AI for Federal Acquisition Officers: What's Actually Approved in 2026

If you are a federal acquisition officer or a contracting officer's representative reading this in April 2026, you are working in the most active period for AI policy in the history of the federal procurement workforce. Two OMB memos have re-shaped what you can use, what you must report, and what you have to write into solicitations. The General Services Administration has published an approved AI tools list and several agencies have followed with their own. The FedRAMP marketplace has finally caught up with frontier-grade AI services. Below is the working acquisition officer's guide.

This piece does not give legal advice. It points you to the actual policy documents, summarizes the requirements in plain language, and gives you a practical checklist you can use before you put any AI service on contract or use one to assist your own work.

What changed in the policy stack

The simple narrative: in 2024, OMB issued M-24-10 to govern how federal agencies acquire and use AI. In 2025, OMB issued M-25-21, which extended the framework, sharpened the requirements for "high-impact" AI use cases, and tightened reporting obligations. Through 2025 and into 2026, GSA published an approved tools list, several agencies named their preferred AI vendors under their own pilot authorities, and FedRAMP authorized a growing number of generative-AI services at Moderate and a smaller number at High.

The net effect on you: if you are buying AI as a service, you must verify that the vendor's authorization level matches the data classification of your use case. If you are using AI to assist your own contracting work, you must confirm the tool is on an approved list and that the data you put in is allowed there.

OMB M-24-10 and M-25-21 in plain English

OMB M-24-10 (March 2024) is the foundational memo. It requires every federal agency to:

• Designate a Chief AI Officer (CAIO).
• Inventory their AI use cases publicly.
• Apply specific minimum risk-management practices for "rights-impacting" and "safety-impacting" AI use.
• Buy AI through processes that consider performance, transparency, security, and the protection of civil rights and liberties.

OMB M-25-21 (October 2025) is the extension. It refined the categories, made reporting more operational, set deadlines for documentation, and added specific obligations for the procurement workforce. The headline requirement for acquisition officers: every contract for an AI product or service must be evaluated against the agency's AI risk-management framework before award. The contract file should reflect that evaluation.

Practically, this means three things for you.

One, document. Your contract file needs a memo or determination that names the use case, the data classification, the authorization status of the tool, and the risk tier. This is no longer optional.

Two, ask vendors the right questions in solicitations. The standard questions: training data provenance, model versioning practices, evaluation methodology, security certifications (FedRAMP level, SOC 2, HIPAA if applicable), and the supplier's incident response plan.

Three, verify before you award. A vendor claim is not a verification. The FedRAMP marketplace is the source of truth for authorization status. Check it.

The GSA approved AI tools list

GSA maintains a list of AI tools approved for federal use. As of April 2026, the list is published on the GSA AI Community of Practice resources and is updated quarterly. Approval is not the same as FedRAMP authorization. The GSA list is a curated set of tools that have been vetted for federal-employee use under specific conditions.

Categories on the list typically include:

• Generative AI assistants for writing, summarization, and document drafting.
• Code generation tools approved for federal developer use.
• Search and retrieval tools authorized to ingest non-public agency content.
• Translation services authorized for limited-distribution content.
• Data analysis copilots for tabular data work.

The list is not a free pass. Each entry carries a usage condition. Some tools are approved only for public information. Some are approved for CUI under specific subscription tiers. Some are approved only with named-user accounts under an agency enterprise license. Read the conditions, not just the name.

FAR and DFARS implications

The Federal Acquisition Regulation has not yet incorporated AI-specific clauses across the board, but several developments are worth knowing in 2026.

Section 889 still applies. If a vendor's AI service is built on covered telecommunications equipment from prohibited sources, you cannot buy it. Ask vendors to certify their inference and training infrastructure.

DFARS 252.204-7012 (Safeguarding CDI) and DFARS 252.204-7019/7020/7021 (NIST 800-171, CMMC) apply to DoD contracts. An AI vendor handling CDI on a DoD contract must meet the same NIST 800-171 controls as any other contractor. CMMC 2.0 Level 2 certification is increasingly the floor for any contractor processing CUI on a DoD contract.

Software supply chain transparency. Executive order requirements for software bills of materials apply to AI services too. Ask for an SBOM for the model artifacts, the inference stack, and the data pipeline.

Data rights matter. The contract should be explicit about who owns the prompts, who owns the outputs, who can use the data for retraining, and what happens to the data at end of contract. Many commercial AI terms of service grant the vendor rights you do not want to grant on a federal contract.

