M-25-21 and M-25-22: The Governing Framework for Federal AI Use and Acquisition

On April 3, 2025, the Office of Management and Budget issued the two memoranda that now define federal AI policy. M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, rescinds and replaces the Biden-era M-24-10 and directs agencies to accelerate AI adoption around three priorities: innovation, governance, and public trust. Its companion, M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government, rescinds and replaces M-24-18 and rebuilds the acquisition side. Both implement Executive Order 14179, Removing Barriers to American Leadership in Artificial Intelligence, signed January 23, 2025.

What M-25-21 requires

The memo applies to executive departments and agencies including independent regulatory agencies, a broader scope than earlier federal AI policy. National security systems are excluded, and portions do not apply to the Intelligence Community.

The structural requirements that matter most for program teams: each agency must retain or designate a Chief AI Officer within 60 days, and the memo’s risk-management core (Section 4 of its implementation appendix) centers on a defined category of high-impact AI: uses whose output serves as a principal basis for decisions with legal, material, or significant effects on individuals. That category carries minimum risk-management practices including pre-deployment testing, impact assessment, ongoing monitoring, and human oversight arrangements.

The lineage is worth noting: M-25-21 keeps Trump-era EO 13960 (Promoting the Use of Trustworthy AI in the Federal Government, 2020) in force and tasks the CAIO Council with advancing its principles. Federal AI governance did not start over in 2025. It was re-anchored.

What M-25-22 requires

The acquisition memo is built on three grounding themes: a competitive American AI marketplace, safeguarding taxpayer dollars through performance tracking and risk management, and cross-functional acquisition teams. Two provisions deserve every vendor’s attention:

First, the training prohibition. Agencies must ensure contracts “permanently prohibit the use of non-public inputted agency data and outputted results to further train publicly or commercially available AI algorithms” absent explicit agency consent. If your commercial model improves on government data by default, your default is now a compliance problem.

Second, anti-lock-in as policy. Solicitations should favor proposals that reduce vendor lock-in through knowledge transfer, data and model portability, clear licensing terms, and pricing transparency. Scoping IP rights “to avoid vendor lock-in” is named explicitly.

Why it matters

The pair establishes something the doctrine community should read carefully: federal AI policy now assumes AI systems are operational actors requiring lifecycle governance, designated accountable officials, defined impact tiers, monitoring, and human oversight, not IT products to be accepted at delivery. That is the same shift the Governed Autonomy Doctrine formalizes for autonomous agents: authorization is the beginning of governance, not the end of it. What the memos do not yet address is runtime behavior, meaning what an AI system may do, across systems, while it operates. As agencies move from predictive models to agentic systems, that gap between acquisition-time control and runtime control is where programs will feel the strain, and where the maturity question, how much autonomy your governance has actually earned, becomes an acquisition question too.