FDE Times: when an AI agent gets it wrong, the approval points and logs FDEs build become the evidence

FDE Times published an analysis on October 23 looking at who carries the blame when an AI agent makes a mistake. It points to Moffatt v. Air Canada, where a British Columbia tribunal rejected the airline's argument that its chatbot was a separate entity, and to Mobley v. Workday, where a California federal court refused to dismiss claims that the AI vendor itself could be liable under an agency theory.

The piece contrasts those rulings with standard SaaS-style contracts, which typically put responsibility for actions on the customer. Citing law firm Bristows, it notes that customers find it hard to prove a vendor breach when an agent errs, and that they should require vendors to provide approval points for high-risk actions. It also flags the EU's new Product Liability Directive, under which missing technical documentation can lead to a product being presumed defective, which turns model, prompt, configuration and update records into evidence.

Why it matters for FDEs: the engineer who connects an agent to a customer's real systems usually sets the thresholds, decides which actions wait for a human, and builds the logs. The article warns against oversight that exists only on paper, escalation thresholds that are never tested in evals, and logs that record "approved" without who, when and which agent version. It suggests treating guardrail work as a CV line and preparing concrete examples of governance and audit-trail work for interviews.