OMB M-25-22: what AI vendors should preserve for procurement review.

M-25-22 is a procurement driver, not a vendor certification.

OMB M-25-22 addresses federal acquisition of AI. For vendors, the commercial consequence is increased scrutiny of system capabilities, risk allocation, data handling, testing, oversight, performance, and ongoing support. Agencies and prime contractors may translate that scrutiny into questionnaires, contract requirements, technical evaluations, and evidence requests.

1. Preserve the system and use-case record

Procurement reviewers need a stable description of what is being purchased and where it will operate.

• system purpose, intended users, and deployment environment;
• model, application, and material vendor dependencies;
• data categories processed and prohibited uses;
• known limitations and out-of-scope functions;
• system owner and accountable business contact.

2. Preserve testing and evaluation evidence

Marketing statements about accuracy, safety, reliability, or bias are weak procurement evidence unless they can be connected to a defined test, system version, dataset, acceptance criterion, and reviewer conclusion.

• test plans and acceptance criteria;
• evaluation datasets and version records;
• performance, robustness, and limitation results;
• independent or internal reviewer conclusions;
• exceptions, failed tests, and remediation decisions.

3. Preserve data and model governance evidence

Reviewers may need to understand how data, models, prompts, retrieval sources, and third-party components are governed across the product lifecycle.

• data-source and licensing records;
• training, fine-tuning, and retrieval documentation;
• model and prompt version control;
• change approval and rollback procedures;
• retention, deletion, and customer-data handling records.

4. Preserve human oversight and operational responsibility

A policy stating that a person remains responsible is not enough. The evidence package should identify the role, authority, escalation path, and records generated when human review occurs.

• defined reviewer and approver roles;
• intervention, override, and escalation procedures;
• representative approval or exception records;
• training and qualification evidence;
• responsibility boundaries between vendor, integrator, and customer.

5. Preserve security and tenant-segregation evidence

Security claims should be tied to the actual architecture and deployment offered to the buyer. Representative records can include authenticated access controls, tenant-isolation design, vulnerability handling, incident procedures, and evidence of relevant testing.

The package should distinguish implemented controls from roadmap items, customer responsibilities, and controls that depend on configuration.

6. Preserve change, monitoring, and support evidence

Federal buyers may need to understand what happens after award. Vendors should be prepared to show how material changes are identified, approved, communicated, and connected to reassessment.

• release and material-change records;
• monitoring scope and review cadence;
• incident and complaint handling procedures;
• support, continuity, and exit responsibilities;
• conditions that invalidate prior testing or evidence.

The procurement-ready evidence package

1. Evidence Inventory: what the vendor can actually supply.
2. Mapping Summary: which artifacts support the defined buyer requirements.
3. Gap Register: unsupported, stale, contradictory, or missing claims.
4. Ordered Remediation Plan: what should be corrected first.
5. Chain-of-Custody Statement: how the evidence was collected, verified, and preserved.

Official sources

• OMB memoranda index
• NIST AI Risk Management Framework