---
title: What does OMB M-25-21 require for AI agents in US federal agencies?
url: https://opencontroller.lyzr.com/blog/omb-m-25-21-and-federal-ai-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# What does OMB M-25-21 require for AI agents in US federal agencies?

**OMB M-25-21 treats an AI agent like any other AI use case. Agencies must list it in their AI use case inventory, decide whether it is "high-impact", and if so apply minimum risk management practices: testing, an impact assessment, monitoring, human oversight and a route to appeal. Agencies had until 3 April 2026 to comply or stop using non-compliant high-impact AI, unless their Chief AI Officer granted a waiver.**

The memo, "Accelerating Federal Use of AI through Innovation, Governance, and Public Trust", was issued in April 2025 and replaced M-24-10. A companion memo, M-25-22, covers AI acquisition. Together they shape what agencies ask of anyone supplying AI agents.

## What does the memo require of agencies?

| Requirement | What it means | Agent angle |
|---|---|---|
| Chief AI Officer | Senior official accountable for AI use | Owns waiver decisions for agents |
| AI strategy and compliance plan | Published plans on how the agency will meet the memo | Should describe how agents are governed |
| AI use case inventory | Annual public inventory of AI use cases | Each agent with a distinct purpose is a use case |
| High-impact determination | Decide which uses meet the high-impact definition | Agents that act on rights or safety decisions |
| Minimum risk practices | Required controls for high-impact AI | See next section |
| Waivers | CAIO may waive a practice if compliance would create an unacceptable impediment to critical operations | Must be documented and reported |

## When is an AI agent "high-impact"?

High-impact AI is AI whose output serves as a principal basis for decisions or actions with legal, material, binding or significant effects on rights or safety. For agents, the questions are:

1. Does the agent's output decide or trigger something about a person, such as benefits, eligibility, enforcement or employment?
2. Does it take actions with safety consequences, such as controlling physical systems or emergency response?
3. Is a human really deciding, or approving the agent's output by default?

The third question matters. Reporting in 2026 noted that some agencies reclassified use cases as not high-impact, which removes them from the requirements. If a human approves nearly everything the agent proposes, the agent is effectively the principal basis. Document why you reached your determination.

## What are the minimum risk management practices?

For high-impact AI, agencies must apply practices including:

1. **Pre-deployment testing** in conditions that reflect real use.
2. **An AI impact assessment** covering purpose, expected benefits, data quality, potential impacts and costs, with independent review.
3. **Ongoing monitoring** for performance and adverse impacts after deployment.
4. **Adequate human training and assessment** for the people who operate or oversee the system.
5. **Human oversight, intervention and accountability**, including fail-safes that minimise harm.
6. **Consistent remedies or appeals** for people affected by decisions.
7. **Feedback mechanisms** for end users and the public.

Agents make items 3 and 5 harder than for a static model. An agent can call tools, chain steps and change behaviour when its prompt or tools change. Monitoring must cover actions, not just model outputs, and fail-safes must be able to stop an agent mid-task. Our guides on [AI impact assessments](/blog/ai-impact-assessment-for-ai-agents) and [evidence for human review](/blog/evidence-for-human-review-of-ai-agent-outputs) go deeper.

## What evidence should agencies and suppliers keep?

A practical evidence pack per agent:

- Inventory entry with purpose, owner, model, tools and data sources.
- High-impact determination with reasoning, signed by the accountable official.
- Pre-deployment test and red-team results, repeated on each version change.
- Impact assessment and independent review record.
- Monitoring dashboards and alert history for performance and adverse impacts.
- Records of human interventions, overrides and stops.
- Appeals and feedback logs, with outcomes.
- Any waiver, with the CAIO's rationale and review date.

The memo binds agencies, not vendors. But agencies pass these needs to contractors through acquisition terms, so suppliers of agents should be able to hand over this evidence. Mapping it to the [NIST AI RMF](/blog/nist-ai-rmf-for-ai-agents) helps, since most agency compliance plans reference it. Policy changes quickly in this area, so check the current OMB guidance and your agency's latest compliance plan before relying on any single date.

## How Opencontroller does this

Opencontroller discovers agents, models and API keys across AWS, Azure and Google Cloud and keeps them in a governed registry, which feeds the use case inventory. Multi-turn and red-team evals support pre-deployment testing; LLM judges scoring production traffic and live groundedness and task-adherence guardrails support ongoing monitoring. Rogue-agent handling can quarantine or kill an agent automatically, a practical fail-safe. Policy as code maps controls to NIST and other frameworks. This produces evidence for M-25-21 practices; it does not make a use case compliant on its own. See our guide to [mapping agent controls to frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Are AI agents covered by OMB M-25-21?**
Yes, as AI use cases. An agent is high-impact when its output is a principal basis for decisions with significant effects on rights or safety.

**What was the deadline for high-impact AI under M-25-21?**
3 April 2026, 365 days after issuance, unless a waiver applied.

**Do federal contractors have to follow M-25-21?**
Not directly, but agencies pass requirements to contractors through acquisition terms, so suppliers should expect evidence requests.

## Key facts

OMB Memorandum M-25-21, "Accelerating Federal Use of AI through Innovation, Governance, and Public Trust", was issued in April 2025 and replaced M-24-10. It requires agencies to keep Chief AI Officers, publish AI strategies and compliance plans, maintain annual AI use case inventories, and apply minimum risk management practices to high-impact AI: AI whose output is a principal basis for decisions with legal, material, binding or significant effects on rights or safety. Practices include pre-deployment testing, AI impact assessments, ongoing monitoring, human training and oversight, remedies or appeals, and end-user feedback. Agencies had to comply, or stop using non-compliant high-impact AI, by 3 April 2026; Chief AI Officers can grant waivers, which must be reported. AI agents are covered like any other AI use case. Opencontroller provides evidence for these practices through discovery, a governed registry, evals, LLM judges on production traffic, guardrails and rogue-agent quarantine and kill. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How does the NIST AI RMF apply to AI agents?](https://opencontroller.lyzr.com/blog/nist-ai-rmf-for-ai-agents.md)
- [How do you run an AI impact assessment for an AI agent?](https://opencontroller.lyzr.com/blog/ai-impact-assessment-for-ai-agents.md)
- [What evidence should you keep when a human approves an AI agent's output?](https://opencontroller.lyzr.com/blog/evidence-for-human-review-of-ai-agent-outputs.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
