What OMB M-25-21 Still Requires After the April 2026 Deadline

Yahia Aboushadi
Director of Brand and Marketing, Pyramid Systems
Reading time:
8 min.

An AI tool that drafts a briefing and one whose score decides who gets a benefit are both covered by OMB M-25-21, the Office of Management and Budget memo on how agencies govern their own AI. Only the second is likely to need a signed impact assessment before anyone relies on it. Every one-time deadline in the memo has now passed, so the questions that matter are what recurs and how a new use case clears the high-impact gate.

Who and what M-25-21 covers

M-25-21, "Accelerating Federal Use of AI through Innovation, Governance, and Public Trust," was issued April 3, 2025 and rescinded and replaced M-24-10. It applies to executive branch agencies, including independent regulatory agencies, for AI they develop, use or acquire. It does not cover AI used as a component of a National Security System, and the Intelligence Community is excluded from its high-impact risk management sections.

It governs AI functionality, not the whole system around it. The definition of AI excludes robotic process automation and systems whose behavior is "defined only by human-defined rules," and the AI system definition excludes common commercial products with embedded AI, "such as a word processor or map navigation system." Section 2 covers innovation, sharing and procurement recommendations, Section 3 governance roles and Section 4 risk management for high-impact AI.

Which OMB M-25-21 deadlines have passed, and what recurs?

All of the one-time deadlines have passed; the last fell on April 3, 2026. The memo counts each deadline in days from its issue date, which puts them on these dates.

Requirement Section Due After that
Retain or designate a Chief AI Officer (CAIO) 3(a)(i) June 2, 2025 Notify OMB within 30 days of a change or vacancy
CFO Act agencies convene an AI governance board 3(a)(ii) July 2, 2025 Standing body
CFO Act agencies publish an AI strategy 2(a) September 30, 2025 One-time
Submit and publish a compliance plan 3(b)(ii) September 30, 2025 Every two years until 2036, and after any update to the memo
Update IT, data, cybersecurity and privacy policies (and, the memo says "should," a generative AI policy) 3(b)(iii), (iv) December 29, 2025 One-time
Document minimum practices for high-impact AI 4(a)(i) April 3, 2026 Applies to every new high-impact use
Publicly report determinations and waivers 4(a)(iv) April 3, 2026 Kept current
Report waivers and determinations to OMB 4(a)(iii) Annually Also within 30 days of granting or revoking a waiver
Publish an AI use case inventory 3(b)(v) Annually Yes

Counting two years from the first compliance plan puts the next round at about September 30, 2027; that date is our arithmetic, not one OMB has announced. The CAIO for a Chief Financial Officers Act (CFO Act) agency must sit at the Senior Executive Service, Scientific and Professional or Senior Leader level, or equivalent. The governance board is chaired at the Deputy Secretary level, with the CAIO as vice-chair, and must include procurement and program office representatives when relevant.

For the 2025 inventory, OMB's reporting guidance set December 22, 2025 for submissions, January 28, 2026 for public posting at each agency's /ai page and April 3, 2026 for public waiver summaries. It also let agencies report 20 kinds of common commercial AI, such as drafting, meeting transcription and code generation, collectively instead of one by one. A use determined to be high-impact still has to be reported individually.

What makes an AI use high-impact under M-25-21?

AI is high-impact when its output "serves as a principal basis" for decisions or actions with a "legal, material, binding, or significant effect" on civil rights, civil liberties or privacy; access to education, housing, credit, employment and other programs; access to critical government services; health and safety; critical infrastructure; or strategic assets. That is the Section 5 definition, and the determination belongs to the agency.

Section 6 adds 15 categories presumed to be high-impact. Several touch civilian programs directly: adjudicating applications for federal benefits and loans, detecting fraudulent use of government services, controlling access to government facilities and deciding the terms of federal employment. An official who concludes that a use in one of those categories is not high-impact must document it in writing to the CAIO, who can revisit it at any time.

Two details decide most arguments. The memo says a high-impact determination "is possible whether there is or is not human oversight for the decision or action," so "a person reviews the output" does not settle it. And footnote 27 says AI can sit inside a high-impact process without being high-impact if its output is not a principal basis for the decision. A summary that helps an adjudicator find documents in a file is a different case from a score the adjudicator is expected to follow. We looked at the same question for acquisition tools in how agencies use AI in federal acquisition.

