GSA’s final LLM acquisition clause, GSAR 552.239-7001, takes effect Oct. 19 as a class deviation signed by Senior Procurement Executive Jeffrey Koses, giving federal AI vendors four days to align their contracts, data flows, and incident processes with a regime that didn’t exist last quarter.

The clause’s scope is narrower than its reputation. Per Crowell & Moring, it attaches only when LLM functionality is a “material feature” and government data is “submitted directly to or produced by the LLM.” Back-office use, or anything “incidental or ancillary,” falls out. That two-part test is the first thing any small vendor on the Multiple Award Schedule should document, because the self-deletion is the product’s cheapest compliance posture when it applies.

Inside scope, the obligations are heavy. Contractors can’t train LLMs on government data, use it to inform advertising, or sell it onward, per Nextgov. Incident reporting runs on a 72-hour clock, though Crowell notes FedRAMP and CISA filings can satisfy it concurrently. Kevin Martin of Government Acquisitions Inc., quoted in Nextgov, flags a 120-day disclosure deadline, embedding and fine-tuned-weight deletion at closeout, and 30 days’ notice before any major model swap. For-cause decommissioning liability is capped at 25% of the affected task or delivery order, per FedScoop.

And then the sentence that reorders every vendor roadmap: “The Government retains the right to suspend use of the LLM at any time.”

GSA has signaled formal rulemaking with a public comment period will follow, and Crowell notes the companion flow-down clauses for the four NIST-aligned roles remain unissued. The clause lands into the same unsettled terrain shaped by Cantwell and Rounds pressing the White House for mandatory AI standards and Trump’s AI liability framework. Procurement, not legislation, is where federal AI policy is actually being written.

Sources

Sources