US Federal AI Acceleration Memo

OMB Memorandum M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust

United States

RAI-US-NA-OMBM2521-2025

M-25-21

Effective: 3 Apr 2025
In Force(In Force)As published at whitehouse.gov · checked 8 Sep 2026

US Federal AI Acceleration Memo is In Force in United States as of 8 Sep 2026, according to whitehouse.gov.

PolicyGovernance and OversightRisk ManagementSafety, Testing, and Evaluation
Export PDF

OMB M-25-21 promotes federal AI adoption with agency governance requirements and safeguards for high-impact uses.

Summary

OMB Memorandum M-25-21 accelerates federal agency AI adoption while safeguarding privacy, civil rights, civil liberties, and safety. It replaces M-24-10 and establishes governance, planning, reporting, and high-impact AI risk-management requirements, without creating public rights or obligations.

Full article

Read full text ↗

Overview

OMB Memorandum M-25-21 was issued by the Office of Management and Budget, Executive Office of the President, and signed by Director Russell T. Vought on April 3, 2025. Effective that day, it rescinds and replaces M-24-10, Advancing Governance, Innovation, and Risk Management for Agency Use of Artificial Intelligence. It promotes better public services, government efficiency, protection of taxpayer resources, human flourishing, economic competitiveness, and national security. This policy memorandum combines mandatory administrative instructions with recommendations. Agencies must follow its Appendix while protecting civil rights, civil liberties, and privacy.

The signed memorandum and Appendix cover executive departments and agencies, including independent regulatory agencies, subject to the statutory agency definition and exceptions. They cover new and existing AI developed, used, or acquired by or on behalf of covered agencies. Requirements attach to AI functionality, not automatically to an entire information system. Exclusions cover AI within National Security Systems and certain external regulation, enforcement-target assessment, general testing-standard, and research activities. Agency AI used to conduct enforcement remains covered even when assessment of AI as an enforcement target is excluded.

Definitions

Section 5 adopts the AI definition in Section 238(g) of the John S. McCain National Defense Authorization Act for Fiscal Year 2019. Its interpretation includes machine learning, deep learning, reinforcement learning, transfer learning, and generative AI. Few parameters, limited training data, or low technical complexity do not alone exclude a system. Autonomous, partially autonomous, and non-autonomous systems may qualify, with or without human oversight. Robotic process automation, systems governed solely by human-defined rules, and systems learning only through exact repetition of an observed practice fall outside this interpretation.

High-impact AI has outputs serving as a principal basis for decisions or actions with legal, material, binding, or significant effects on civil rights, civil liberties, privacy, access to education, housing, insurance, credit, employment or other programs, critical government resources or services, health and safety, critical infrastructure or public safety, or strategic assets or resources. An AI model is an information-system component using computational, statistical, or machine-learning techniques to produce outputs from inputs; model weights are numerical parameters shaping those outputs. A significant modification changes an application or deployment context enough to make earlier evaluations, training, or documentation misleading. AI maturity is organizational capacity to adopt AI responsibly, manage risks, and comply with relevant authorities.

Governance and Institutional Framework

Every covered agency must retain or designate a Chief AI Officer within 60 days under Appendix Section 3(a)(i). A CIO, CDO, CTO, or comparable official may serve with significant AI expertise. CFO Act agency CAIOs must occupy Senior Executive Service, Scientific and Professional, Senior Leader, or equivalent positions; other CAIOs must be at least GS-14 or equivalent. Sufficient authority and senior placement must support regular engagement with leadership. Duties include advising agency heads, promoting responsible adoption, maintaining inventories, coordinating legal compliance, overseeing high-impact AI, and supporting workforce readiness and AI resource planning. Agencies must notify OMB within 30 days of a CAIO change or vacancy.

CFO Act agencies must convene AI Governance Boards within 90 days; existing bodies may satisfy this duty. A Deputy Secretary-level official or equivalent chairs the board, with the CAIO as vice-chair. Relevant functions include technology, cybersecurity, data, budget, legal counsel, privacy, civil rights, and civil liberties. OMB must convene and chair a Chief AI Officer Council within 90 days, including CFO Act agency CAIOs, designated White House and intelligence representatives, and other agencies selected by the chair. The Council coordinates implementation and shared resources and sunsets five years after issuance unless otherwise authorized by the OMB Director.

Key Focus Areas

Innovation measures remove unnecessary barriers to responsible deployment and improve existing investments. CFO Act agencies must publish AI strategies addressing enterprise maturity, impactful use cases, infrastructure, data quality and traceability, workforce capability, risk governance, and future investment. Agencies must strive to scale existing governance tools rather than add bureaucracy. They must proactively share custom-developed code, models, and model weights used in active AI applications across government, subject to legal, contractual, security, privacy, safety, and operational exceptions. Public open-source release is required where practicable and exceptions do not apply.

