GSA's AI Clause (GSAR 552.239-7001): Final Rule Issued Sept 28 — Effective Oct 19, 2026
Update (October 12, 2026): One week before the effective date, new detail from Winvale's contractor briefing: Contracting Officers already have discretion to use the clause before October 19 — for new solicitations, and at their discretion for open solicitations and awarded contracts. GSA has also listed the clause in the advance notice for MAS Solicitation Refresh #34, whose draft was posted in early October 2026 and which is expected to be released in November 2026; the Refresh formally wraps the clause into a new GSAR Part 539 ("Creating a new AI-related clause for use in GSA contracts, including Schedules, to protect Government Data and to mitigate other risks presented by LLM systems"), applying across the full MAS solicitation — not just IT contractors.
"GSAR clause 552.239-7001, 'Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems' is a class deviation that becomes effective on October 19, 2026, but Contracting Officers (COs) can theoretically start using it now." — Winvale
How the final clause evolved under industry pushback (the scope test the watch tracks):
- Narrower scope: the original draft applied to contractors' AI use generally, even internal; the final applies only when the government is procuring an LLM, generative assistant, chatbot, agentic system, or LLM-enabled document/productivity tool where government data will be submitted directly to or produced by the LLM.
- "Self-deleting" exclusions: unless the CO says otherwise, the clause imposes no obligation when LLM use is limited to internal back-office/operational systems not delivered to or accessed by the government (even if incidentally processing government data), or when LLM functionality is incidental/ancillary to a commercial product whose primary purpose is not AI.
- Definitions narrowed: Data Inputs and Data Outputs definitions tightened; metadata and logs are now excluded from Data Outputs.
- Flowdowns tied to NIST AI RMF task categories: subcontractor flowdown applies to primes' subs performing AI Design, AI Development, AI Deployment, or Operation and Monitoring tasks; primes remain responsible and must use best efforts to vet AI subs.
- "Unbiased AI Principles" removed: replaced with "reasonable efforts" to design/train/configure the LLM to prioritize accuracy, scientific inquiry, and objectivity — but the government reserves the right to run automated assessments of the LLM for bias, truthfulness, safety, and other factors (evaluation access survives).
- Incident reporting narrowed: trigger limited to incidents affecting an LLM used in contract performance that may affect government data; FedRAMP-authorized systems can satisfy it via existing FedRAMP or CISA reports in certain cases.
- Liability cap: contractor liability for decommissioning costs after a for-cause termination is capped at 25% of the affected task or delivery order.
Status of the watch: the trigger (post-Oct-19 solicitations applying the two-part material-feature + direct-government-data test, or reporting on how the scope test and flow-downs are applied in practice) has not yet fired — the effective date is Oct 19, 2026, and no post-effective-date solicitations have been reported yet. The clause functions as GSA policy via class deviation until codified through formal rulemaking with a public comment period, or rescinded.
Founder takeaway: federal (and federal-adjacent) buyers now have a template for LLM data-safeguarding terms — no training on government data, no advertising/selling use, encryption and logging, 72-hour-style incident notice, evaluation access, deletion. Vendors that build government-ready terms now (retention, training restrictions, incident notice, model-change notice) will be positioned for the November Refresh #34 mass mod and the solicitations that follow; the same clause architecture is likely to propagate into commercial procurement playbooks (see The 2026 AI Procurement Governance Mandate: COSO Controls and California's Vendor Certifications and The 2026 AI Procurement Playbook: Weighted Rubrics, TCO Realities, and Non-Negotiable Contract Clauses).