Contract Negotiation

State of play.

• The Anthropic-Pentagon contract collapse remains the defining precedent for AI vendor governance risk in federal contracting. The Department of War terminated its $200 million partnership after Anthropic refused to remove safety guardrails, designated Anthropic a "supply chain risk," and OpenAI captured the replacement contract — with Anthropic's injunction against the designation still pending .
• Vendor "service improvement" clauses have been reclassified as AI risk vectors. Law firms including Kilpatrick Townsend, Consilium Law, and SiLaw have published redlining guides demanding explicit "no training," "no commingling," and "no retention" clauses in MSAs — with the compliance gap between regulated government contracts and commercial agreements now a recognized exposure .
• Data license scope in AI contexts is headed to a federal court ruling. The Fastcase v. Alexi case — turning on whether "internal research purposes" restrictions cover AI model training and public deployment — will set the template for how courts read use-restriction clauses when a licensee's product evolves .
• Engagement-letter arbitration clauses are being construed against firms when framed around fee disputes. A bankruptcy court held that Reed Smith's broad arbitration language was limited to billing disputes by its placement under a fee-dispute heading — leaving RICO and fraud claims in court .
• For counsel advising AI vendors, enterprise technology buyers, or law firms reviewing engagement terms, the practical baseline is: vendor data-use clauses require active redlining at every renewal cycle; federal AI contracts carry demonstrated mid-term termination risk; and arbitration clauses in engagement letters must explicitly cover misconduct claims or courts will read them narrowly.

Where things stand.

• AI agent transaction governance has no settled legal framework — the AAA's Legal Context Protocol is the first industry attempt to fill it. The Legal Context Protocol specifies contractual terms within a broader stack that includes payment and identity protocols; most agent-to-agent transactions currently lack verifiable terms, identified jurisdiction, or any recourse mechanism .
• Agentic AI is reshaping in-house contract management workflows. The market is shifting from copilot tools toward autonomous agent workflows capable of handling intake, clause extraction, redlining, routing, and status tracking end-to-end — with vendors including Ironclad, Sirion, Thomson Reuters, and LexisNexis positioning autonomous exception-flagging as the next efficiency frontier .
• Federal AI contracting now carries a demonstrated supply-chain-risk designation risk. The Anthropic-Pentagon dispute — termination of a $200 million contract, supply-chain-risk designation, and a presidential order barring federal agency use — sets a template for how the administration can pressure technology vendors on AI governance without new legislation .
• "Internal use" restrictions in data licenses are under direct judicial scrutiny. Fastcase's suit against Alexi seeks destruction of both datasets and AI model weights, treating model weights as infringing derivative works — a theory that, if accepted, would give data licensors powerful leverage to constrain AI applications of licensed material and force explicit AI training carve-outs in future agreements .
• Vendor SaaS and AI contracts default to opt-out consent structures for model training. The industry-wide pattern — "service improvement" language covering secondary data exploitation — creates exposure for companies operating under AI legislation or in regulated jurisdictions; the compliance gap between government contracts and commercial MSAs is now documented .
• Engagement-letter arbitration clauses are narrowly construed when framed around fee disputes. The Reed Smith ruling in the Eletson RICO matter establishes that broad "arising out of or relating to" language is undermined by placement under a fee-dispute heading — courts will find no meeting of the minds to arbitrate statutory and tort claims through provisions designed for billing disputes .
• Deferred-control franchise ownership structures require multi-year legal architecture. The White Sox investment agreement — capital injections in 2025 and 2026, an option to acquire controlling interest beginning in 2029, and a second option after the 2034 season — illustrates how layered capital commitments and deferred control transfers must be structured to align incentives while preserving current ownership through a defined transition window .

Latest developments.