Risk controls are proportionate to use. Section 6 identifies presumed high-impact purposes, including safety-critical infrastructure, medical functions, law-enforcement assessments, immigration decisions, benefit eligibility, federal employment decisions, and legally consequential translation. Integration into a consequential workflow alone does not necessarily satisfy the principal-basis test. Procurement recommendations address adequate government data rights, protection against unauthorized disclosure or commercial training use, competition, interoperability, and avoidance of vendor lock-in. Workforce recommendations support practical training and recruitment of people experienced in designing, deploying, and scaling AI in high-impact environments.

Implementation Framework

Section 4(b) requires pre-deployment testing and risk-mitigation plans for high-impact AI reflecting expected real-world outcomes and benefits. If source code, models, or data are unavailable, agencies must use alternatives such as querying a service or giving evaluation data to a vendor. A documented pre-deployment impact assessment must cover purpose and benefit, data quality and model capability, potential public impacts, reassessment arrangements, costs and savings, independent review, and signed risk acceptance. The independent reviewer must be within the agency and uninvolved in development; review comments must reach the risk-accepting official.

Accountable officials at the lowest appropriate level may identify, assess, mitigate, and accept risks, but specified CAIO waiver duties cannot be delegated downward. A CAIO may waive individual minimum practices for a specific application or component only through a written, system-specific and context-specific assessment finding that compliance would increase overall safety or rights risks or unacceptably impede critical operations. Limited-scale, limited-duration pilots may be exempt with centrally tracked CAIO certification, opt-in and opt-out participation where possible with adequate notice, and minimum practices where practicable. Neither exception displaces applicable law or general information-system authorization requirements.

Monitoring and Evaluation

Agencies must conduct ongoing testing and, where feasible, periodic human review of high-impact AI for adverse performance and security effects, including possible privacy, civil-rights, or civil-liberties violations. Monitoring must detect unforeseen circumstances and changes in systems, deployment contexts, and associated data. Agencies must mitigate risks, maintain suitable documentation, and develop traceability and transparency processes where possible. Impact assessments require appropriate lifecycle updates and reassessment arrangements after significant modifications. Operators must receive sufficient periodic training, assessment, and oversight to interpret outputs, act appropriately, and manage system-specific risks.

CAIOs must establish measurement and evaluation processes, centrally track applications and determinations, and oversee compliance. Agencies other than the Department of Defense and Intelligence Community must inventory AI use cases at least annually, submit inventories to OMB, and publish public versions. OMB may request documentation through accountability reviews, inventories, or other reporting. Waivers require central tracking, annual certification, and reassessment after significant contextual changes; grants and revocations must be reported to OMB within 30 days under Section 4(a)(iii). Determinations rebutting high-impact presumptions are supplied to OMB upon request. Agencies must publish lawful summaries of determinations and waivers, or indicate that none are active.

Penalties, Liability, and Appeals

The memorandum creates no monetary fines, criminal penalties, private damages scheme, or new public rights or obligations. Administrative correction requires documentation of minimum high-impact practices within 365 days and safe discontinuation of noncompliant AI functionality, subject to qualifying pilots and authorized waivers. Agencies must plan discontinuation when high-impact AI performs improperly and cease use when adequate mitigation is impossible. Applicable laws, OMB policies, existing agency authorities, and liability or review arrangements remain in place.

Section 4(b)(vi) requires timely human review and opportunities to appeal negative impacts when appropriate, not an unconditional entitlement for every AI-supported decision. Agencies may adapt existing processes consistently with law. Remedies should avoid unnecessary individual burdens and follow administrative-burden guidance. Human oversight and intervention must fit the use case; appropriate fail-safes are required when practicable and consistent with existing practices. Agencies must provide end-user and public feedback opportunities where appropriate. Section 3(b)(iv) recommends, rather than mandates, a generative-AI acceptable-use policy with safeguards and oversight within 270 days; its inclusion in the consolidated action table does not change that wording.

Relationship to Other Instruments

M-25-21 implements priorities associated with Executive Order 14179, Removing Barriers to American Leadership in Artificial Intelligence, signed January 23, 2025. It operates alongside Executive Order 13960, the AI in Government Act of 2020, and the Advancing American AI Act. Replacing M-24-10 does not repeal underlying statutes or displace general federal information policy. The Appendix requires coordination with officials responsible for relevant compliance domains. OMB Circular A-130 information-system authorization remains separate from AI risk acceptance; privacy impact-assessment duties remain relevant under existing E-Government Act guidance.