FedRAMP and the AI tools that have it

FedRAMP authorization is the single most important compliance signal for an AI service in the federal market. The marketplace at marketplace.fedramp.gov is the authoritative source. As of April 2026, the publicly-listed AI authorizations include both major cloud-vendor offerings (Microsoft 365 Copilot for Government, Google Gemini for Workspace Government, AWS Bedrock for GovCloud) and several specialized vendors. Authorization levels matter:

• FedRAMP Low: public information only.
• FedRAMP Moderate: CUI authorized in most cases. The most common authorization level for federal AI use.
• FedRAMP High: required for high-impact systems including financial, law enforcement, and certain healthcare workloads.
• DoD IL2 / IL4 / IL5 / IL6: required for DoD-specific workloads. IL5 is the floor for most CUI work in DoD; IL6 is required for classified.

Verify before you award. Verify before you use. The FedRAMP marketplace lists the authorization level, the date, the agency that sponsored the authorization, and the scope. A vendor saying "we are FedRAMP authorized" without specifying level and scope is not enough.

IRS Pub 1075 crossover

If your acquisition involves federal tax information (FTI), IRS Publication 1075 imposes additional security requirements that overlap with but extend beyond FedRAMP. AI services that touch FTI must meet 1075's controls including specific encryption-at-rest, access controls, audit logging, and personnel screening. Most general-purpose generative AI services are not yet 1075-compliant in 2026. If your contract touches FTI, you need either an FTI-specific service or a deployment pattern that excludes FTI from the AI service entirely.

The pattern many agencies adopt: a redaction step before AI ingestion. Strip identifiers, mask amounts, and use the AI only on de-sensitized text. This is a workable pattern but it must be documented and the redaction step itself must be reliable.

Do-this / don't-do-this checklist

Practical workflow for the next acquisition

If you are about to buy an AI service in 2026, here is the workflow that keeps you on solid ground.

1. Define the use case in writing. Two paragraphs: what the AI is doing, what data it touches, who uses the output, and what happens if it goes wrong.
2. Classify the data. Public, FCI, CUI, FTI, classified, or other? Do not skip this.
3. Tier the risk. Is this a "rights-impacting" or "safety-impacting" use under M-24-10? If yes, additional controls apply.
4. Find candidate vendors. Filter by authorization level matching your data class.
5. Issue solicitation with the right clauses. Section 889, FedRAMP authorization at the required level, data rights, SBOM requirement, incident response.
6. Verify before award. Pull authorization from the marketplace, save the screenshot, file it.
7. Document. Memo for the file: use case, data class, risk tier, vendor selection rationale, FedRAMP verification.
8. Train your end users. The contract is only as good as the users who follow the rules. Brief them on what data is and is not permitted.

Frequently asked questions

Q: Can my staff use ChatGPT to draft acquisition documents?

Only if your agency has approved a specific ChatGPT product (e.g., ChatGPT Enterprise under an agency enterprise agreement) for that use, and only with data classified within the authorization scope. Public, free ChatGPT is not authorized for non-public agency content.

Q: Is GitHub Copilot approved for federal developers?

GitHub Copilot for Business and GitHub Copilot Enterprise have specific federal offerings. Verify the authorization level for your agency before allowing it on agency code that includes CUI.

Q: Where is the GSA approved AI tools list published?

GSA's AI Community of Practice publishes resources at GSA.gov AI pages. The list is updated quarterly. Confirm with your agency CAIO that the list applies to your agency or that your agency has its own list.

Q: Does FedRAMP cover the AI model itself or just the cloud service?

FedRAMP authorization covers the cloud service offering, which includes the inference infrastructure, data handling, and operational controls. Model-specific evaluations (training data, bias testing, evaluation methodology) are typically addressed in the agency's own AI risk assessment, layered on top of FedRAMP.

Q: What if I need an AI tool that does not have FedRAMP yet?

Two paths. One, ask the vendor about their FedRAMP roadmap and timeline. Two, work with your CAIO and CIO to use the FedRAMP "agency authorization" pathway if the use case justifies it. Do not bypass authorization on the assumption it will arrive later.

Q: Can I use AI to read responses to a solicitation?

You can use AI to assist drafting questions, summarizing technical content, or organizing information. Source-selection decisions remain human. The agency's source-selection plan should be specific about where AI is used in the evaluation process and how the human evaluator validates AI-assisted outputs.