Skip to main content
Simkins & Elgazar
Contact us

GSA's New AI Clause: What It Asks of Federal Contractors

GSA clause 552.239-7001 applies when you sell an LLM tool that handles government data. It bans training on that data and sets 72-hour incident reports.

By Ahmed Elgazar. Published . 8 min read.

Topics: GSA, federal contractors, AI governance, NIST AI RMF, large language models

GSA's new clause, GSAR 552.239-7001, applies when a GSA contract buys an AI tool built on a large language model (LLM) and government data goes into it or comes out of it. If the clause is in your contract, you can't use that data to train models or for your own business, you report incidents within 72 hours, you give 30 days' notice before you swap or materially change a model, and you pass the rules down to the subcontractors doing the AI work. GSA's deviation memo lists October 19, 2026 as the effective date for the part of its regulation where the clause sits.

It also borrows its subcontractor roles from the NIST AI RMF, which matters if you've been treating the framework as optional.

Where did this clause come from?

GSA put a draft out for comment in the Federal Register on June 17, 2026, asking for feedback "before taking future action (e.g., deviation and/or formal rulemaking)." The final text is now in part 552 of GSA's overhauled acquisition regulation, issued as a class deviation under memo RGO-2026-01 and dated September 2026.

Attachment A of that memo lists 10/19/2026 as the effective date for part 552 and says GSA contracting activities "may use the deviated language immediately, but must begin using the deviated language no later than the date listed." The memo adds that "Formal rulemaking will follow including a full notice and comment period."

GSA's notice of planned MAS changes lists the new clause and says "GSA anticipates that MAS Refresh #34 will occur in November 2026."

Does it apply to my contract?

Only when LLM features are part of what the government is buying. The clause applies to an LLM, chatbot, agentic system or LLM-enabled tool "where LLM functionality is a material feature and Government Data will be submitted directly to or produced by the LLM."

Two cases are carved out unless the contracting officer says otherwise. The clause is "self-deleting" when LLM use is limited to your "internal business, back-office, operational, or performance-support systems" that aren't delivered to or accessed by the government, even if those systems touch government data along the way. It also drops out when the LLM feature is incidental to a commercial product whose main purpose isn't AI, and the government doesn't require it, use it or run its data through it.

So a company selling an AI assistant to an agency is inside the clause. Under that wording, a team using an AI tool internally to help draft a deliverable falls outside it. That internal use still needs your own rules. Our post on writing an AI acceptable use policy for a small team covers that side.

What does the clause require?

These are the obligations most likely to change how you work.

ObligationClause paragraphWhat to have ready
No use of government data to train or improve an LLM, inform your marketing or business decisions, or sell it(e)(3)Vendor terms that match, for every model in the stack
Disclose the LLMs used on the contract, due within 120 days of starting work if the contract sets no date(f)(1)A model list, with model or system cards
Report incidents affecting government data within 72 hours of discovery, keep logs and forensic artifacts at least 90 days, and file the CISA incident form(f)(5)An incident runbook with the contracting officer and CISA steps in it
Give 30 calendar days' written notice before adding, replacing or materially changing an LLM or a provider in the chain(i)(1)Someone who watches vendor model changes
When a third-party provider deprecates a model you don't control, tell the contracting officer within five business days of their notice(i)(2)Subscriptions to your providers' deprecation notices
Let the government export its data in open or standard formats, with no added fees or lock-in terms(h)A tested export
At contract end, delete government data and custom work, including "fine-tuned model weights, embeddings, indexes, and caches"(e)(4)(vii)A deletion procedure and a written record of it
Flow the clause down to subcontractors doing AI work and use best efforts on due diligence(a)(2), (d)A subcontractor list with each one's AI role

The clause also beats commercial paperwork. If it conflicts with your terms, including "terms of service, click-wrap, browse-wrap," the text is plain: "this clause takes precedence."

If you fall short, the government can suspend use of the LLM at any time. If it terminates for cause after written notice that you failed to fix, you owe reasonable decommissioning costs, capped at 25% of the value of the affected task or delivery order.

Do federal contractors now have to follow the NIST AI RMF?

Not as a blanket rule. NIST publishes the AI RMF as a voluntary framework, and this clause doesn't order anyone to adopt it. It does use the framework in two places that matter.

First, the flowdown follows the framework's roles. You must flow the clause to any subcontractor that performs "AI Design, AI Development, AI Deployment, or Operation and Monitoring tasks, as described in NIST Artificial Intelligence (AI) Risk Management Framework (RMF) 1.0, Appendix A". Those four headings come straight from Appendix A of NIST AI 100-1, which describes the tasks in each one.

Second, on request you must provide "System documentation consistent with NIST AI Risk Management Framework guidelines," along with information the government needs "to complete an LLM Impact Assessment required by OMB M-25-21".

The clause tells you what due diligence evidence counts. A subcontractor attestation should be "supported where practicable by evidence that is independently verifiable rather than by attestation alone". Existing artifacts such as model cards, SOC 2 reports, FedRAMP packages and ISO/IEC 27001 materials can count, and you don't have to create duplicate documents when those already cover it.

So a contractor selling LLM tools through GSA needs a subcontractor list sorted by those four roles, and documentation it can hand over that maps to the framework.

What happens when nobody vets the AI vendor?

This is a public case, not one of our engagements. In December 2023 the FTC sued Rite Aid over the facial recognition system it ran in hundreds of stores from about 2012 to 2020, which it got from two outside vendors. The setting is retail, but the failure is close to what paragraph (d) of the GSA clause asks primes to prevent: taking on an AI vendor without checking its work.

The FTC's complaint alleges that "Rite Aid did not ask its first vendor for any information about the extent to which the technology had been tested for accuracy". The vendor's contract said: "[VENDOR] MAKES NO REPRESENTATIONS OR WARRANTIES AS TO THE ACCURACY AND RELIABILITY OF THE PRODUCT". When Rite Aid moved to a second vendor, it "once again did not seek, and did not receive, any test results or other information about, assurances of, or evidence of accuracy".

The complaint says employees "recorded thousands of false positive match alerts between December 2019 and July 2020", and that in numerous instances people were put under surveillance, asked to leave or reported to police because of false matches. It also says Rite Aid broke an earlier FTC order because it "routinely failed to use reasonable steps in selecting and retaining service providers capable of appropriately safeguarding personal information they received from Rite Aid".

Rite Aid neither admitted nor denied the allegations. The order filed with the complaint bars it "for five (5) years" from using any facial recognition or analysis system in its stores or online, and requires it to delete the photos and videos it collected plus "any data, models, or algorithms derived in whole or in part therefrom."

Rite Aid signed a vendor contract that disclaimed accuracy and asked for no test results. The GSA clause puts best-efforts due diligence on the prime and wants evidence beyond a signed attestation where practicable. Your AI vendor's answers, and the documents behind them, become part of what you may have to show the government.

Where do you start?

Find the GSA contracts and orders where an LLM feature is part of what the agency buys. For each one, list every model and provider in the chain, tag each subcontractor with its AI RMF role, and collect the documents they already have. Then check each provider's terms against paragraph (e)(3) and set up the 72-hour incident path before you need it.

If you need help with the governance side of this, our AI solutions page covers the AI governance, policy and documentation work we do.

Sources

Talk to a principal.

Small, direct and responsive. Both principals do the work on every project; no layered overhead.

Send a messageDownload the capabilities statement