI use cases at least annually and post public versions; compliance plans continue on a two-year cycle through 2036. Publication proves management attention, not that any listed use case is safe or effective.
M-25-21 says a use can be high-impact with or without human oversight, and presumed categories exit only through written notice to the chief AI officer. A human-in-the-loop label settles nothing.
Pre-deployment testing, a lifecycle impact assessment with independent review and signed risk acceptance, monitoring, trained operators, intervention, and appeal where appropriate. Noncompliant uses must be safely discontinued. Waivers exist, but only as written, tracked, publicly summarized decisions.
M-25-22 asks the agency to frame foreseeable uses and make the initial likely-high-impact call; suppliers provide documentation to support that assessment, not a certification. GAO's count of reported federal use cases nearly doubled to more than 1,100, which shows scale, not proven safety or value.
The verified base establishes what the memos require and where the duties sit. It does not establish that the required artifacts reduce real-world errors, rights harms, or safety incidents; the briefing holds that outcome question open. It also cannot show whether the promised human review and appeal paths work for the people they cover, and GAO reviews have found inventory quality and completeness uneven. Read published counts and plans as a floor for questions, not proof of safe operation.
First moves before hiring anyone
Capture the intended decision, principal-basis analysis, owner, scope and data boundary, high-impact result, evidence, and review date. The record turns the inventory into a control surface instead of a spreadsheet.
Mark M-25-21 must language, should language, exemptions, pilots, and waiver candidates. Never package optional governance features as if the memo requires them.
Use M-25-22's foreseeable-use analysis to request the supplier documentation an impact assessment needs. The determination, risk acceptance, and stop decision stay with the agency.
Test against real-world outcomes, name the independent reviewer and risk acceptor, train operators, define intervention and appeal, and rehearse safe discontinuance. Human in the loop is not a design specification.
Request capability limits, data provenance and handling, performance results, test access or output evidence, monitoring support, and the M-26-04 contract terms where a large language model is procured. Treat any vendor compliance assertion as a claim to evaluate.
Start with one consequential use case and run it through the determination and evidence test. If the gap is material, widen to a readiness look at the inventory, determinations, waivers, and stop paths across the portfolio. Build the operating cadence only when the sponsor wants it run.
Claim ledger
- OMB rescinds, supersedes, materially amends, or issues binding implementation instructions for M-25-21, M-25-22, or M-26-04.
- A new executive order changes the agency AI governance baseline.
- OMB's detailed inventory, compliance-plan, or public-summary instructions change the required fields or the disclosure boundary.
- GAO, an inspector general, or OMB publishes cross-agency evidence that determinations, waivers, or minimum practices diverge from this plain reading.
- A well-designed government study measures whether the required assessment, oversight, or appeal controls change rights, safety, accuracy, or service outcomes.