On October 19, 2026, the General Services Administration's AI clause takes effect. GSAR 552.239-7001 sets the rules for how contractors handle government data when a large language model (LLM) product is bought under a GSA contract, order or blanket purchase agreement. If your agency buys AI tools through GSA vehicles, or your company sells them, here is what the clause requires, who it covers and what to prepare.
Reading time:
5 min.
What is GSA's new AI clause?
GSAR 552.239-7001, "Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems," is a clause in the GSA Acquisition Regulation. GSA issued it as a class deviation through RGO-2026-01, the memo behind its regulation overhaul, signed by Senior Procurement Executive Jeffrey A. Koses. The full text is in the RGO-2026-01 package on Acquisition.gov (GSAR part 552).
The clause started as a draft. GSA published it for comment in the Federal Register on June 17, 2026, held a listening session on July 14 and took comments through August 3. That notice said the clause may be used in GSA's governmentwide contracts, such as the Federal Supply Schedule, governmentwide acquisition contracts (GWACs) and OASIS+.
When does the GSA AI clause take effect?
The clause takes effect October 19, 2026. The RGO memo gives contracting officers discretion to put GSAR changes in place before the effective date, and discretion over whether to apply them to open solicitations and awarded contracts.
A deviation stays in force until GSA rescinds it or writes it into the GSAR, and the memo says formal rulemaking will follow with a full notice and comment period. For schedule holders, Winvale reports that GSA listed the clause in the advance notice for MAS Solicitation Refresh #34, expected in November 2026.
Which AI purchases does it cover?
The clause covers purchases where LLM functionality is a material feature and government data is submitted to or produced by the model. Generative assistants, chatbots, agentic systems and LLM-enabled document and productivity tools all fall in that description.
Unless the contracting officer identifies otherwise, two carve-outs apply:
- Internal use. LLMs a contractor uses only for its own back-office, operational or performance-support work, not delivered to or accessed by the government, even if they incidentally touch government data.
- Incidental AI. LLM features that are incidental or ancillary to a product whose primary purpose is not AI.
Crowell & Moring's analysis, published on Mondaq, notes the clause doesn't say whether both carve-outs must apply or whether either one is enough. That is a question for the contracting officer, not something to settle on your own.
What does the clause require?
The main obligations, with the paragraph of the clause where each appears:
| Area | What the clause requires (paragraph) |
|---|---|
| Government data | No training, fine-tuning or improving an LLM with it, no using it to inform advertising or business decisions, no selling or licensing it (e)(3) |
| Data handling | Encrypted transmission and audit logging among the required safeguards (e)(4) |
| Disclosure | Disclose the LLMs used, within 120 days of starting work if the contract sets no date (f)(1) |
| Origin and control | Maximize use of U.S.-based LLMs and avoid models under adversary-government control (f)(2) |
| Incidents | Notice within 72 hours, with daily updates; FedRAMP or CISA reports can satisfy it in some cases (f)(5) |
| Model changes | 30 calendar days' advance notice before adding or replacing LLMs or providers, a concurrent-access period for successor models, and notice to the contracting officer within 7 calendar days of identifying a change that materially increases bias, weakens guardrails or degrades performance (i) |
| Output quality | Reasonable efforts so the LLM prioritizes accuracy and objectivity and acknowledges uncertainty (j)(1) |
| Government testing | Automated assessments for bias, truthfulness, safety, unsolicited ideological content and other factors (j)(2) |
| Suspension and termination | The government may suspend use at any time; after a for-cause termination for unremedied non-compliance, decommissioning-cost liability is capped at 25% of the affected order's value (j)(3) |
| Closeout | Delete government data, including fine-tuned weights, embeddings, indexes and caches (e)(4) |
| Subcontractors | Flows down to subcontractors doing AI design, development, deployment or operation tasks that handle government data; the prime uses best efforts in due diligence on them (a)(2), (d)(1) |
From our delivery teams: Draw the data flow before anyone asks for it. A one-page diagram showing where government inputs go, which model touches them, where outputs and logs live and who can see them answers most of what this clause will raise.
