Air Canada's chatbot invented a refund policy that didn't exist. A tribunal made the airline pay. Before you ship your AI agent, find out what it promises when nobody is watching.
Moffatt v. Air Canada (2024 BCCRT 149) · the $1 Tahoe bot · DPD · NYC MyCity
These aren't hypotheticals. Every scenario in our test suite is reconstructed from a documented, public failure.
A bot offered a bereavement fare it had no authority to offer. The company was held liable — disclaimers did not save them. Does yours invent commitments too?
A dealership bot agreed to sell a $76,000 vehicle for $1. After it went viral, it absorbed 3,000+ copycat jailbreak attempts from strangers trying the same trick.
A customer says "I dispute this transaction." Under Regulation E that can start a formal error-resolution clock. Your bot answers politely. Nobody in disputes ever hears about it.
When a customer says something legally operative — "I dispute this charge", "I want my data deleted", "I revoke authorization" — does your system recognise it, preserve it, and route it to a human?
Regulation E — error resolution Regulation Z — 60-day billing notice GDPR Art. 15 / 17 CCPA opt-out
Miss the routing and you have a problem that no dashboard will ever show you. We test for exactly this.
Not a dashboard. Not a subscription. A document — with transcripts — that tells you where your bot crosses the line before a customer finds out.
False promises, refund & policy invention, legal routing, escalation dead-ends, brand self-harm, price & commitment errors.
Full conversation logs, exact trigger phrases, and the step where your bot crossed the line — reproducible by your team.
Each finding mapped to the precedent or regulation it violates, so legal and CX can triage in one meeting.
Concrete guardrail and escalation-path recommendations, written for the team that owns the bot — not a generic checklist.
No subscription, no integration, no access to your customer data.
Single bot, 30 scenarios, delivered within 5 business days.
Up to 3 bots or locales, same scenario suite, consolidated report.
We re-run the same suite after your fixes and verify each finding is closed.
Not a subscription. You pay once, you get a report. If we find nothing, the report says so — and that's worth knowing too.
No. We test with synthetic users against your public chat widget or a sandbox account you create for us. We never read real customer conversations, never connect to your backend, and never train on your data.
Any text-based support bot reachable through a web chat widget or API — Intercom, Zendesk, Drift, custom-built. We don't test voice agents (that's a different discipline).
They sell subscription platforms for teams that build and operate voice agents at scale. We sell a one-time report for teams that just shipped a text bot and want to know their legal exposure before launch. Different buyer, different moment, different deliverable.
Then you get a documented, scenario-by-scenario pass record — the same artifact your legal team would want after an incident, except you have it before one.
Built by an engineer who has spent years on both sides of this: shipping AI agents and reading the incident reports after they go wrong. The scenario library is maintained as new cases and rulings emerge.