AI liability tightens as contracts get tougher

The gist

AI liability just got real: courts, regulators, and contracts are making companies pay up when their bots go rogue.

What to know

Courtroom Crackdowns on AI Errors

Judges and regulators are holding companies fully liable for AI mishaps, rejecting excuses about autonomous systems and triggering investigations and lawsuits across industries.

By September 2026, liability for AI deployments was being clarified not as an abstract governance principle but through enforceable decisions and active investigations across sectors. ZDNET noted that in Moffatt v. Air Canada, the British Columbia Civil Resolution Tribunal held Air Canada fully responsible when its chatbot wrongly promised a refund, rejecting the airline’s attempt to treat the bot as a third-party integration; the same piece said the NTSB began investigating Waymo robotaxis in January for “illegally blowing past a school bus stop sign in Austin, Texas,” while NHTSA opened a related probe that led to a recall of more than 3,000 vehicles.

The pressure intensified through late September and October as lawmakers and litigants sharpened who pays when AI-driven systems cause harm. AeroXplorer reported that “A Louisiana family filed a federal lawsuit against American Airlines in April 2026” alleging an agent falsely claimed a flight was oversold after learning the boy’s mother is deaf, while Forbes said that, in late-September 2026, Treasury Secretary Scott Bessent argued AI labs should not get “a liability exemption,” and that two weeks after his September 15 remarks Senator Josh Hawley said, “If you break it, you pay for it,” before introducing an October 1 bill extending liability to operators and developers whose agents recklessly cause hacking damage.

Sources

Contracts Get AI-Specific Teeth

Enterprise buyers are tearing up old software templates, demanding performance guarantees, insurance, and risk-sharing as AI deployments expose them to new liabilities.

The shift in AI contracting is driven by a harder recognition that enterprises deploying AI systems retain legal responsibility even when a vendor owns the model or the underlying technology. SecurityBrief Australia frames AI risk as a deployer governance issue, citing IBM’s 2025 breach research that 63% of organisations had no AI governance policy in place and that, among those experiencing an AI-related security incident, 97% lacked proper AI access controls, which is why one company said it aligned its AI governance framework with ISO/IEC 42001 so the system is held to a defined, audited standard.

That same logic is pushing contracts beyond generic software terms toward detailed risk allocation, indemnities, data residency clauses, and insurance products, because model provenance does not end the customer’s exposure after signature. Morgan Lewis notes that six months later a provider may upgrade or replace the underlying model, while Startup Fortune says a wave of 2024 and 2025 chatbot deployments overpromised and underdelivered, making guarantee clauses table stakes for anything north of six figures; hence outcome-based pricing like Intercom’s Fin at $0.99 only when it resolves a conversation, and the pressure around Decagon’s median enterprise contract of $432,750 a year, about $36,063 a month, because at that price a CFO wants a mechanism tying spend to performance.

Sources

Procurement Blocks Over Data Chains

AI deals now stall or die in procurement as buyers insist on granular data flow transparency, subprocessor disclosures, and enforceable residency clauses to satisfy regulatory scrutiny.

Procurement has become the choke point for AI deals because buyers now inspect data flows that older SaaS paperwork never covered. Startup Fortune reports that “the clause needs four things, and the order matters because procurement reads top to bottom and stops asking questions once satisfied,” starting with an explicit transfer mechanism, while legal teams also demand where inference runs, where logs are stored, and that a clause that survives legal review must include “Third, every subprocessor in the chain, by name, because GDPR Article 28 makes the vendor liable for its subprocessors' location too.”

Those demands are no longer theoretical: Startup Fortune says startups “keep losing months of a signed deal,” and that drafting the clause before the first EU call can cut a cycle that usually lasts eight to twelve weeks by three to four weeks. In one case, a Series A company selling an AI agent to a German insurer spent eight weeks in procurement, won a verbal yes, then saw legal review kill the deal over four sentences in the data processing addendum, as Schrems II-era scrutiny after “the 2020 European Court of Justice ruling that struck down the EU-US Privacy Shield” put “every cross-border data transfer under real scrutiny,” making vague residency promises unacceptable.

Sources

Insurance Lags, Indemnity Rises

With traditional insurance failing to cover AI risks, contracts are filling the gap with indemnity clauses and capped liabilities to ensure someone pays when AI goes wrong.

The market is reaching for insurance and indemnity because AI failures sit awkwardly inside older coverage and liability models. Latent Space noted that AI creates “a lot of uncertainty,” making insurance “hard to do early on” even as that same ambiguity increases demand for risk transfer, while Weighty Thoughts described a policy debate in which frontier-lab proposals said “not liability, not damages, not insurance…” even as Anthropic’s August 2024 letter to Governor Newsom floated “strengthening liability for catastrophes” and “regulation of insurance” as mechanisms to manage harms existing frameworks do not cleanly absorb.

That gap is being filled in contracts first, with indemnities and tailored caps standing in for clean doctrinal answers. Startup Fortune reports that “Indemnification caps on AI agent contracts typically run 1x to 3x the annual contract value, not uncapped liability,” while enterprise buyers often “use the DPA to set a separate breach liability cap, commonly 3x to 5x fees paid,” because skipping that work can create “uncapped exposure”; Microsoft’s September 2023 Copyright Commitment became a model by promising to defend Copilot customers against output IP claims, and the Air Canada chatbot ruling showed why buyers want someone financially on the hook when an agent “invented a refund policy.”

Sources

Europe and US Draw Liability Lines

New EU directives and US legal moves are erasing ambiguity, making it clear that AI developers and operators face direct accountability—regardless of an agent's autonomy.

The legal direction in late 2026 is toward clearer, harder-edged AI accountability, not ambiguity. In the EU, that tightening is visible both at the bloc level and inside member states: the Revised Product Liability Directive extends strict product liability to software and AI-enabled products with a December 2026 transposition deadline, while Italy has gone further by building on Law No. 132/2025 to create “an additional domestic regulatory architecture alongside the EU rules,” including civil and criminal exposure that makes liability more explicit and more enforceable in practice.

Outside the EU, regulators and courts are also closing off the idea that AI autonomy weakens legal attribution. Late-2026 discussions increasingly hold that existing liability and consumer-protection frameworks are inadequate for autonomous systems, and on September 25, 2026, FTC Chairman Andrew Ferguson said at Reuters Momentum AI that “developers who instruct AI agents bear liability for any resulting harm”; California AB 316 likewise forecloses any defense that an AI system “autonomously” caused harm, while Amazon v. Perplexity on August 4, 2026 left tort and contract theories against AI developers available even after narrowing other claims.

Sources

Part of these trends

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.