What changed from the June draft?
GSA received more than 75 public comments over the summer, according to FedScoop, and the final clause changed several provisions that drew pushback. Most of the draft's "unbiased AI principles" language and its ban on embedding partisan or ideological judgments came out. The draft's four role-based companion clauses were replaced by one flowdown tied to AI tasks and government data. The suspension right grew: the June draft allowed suspension only until performance problems were fixed.
The clause takes its definition of an LLM from Executive Order 14319. Our post on the federal unbiased AI executive order covers that background.
What should agencies and AI vendors do now?
For acquisition teams buying AI:
- Ask the contracting officer whether the clause goes into new orders now or after the refresh, and how it applies to open awards.
- Ask vendors for a data flow description: what government data reaches the model, where it is stored and how training is prevented.
- Decide who will use the government's testing and suspension rights, and on what evidence.
For companies selling AI to the government:
- Inventory every offering with LLM functionality, note which ones might be a material feature versus incidental, and confirm the reading with your contracting officer.
- Check logging, encryption, retention and deletion settings against the clause, including embeddings and fine-tuned weights.
- Set up the notice processes now: a 72-hour incident path, a 30-day model change notice and a list of subcontractors that handle government data.
If you hold work on GSA contract vehicles, watch for the clause in upcoming modifications and refreshes, which is how the memo says existing schedule contracts are updated.
The Bottom Line
The clause turns good AI data hygiene into contract terms: know where government data goes, keep it out of training, log what the model does and be ready to show your work. Whether it applies to a given order is the contracting officer's call, so the practical step now is mapping your AI data flows before the question comes up. If your team wants a second set of eyes on that map, talk to our AI team.
FAQ
What is GSAR 552.239-7001?
It is a General Services Administration Acquisition Regulation clause titled "Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems." GSA issued it as a class deviation through its regulation overhaul memo RGO-2026-01, and it takes effect October 19, 2026. It sets rules for how contractors handle government data processed by large language models.
Does the GSA AI clause apply to existing GSA schedule contracts?
GSA's deviation memo gives contracting officers discretion over applying the changes to open solicitations and awarded contracts, and says existing schedule contracts are updated through the usual refresh and mass modification process. Winvale reports the clause is listed for MAS Solicitation Refresh #34, expected in November 2026. Your contracting officer decides how it applies to your contract.
Can contractors train AI models on government data under the clause?
The clause lists training, fine-tuning or otherwise improving a large language model with government data as a prohibited use, along with using it to inform advertising or business decisions and selling or licensing it. Other uses are limited to what the contract allows or the contracting officer authorizes in writing.
Does the clause cover AI tools a contractor uses internally?
Generally not. Unless the contracting officer identifies otherwise, the clause does not apply to large language models a contractor uses only for internal or back-office work that the government does not buy or access, or to AI features that are incidental or ancillary to a product whose primary purpose is not AI.
Did GSA keep the "unbiased AI" requirements?
Mostly not. FedScoop reports the final clause removed most of the June draft's "unbiased AI principles" language. What remains is a duty to use reasonable efforts so the model prioritizes accuracy and objectivity, plus the government's right to run automated assessments for bias, truthfulness, safety, unsolicited ideological content and other factors.
Sources
- GSA, RGO-2026-01 deviation memo and GSAR part 552 text (memo January 15, 2026; clause dated September 2026)
- Federal Register, GSAR Acquisition of Information and Communication Technology, request for comments (June 17, 2026)
- FedScoop, GSA issued an AI acquisition policy for new contracts. Here's what it says
- Nextgov/FCW, GSA memo to set AI-specific acquisition rules
- Winvale, GSA's New AI Clause Is Here
- Crowell & Moring via Mondaq, GSA Issues Final Rule On Large Language Model Procurements
