AI Agent Liability in 2026: Lessons from Air Canada, Workday, and Nippon Life v. OpenAI

Three Signals on AI Agent Accountability

The following table summarizes the three matters discussed in this analysis and their current standing as legal precedents.

Air Canada: Negligent Misrepresentation Applied to Chatbot Output

In February 2024, British Columbia's Civil Resolution Tribunal decided Moffatt v. Air Canada, one of the first adjudicated rulings addressing liability for an AI chatbot's output. A customer relied on incorrect bereavement-fare information provided by Air Canada's website chatbot. Air Canada argued the chatbot was a separate legal entity responsible for its own statements. The tribunal rejected that argument and held Air Canada responsible for all information on its website, whether it came from a static page or an automated chatbot. It found the airline liable for negligent misrepresentation and ordered payment of damages and tribunal fees.

The monetary judgment was modest, consistent with a small-claims-level dispute, but the reasoning is significant. The tribunal did not need a new AI-specific statute to reach its conclusion; it applied existing negligent misrepresentation doctrine directly to the chatbot's output, treating it as an extension of the company's own representations.

Workday: Vendor Liability for Automated Hiring Decisions

Mobley v. Workday, Inc., filed in the Northern District of California, presents a related but distinct failure pattern. The plaintiff alleges that Workday's AI-powered applicant-screening tools produced disparate-impact discrimination against applicants based on race, age, and disability, in violation of Title VII, the ADEA, and the ADA.

Workday sought dismissal on the theory that it could not be held liable because it is a software vendor, not the employer making hiring decisions. The presiding judge rejected that argument and allowed the claims to proceed, on the theory that Workday could be acting as an agent of the hiring employers that deployed its tools. The litigation remains active and unsettled.

The ruling signals that vendor status alone does not insulate a company from liability for what its automated tools produce. The underlying failure point is the absence of traceable decision criteria and human review checkpoints in the screening workflow, which leaves both the vendor and the deploying employer without a documented basis to show the outcome was not discriminatory.

Nippon Life v. OpenAI: An Unverified Signal, Not a Precedent

Nippon Life v. OpenAI has circulated in discussions of AI liability, but publicly available court records and reporting did not contain sufficient verified detail to describe its allegations, legal theory, or current status. Rather than speculate, this analysis treats the matter as an unconfirmed signal to monitor rather than a settled precedent.

Runtime Governance Mechanisms That Address These Failure Points

Both litigated matters point to a common underlying gap: organizations deployed AI systems that produced outputs or decisions without maintaining a verifiable record of what the system was authorized to do, under what conditions, and on whose approval. Runtime governance closes this gap through four interconnected controls.

• Agent identity: Each AI agent is assigned a persistent, auditable identity so every action can be traced to a specific system operating under a specific authorization scope.
• Permissioning: Agents are granted only the capabilities required for a defined task. Scope limitations are enforced at runtime, not merely documented in design specifications.
• Audit trails: Every output, decision, or action is logged with sufficient fidelity to reconstruct what the agent did, what inputs it received, and which policy governed its behavior at the time.
• Policy enforcement: Organizational policies, including content constraints, escalation rules, and prohibited action categories, are applied dynamically at runtime rather than treated as static training-time guardrails.

In the Air Canada scenario, a policy enforcement layer would have prevented the chatbot from stating a fare policy that did not match the current approved policy document. In the Workday scenario, traceable decision criteria and configured human review checkpoints would provide the documentation needed to defend or audit screening outcomes.

Questions AI Governance Leaders Should Be Able to Answer

Based on the failure patterns in these cases, organizations operating AI agents should be prepared to answer the following questions, ideally with documentary evidence rather than general process descriptions.

• Can you identify, for any given agent output, which policy version governed the agent's behavior at the time that output was produced?
• Is there a log entry that records what the agent said or decided, what inputs it received, and which authorization scope it was operating under?
• If a screening or recommendation decision is challenged, can you produce the criteria the system applied and a record of any human review that occurred?
• Do your vendor contracts allocate liability for AI-driven decisions, and do those allocations reflect the agency theory now being applied by courts?
• Can you demonstrate that your chatbot or agent's outputs are bounded by currently approved policies, not training-time snapshots that may be outdated?

What This Means for AI Agent Governance in 2026

Taken together, the Air Canada and Workday matters show that existing law extends readily to AI-driven outputs and decisions without waiting for AI-specific legislation. Courts and tribunals are evaluating whether the deploying organization, or a vendor acting on its behalf, exercised reasonable control over what its AI systems said or decided.

The common thread is not the model's internal architecture but the absence of a verifiable connection between an agent's action and organizational oversight of that action. Enterprises operating AI agents should expect this pattern to continue: liability will turn on whether an organization can demonstrate control, documentation, and reasonable diligence, regardless of whether an AI system nominally made the decision.