What must a high-impact use case have before it goes live?

Section 4(b) sets seven minimum practices. Under 4(a)(i), a high-impact use that doesn't meet them must be "safely" discontinued, unless the CAIO has waived a practice or the use qualifies as a certified pilot. The practices:

  1. Pre-deployment testing and a risk mitigation plan that reflect real-world outcomes. Without access to source code, models or data, the agency must test another way, such as querying the service or giving the vendor evaluation data.
  2. An AI impact assessment, completed before deployment, covering purpose and expected benefit, data quality and fitness, privacy and civil rights impacts, reassessment schedule, costs, an independent reviewer's findings and the signature of the official accepting the risk.
  3. Ongoing monitoring for performance, security and adverse impacts.
  4. Periodic training and assessment for the operators who act on the output.
  5. Human oversight, intervention and accountability, with a fail-safe where practicable.
  6. Timely human review and a chance to appeal for people affected by AI-enabled decisions.
  7. A way for end users and the public to give feedback that informs the agency's decisions.

The independent reviewer must be someone in the agency who "has not been involved in the development," and their comments go with the assessment to the person signing. OMB's 2025 reporting guidance turned the practices into inventory fields: deployed high-impact uses report whether each is in place, and the general fields ask whether a use has an associated ATO and whether the system was bought from a vendor, developed under contract or built in-house.

What the 2025 inventory shows

OMB's consolidated 2025 repository, as of April 13, 2026, lists 3,611 individually reported use cases from 41 agency submissions (56 submissions in all, counting those that reported only commercial tools or no AI use), with 445 marked high-impact. The totals cover every stage, from pre-deployment to retired, so they are not a count of systems in use. Nextgov/FCW reported the 3,611 figure as a 105% increase over the 1,757 reported for 2024; because the 2025 guidance let agencies report common commercial AI collectively, counts from different years may not compare directly.

The high-impact count is concentrated. Veterans Affairs reported 215, the Department of Justice 114 and Homeland Security 55, while Commerce and Interior, with more than 200 use cases each, reported none. These are each agency's own determinations, which the memo leaves to the agency, so similar tools can be classified differently from one agency to the next. A vendor selling one product to several agencies should expect several review paths.

How does M-25-21 relate to the ATO and to contracts?

They run side by side. Footnote 22 says AI risk review "is separate from, and does not supersede," the information system authorization under OMB Circular A-130, and footnote 29 says AI systems must still meet A-130's ATO requirements. A system with a current Authority to Operate can still need its AI functionality assessed, and an approved impact assessment does not authorize a system. Our post on how long an ATO takes covers the authorization side.

On contracts, Section 2(d) recommends that agencies keep rights to government data and improvements to it, consider terms that prevent vendor lock-in and bar use of agency data to train a vendor's commercial offerings without express permission. In M-25-21 these are recommendations; its companion memo, M-25-22, covers how agencies buy AI.

From our delivery teams: Start the AI impact assessment when the system security plan starts, not after the ATO package is done. The two can have different reviewers and signers, and the independent reviewer's comments must reach the official accepting the risk before that signature.

Questions to settle before an AI use case goes to review

Most delays come from settling these late. Ask them early, of the program office, the CAIO's process and the vendor:

  • Is there a written high-impact determination, made through the agency's CAIO process, before the design is fixed?
  • If this is a pilot, is it limited in scale and duration, and is the CAIO's certification tracked centrally?
  • Can the agency test the AI with its own data if code and model weights can't be shared?
  • Does the vendor document capabilities, limits, intended use and training data provenance, the items Section 2(d)(ii) suggests agencies track?
  • Can an existing human review or appeals process be extended to AI-enabled decisions, as the memo allows?
  • Is the use in the inventory, with the ATO and vendor fields filled in?
  • What do the data rights and training terms in the solicitation say? How they apply is for the contracting officer to decide.

For delivery support on federal AI projects, see our analytics and AI services.