Section 2(c) identifies M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government, issued April 3, 2025, as the companion procurement memorandum. M-25-21 references M-16-21 for reusable and open-source code, M-25-05 and the OPEN Government Data Act for data inventories and sharing, and M-21-06 for regulation of AI outside agency use. National Security Systems operate under other policies, including defense and intelligence instruments identified in its footnotes. Requirements and recommendations are severable: a judicial stay or injunction affecting one provision or application is not intended to disable the rest. Nonconflicting agency-specific principles may continue.

International Alignment

M-25-21 is a domestic federal AI-use policy, not a treaty, international standard, or general private-sector regulatory regime. It establishes no cross-border conformity recognition, international certification, foreign-market access requirement, named foreign-framework mandate, or mutual-recognition arrangement. Its international dimension concerns American AI leadership, competitiveness, national security, and coordination. CAIOs may represent agencies in relevant international bodies, without a general duty to obtain foreign approval or demonstrate compliance with another jurisdiction's AI legislation. Safeguards and administrative duties remain tied to covered agency uses and United States authorities.

Consistent with applicable law, the memorandum favors buying American and maximizing AI products and services developed and produced in the United States. Code sharing and release remain subject to intellectual-property, export-control, classified-information, confidentiality, privacy, safety, and operational constraints. These exceptions do not create an international data-transfer authorization system. Government-wide interoperability, shared resources, and standard-setting participation support broader coordination.

Implementation Timeline

Calendar deadlines below are calculated from April 3, 2025 using the memorandum's stated periods, rather than separately printed calendar dates.

MilestoneDateRequirements
Issuance and replacement of M-24-102025-04-03Signed and effective; implementation periods begin.
Retain or designate CAIO2025-06-0260 days; every covered agency, Section 3(a)(i).
Convene agency AI Governance Boards2025-07-0290 days; CFO Act agencies, Section 3(a)(ii).
Convene Chief AI Officer Council2025-07-0290 days; OMB, Section 3(c)(i).
Publish AI strategies2025-09-30180 days; CFO Act agencies, Section 2(a), with specified Defense and Intelligence Community use-case content exceptions.
Submit and publish compliance plan or no-covered-AI determination2025-09-30180 days; Section 3(b)(ii). Repeat every two years until 2036; memorandum updates also trigger a 180-day requirement.
Revisit internal policies and update as necessary2025-12-29270 days; IT infrastructure, data, cybersecurity, and privacy, Section 3(b)(iii).
Recommended generative-AI policy2025-12-29270 days; agencies should develop the policy, Section 3(b)(iv).
Document minimum high-impact practices2026-04-03365 days; Section 4(a)(i), excluding Intelligence Community elements. Safely discontinue noncompliant functionality, subject to exemptions and waivers.
Publish determinations and waiver reporting2026-04-03365 days in the consolidated action table; publication subject to law and government-wide policy.
Chief AI Officer Council sunset2030-04-03Five years unless otherwise authorized by the OMB Director; this sunset does not apply to the whole memorandum.

Sources and References

Requirements for a company

What an organisation has to do under US Federal AI Acceleration Memo, at a glance. Not legal advice — the table below gives the provision and deadline for each item.

Must do

24
  • +18 more in the table below

Must not do

0

Nothing in this category.

Should do

0

Nothing in this category.

Should not do

0

Nothing in this category.

Who must do what

The obligations under US Federal AI Acceleration Memo, most serious first. Not legal advice — verify against the official text before relying on it.

#WhoRequirementBy whenWhereSeverity
1within 180 days of the issuance of this memorandum—Important
2——Important
3——Important
4——Important
5——Important
6——Important
7——Important
8within 60 days of the issuance of this memorandum—Important
9within 30 days—Important
10——Important
11within 90 days of the issuance of this memorandum—Important
12within 180 days of the issuance of this memorandum or any update to this memorandum, and every two years thereafter until 2036—Important
13within 270 days of the issuance of this memorandum—Important
14at least annually—Important
15within 365 days of the issuance of this memorandum—Important
16——Important
17within 30 days of granting or revoking any waiver—Important
18——Important
19before deploying any high-impact AI use case—Important
20——Important
21——Important
22——Important
23——Important
24——Important

© Regulations.AI — created on 20 May 2026 using Gemini 3.6 Flash · updated on 4 Nov 2026 · reviewed against official sources on 8 Sep 2026 using Gemini 3.6 Flash