A month earlier, a DPD delivery chatbot in the UK had sworn at a customer and written a poem criticizing its own employer, after a user jailbroke it. Around the same time, a Chevrolet dealer's ChatGPT-powered bot agreed to sell a $76,000 Tahoe for one dollar. None of these were model failures in the research sense. They were governance failures - systems with no guardrails, no grounding, and nobody reading the transcripts.

Every AI support tool still ships with a demo of its best answers. Nobody demos the worst ones. But customers don't meet your average answer. They meet the one answer they got, and that answer is your brand now - and, as Air Canada learned, potentially your legal liability.

The fix isn't a bigger model. It's a reading habit: twenty random transcripts a week, scored on three questions. Was it correct - grounded in your actual policies, not improvised? Was it complete - did it resolve, or merely respond? Did it know when to hand over to a human, or did it bluff past its competence? Teams that run this audit find the same patterns every time: the bot invents policies instead of quoting them, dodges pricing questions, and escalates three messages too late.

This is where AI chat done right earns its keep. A chatbot grounded in your own knowledge base answers from what you actually wrote - your return windows, your prices, your terms - not from what the model guessed. Confidence thresholds route uncertainty to a human before it becomes a tribunal exhibit. And every conversation stays readable in transcripts you can actually audit, which is the entire point: the companies that survive the chatbot-liability era won't be the ones with the smartest models. They'll be the ones that read what their bots said.

Your chatbot's worst answer is already out there, talking to a customer while you read this. Go read it before a tribunal does.

Oscar Chat was built for teams that can't afford a wrong answer - ecommerce brands, Shopify and WooCommerce stores, clinics and service businesses answering customers in 50+ languages. The bot grounds every answer in your own knowledge base, hands off to a human the moment confidence drops, and keeps every transcript readable for your weekly audit. There's a free plan to start, no credit card required: https://www.oscarchat.ai