OMB M-24-18 Federal Contractor AI Compliance Checklist (2026)

Contractors must disclose to the contracting agency all AI used in performance, sufficient to support the agency's AI Use Case Inventory required under EO 14110 and M-24-10. The disclosure must cover the AI vendor or developer, the AI's purpose in the contract, the data classes the AI processes, and whether the AI is rights-impacting or safety-impacting.

Practical preparation: maintain an internal AI bill of materials (AIBOM) for every system that touches a federal contract. The AIBOM should list every model component (foundation model, fine-tunes, retrievers, classifiers, guardrails), provenance, training data summary, and version. The AIBOM is the artefact you hand to the contracting officer; if you cannot produce one in 24 hours, you are not M-24-18 ready.

For each disclosed AI use, contractors must tag whether it is rights-impacting, safety-impacting, both, or neither, using the definitions in OMB M-24-10 Appendix I. Tagging accuracy matters because the rights- or safety-impacting tag triggers all of the heaviest downstream obligations - testing, monitoring, individual notice and appeal, and impact assessment.

Practical preparation: train your solutions architects and proposal team to apply the M-24-10 tagging rubric before contract signature. Document the rationale per use case. Expect the contracting officer to challenge a 'neither' tag on any AI that touches benefits adjudication, hiring, law enforcement, healthcare triage, identity verification, or critical infrastructure control.

For rights- and safety-impacting AI, contractors must provide pre-deployment testing evidence covering performance, fairness, robustness, and security. Testing must be conducted with data representative of the agency's actual operating context, not the model's general benchmarks.

Practical preparation: assemble a pre-deployment test pack per offering, including: benchmark performance with confidence intervals; subgroup fairness analysis across protected classes; adversarial robustness against the OWASP LLM Top 10 and NIST AI 600-1 Generative AI risk taxonomy; prompt injection and data exfiltration tests; and an explicit statement of training data sources sufficient to support the agency's bias evaluation.

Contractors must provide system documentation sufficient for the agency to perform its own NIST AI RMF MAP and MEASURE functions. The required artefacts typically include a model card, a system card, a data card or data nutrition label, and an integration architecture diagram showing where the AI sits relative to agency data, decisions, and human oversight.

Practical preparation: maintain a current model card per model (following Google's model card template or equivalent), a current system card per offering (following Anthropic and OpenAI system card patterns), and a data card per training dataset. Version these artefacts and treat them as contract deliverables, not marketing collateral.

Contractors must disclose and contractually agree to: where agency data is processed (data residency); how long it is retained; whether it is used to train or improve any model; whether it is shared with subprocessors; and whether it is subject to lawful access requests. Most agencies will require an affirmative no-training default and explicit consent for any improvement use.

Practical preparation: stand up a federal data handling addendum that defaults to: no training use; US-region processing; defined retention (typically 30 to 90 days for logs, zero for prompt and response content); enumerated subprocessor list with prior notice for changes; and FedRAMP-aligned encryption and access controls. Reference the addendum in every covered offering.

For rights- and safety-impacting AI, contractors must support human review of AI outputs, with the ability for an agency official to override, reverse, or escalate the decision. The system must record the reviewer identity, action, and rationale in an audit-grade log.

Practical preparation: ensure the offering supports a review or approval workflow with role-based access, decision logging, and rollback. If the offering is a model API rather than an application, document the recommended human-in-the-loop pattern and provide reference architecture for agency implementation.

For rights-impacting AI, contractors must support agency obligations to notify affected individuals that AI was used in a decision affecting them and to provide an appeal path. The notice and appeal requirement flows from EO 14110 and is reinforced in M-24-10.

Practical preparation: provide template notice language, an explanation interface (what factors drove the decision and how a human reviewer can revisit it), and an API or workflow integration that captures appeal requests in the same audit log as the original decision.

Contractors must support post-deployment monitoring of model performance, fairness, and security, and must report incidents within agency-specified timelines. Typical agency contract language requires 72-hour notification of a confirmed AI incident affecting agency data, decisions, or systems.

Practical preparation: instrument the offering with continuous performance, fairness, and drift monitoring; expose monitoring data to the agency via dashboard or API; and maintain a published incident response procedure with named contacts. Map the procedure to the agency's CISO and Chief AI Officer contact list per contract.

M-24-18 requires the contractor to flow down equivalent obligations to subcontractors whose work touches covered AI. In practice, this means cloud providers, foundation model vendors, data labelling vendors, and any third-party API used in the AI pipeline.

Practical preparation: review every upstream contract for the AI offering. Foundation model vendors must accept the same data handling, no-training, and incident reporting terms you accept with the agency. If your foundation model vendor cannot accept those terms, you have a M-24-18 readiness gap that must be disclosed.

For rights- and safety-impacting AI, the agency must conduct and publish an AI Impact Assessment under M-24-10. Contractors must provide the technical inputs - the model card, system card, data card, testing results, and monitoring plan - sufficient for the agency to complete the assessment.

Practical preparation: pre-package the AI Impact Assessment input pack as a single deliverable, mapped against the agency template. The CIO Council and AI.gov publish reference templates that most civilian agencies have adopted with minor variations.

Contractors must support orderly decommissioning, including return or destruction of agency data, deletion of fine-tunes or embeddings derived from agency data, and certification of destruction. This obligation often surprises vendors who assumed standard data deletion clauses were sufficient.

Practical preparation: extend the offering's data deletion procedure to cover model artefacts derived from agency data (fine-tune weights, embedding indexes, retrieval caches). Document the deletion timeline (typically 30 to 90 days) and the certification artefact.