How to Prevent AI Support Liability
To prevent legal liability from AI customer support, operate on one principle: you own every word your bot says. In the Air Canada case (Moffatt v. Air Canada), a tribunal held the company liable when its chatbot invented a bereavement refund policy. The defenses are concrete: ground answers in real policy, compute commitments in code, disclose the AI, keep risky actions on a human-only list, and monitor reopen rate to catch wrong answers fast.
The precedent every team should know
The Air Canada tribunal rejected the argument that the chatbot was a separate entity the company was not responsible for. A bot answer is a company answer, legally. That single ruling means any invented policy, price, or promise your AI makes can bind you, which reframes hallucination from an embarrassment into a liability exposure.
The concrete defenses
- Ground or refuse: never state a policy, price, or refund the source material does not support.
- Policy in code: compute refund amounts and eligibility windows deterministically, not from the model.
- Forbidden-action list: route refunds over a threshold, legal, and account changes to a human.
- Disclose the AI: label it clearly so expectations and accountability are set.
- Monitor reopens: a rising reopen rate catches wrong answers before they compound.
How Resolve reduces the exposure
Resolve is built so hallucination is structurally unlikely: grounded answers, business logic in code, a supervisor check before send, a hard forbidden-action list, and a decision card for anything uncertain. Reva refuses to guess when the knowledge base lacks an answer, and the AI is always labeled as AI. The design turns "you own what the bot says" from a risk into a manageable, auditable process.