An agent can quote the wrong refund policy or take an action outside its permissions. The company deploying it then has to explain what happened, who was affected and what protection applies.
In her Substack essay, Ines Boutemadja uses the Air Canada chatbot dispute to introduce that responsibility. She argues that buying a policy leaves unanswered questions about how an agent was tested and what happens when it fails.
Her proposed approach combines documented testing, contractual performance commitments and insurance with an agreed scope. Each needs its own evidence and terms; a testing record alone does not establish coverage.
The practical checklist includes permission boundaries, human approval requirements, adversarial test results, production monitoring and an incident log that records fixes. These are records an AI vendor can bring to an enterprise procurement review or underwriting discussion.
Existing E&O and cyber coverage should be checked against the actual agent workflow. Insured services, exclusions and conditions determine whether a particular loss is covered.