The Bottom Line

With its one-time deadlines behind it, M-25-21 now works as a standing gate for new AI uses, and the gate is the high-impact determination. Program teams should get that determination in writing early and plan the impact assessment beside the ATO package. Vendors should be ready to support agency testing and inventory reporting.

FAQ

Is OMB M-25-21 still in effect?

Yes. As of October 28, 2026, OMB's memoranda page still lists M-25-21, and no later memo on that page rescinds it. It replaced M-24-10, the earlier AI governance memo, when it was issued on April 3, 2025.

Does M-25-21 apply to the Department of Defense?

In part. It does not cover AI used as a component of a National Security System, the Department of Defense (DoD) is exempt from inventorying individual use cases, and DoD is excluded from the requirement to publish an AI strategy. Elements of the Intelligence Community are excluded from the high-impact risk management sections.

Does M-25-21 apply to contractors?

The memo governs agencies' own use of AI, including AI "developed, used, or acquired by or on behalf of" an agency, and says it creates no rights or obligations for the public. Contractors encounter it through contract terms and the testing and documentation an agency asks for. What a given contract requires is for the contracting officer to determine.

Can a Chief AI Officer waive the minimum practices?

Yes, for a specific AI application, after a written, system-specific risk assessment finding that a practice would increase overall risk to safety or rights or create an unacceptable impediment to critical operations. The CAIO cannot delegate that authority, must recertify each waiver annually and must report a waiver to OMB within 30 days of granting or revoking it.

Are AI pilots exempt from M-25-21?

Pilots are exempt from the minimum risk management practices if the pilot is limited in scale and duration, the CAIO has certified it and that certification is tracked centrally, people who interact with it can opt in and out where possible, and the practices are applied where practicable. Pilots still belong in the use case inventory.

Sources

  1. OMB Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust (April 3, 2025)
  2. OMB, 2025 Federal Agency AI Use Case Inventory (repository summary as of April 13, 2026)
  3. OMB, Guidance on 2025 Agency Artificial Intelligence Reporting (copy published by FedScoop)
  4. OMB Memorandum M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government (April 3, 2025)
  5. Nextgov/FCW, Agencies report over 3,000 AI use cases in 2025 (April 16, 2026)

How we researched this. Every number, date and requirement here is checked against the sources listed above, primary documents first, and the post is fact-checked again before publication. When a rule, date or status changes, we update the post and log the change below. Read our editorial standards or send a correction.

Yahia Aboushadi
About the author. Yahia Aboushadi is Director of Brand and Marketing at Pyramid Systems, an employee-owned federal IT firm in Fairfax, Virginia, that has delivered for federal agencies since 1995. . Connect on LinkedIn.

Continue reading

Related Insights

More from Pyramid Systems on federal AI and acquisition.

Two people at a gray table reviewing a printed form, one pointing at a line, with a pen and a tablet nearby

Acquisition

15 November 2026
Best Value Tradeoff vs LPTA, Plus the Third Approach the FAR Overhaul Named

Best value tradeoff vs LPTA under the rewritten FAR Part 15: the six civilian LPTA conditions, DoD's eight, the new highest technically rated approach and the file.

Read more
Source code on a monitor in a dark room, the kind of software that runs agentic AI in government

Federal AI

12 November 2026
Who Signs Off When an AI Agent Acts? Agentic AI in Government Under OMB, GSA and NIST

Agentic AI in government: what OMB M-25-21, GSA's AI clause and NIST's 10 agent identity tenets ask for when software acts, and where human approval belongs.

Read more
A person taking notes with a stylus at a dark desk, the kind of work FAR Part 10 market research asks agencies to document

Acquisition

08 November 2026
FAR Part 10 Market Research Under the Overhaul, and the Plan to Move It to Part 7

FAR Part 10 market research under the Revolutionary FAR Overhaul is now one short section, and a proposed rule would move it to FAR 7.2. What to document now.

Read more
A tall stack of hardbound books on a desk, the kind of reference shelf the FAR overhaul is rewriting

Acquisition

31 October 2026
Where the FAR Overhaul Stands: Deviations, Proposed Rules and FAR Companion 3.0

FAR Companion 3.0, agency class deviations and eight proposed rules: which one governs a buy under the FAR overhaul, and which parts still wait.

Read more