• Vendor "service improvement" clauses reclassified as AI risk vectors; Kilpatrick Townsend, Consilium Law, and SiLaw publish redlining guides demanding explicit "no training," "no commingling," and "no retention" language in MSAs
• Agentic AI contract management moves from copilot tools to autonomous end-to-end workflows; Ironclad, KPMG, Thomson Reuters, LexisNexis, Sirion, and Leah among vendors positioning autonomous exception-flagging as the next efficiency frontier
• Reed Smith loses motion to compel arbitration in Eletson RICO matter; Judge Mastando holds that broad arbitration language placed under a fee-dispute heading does not cover statutory and tort claims — no meeting of the minds to arbitrate misconduct allegations
• White Sox deferred-control investment agreement structured by Covington's David Simon — multi-year capital tranches with options to acquire controlling interest in 2029 and after the 2034 season
• WestJet and CUPE Local 8125 reach tentative wage agreement ending 4,400-flight-attendant strike; dispute centered on compensation for ground duties and pre-departure work under WestJet's credit-hour system
• Bronx Defenders tentative contract agreement closes the summer wave of NYC public defender labor actions; deal includes social workers receiving higher pay than attorneys, student-loan assistance, expanded leave, and 2-3% cost-of-living increases for fiscal 2028-2029

Active questions and open splits.

• "Internal use" restrictions and AI model training — what the contract actually covers. Fastcase v. Alexi will determine whether a licensee's pivot from human-reviewed services to AI-driven platforms constitutes a material breach of use restrictions, and whether model weights trained on licensed data are derivative works subject to destruction — the answer reshapes drafting standards for every data license that predates the generative AI era .
• Vendor data-use clauses — opt-in versus opt-out as the default. The industry-wide pattern of opt-out consent for model training leaves commercial MSAs structurally exposed relative to government contracts; whether courts will treat passive "service improvement" language as sufficient consent for AI training — or require explicit opt-in — is unresolved .
• Engagement-letter arbitration clause scope — fee disputes versus misconduct claims. The Reed Smith ruling establishes that heading and placement can defeat broad "arising out of or relating to" language; whether other courts adopt this contextual reading, and how firms should restructure engagement letters to cover statutory and tort claims, is now an active drafting question .
• Supply-chain-risk designation as a contract termination mechanism. The Anthropic precedent leaves open whether a supply-chain-risk designation survives judicial review, what procedural protections attach before designation, and how AI vendors should draft force majeure, termination-for-convenience, and governance-change clauses in federal contracts to account for this risk .
• Enforceability of AI agent transactions under existing contract law. The Legal Context Protocol is an industry attempt to fill a gap that courts and regulators have not yet addressed — whether LCP-embedded terms will be treated as enforceable agreements, and under which jurisdiction's law, is entirely open .
• Agentic AI contract review — liability for autonomous exception-flagging. As in-house teams shift to autonomous workflows that flag only exceptions for human review, the allocation of responsibility for errors in AI-reviewed agreements — between vendor, in-house counsel, and outside counsel — has no settled standard .
• Aviation labor contracts — all-duty-time compensation as an emerging standard. The WestJet dispute over ground-duty compensation under a credit-hour system signals a structural shift in how cabin crew contracts are negotiated; whether all-duty-time compensation becomes the industry baseline — and how existing contracts must be revised — is unresolved across the sector .

What to watch.

• The D.C. federal court ruling in Fastcase v. Alexi — if the court accepts the model-weights-as-derivative-works theory, every data license with pre-AI "internal use" language becomes a renegotiation or litigation candidate .
• Whether Anthropic's injunction against the supply-chain-risk designation survives — the ruling will define the procedural and substantive limits of this enforcement mechanism for all AI vendors with federal contracts .
• Whether the Reed Smith ruling prompts law firms to restructure engagement-letter arbitration clauses to explicitly cover misconduct, RICO, and fiduciary claims — and whether bar association ethics committees issue guidance on the required scope .
• Whether the vendor data-training-clause redlining guides published by Kilpatrick Townsend, Consilium Law, and SiLaw produce market-standard "no training" language that becomes a baseline expectation in MSA negotiations .
• Whether agentic AI contract management deployments produce documented error rates or liability incidents that force vendors to define audit rights and error-allocation terms — and whether market-standard provisions emerge from the renewal conversations now underway .
• Whether the WestJet all-duty-time compensation settlement terms, once disclosed, prompt other cabin crew unions to demand similar restructuring — and how aviation employers respond in contract language .