> "We deployed an AI assistant on top of our docs and it hallucinates — or worse, it confidently returns the wrong answer." ^pain A collective consumption pain. The org's artefacts may be well-written, well-organised, and well-maintained for human reading — and still be unusable by an agent because the structure is implicit, the format is presentational, or the tacit context every human reader brings is missing. The cost is invisible until the agent returns a wrong answer; then it cascades through every downstream consumer of that answer and, eventually, through the org's confidence in agent-assisted work at all. ## Discovery questions - "Have you deployed any AI assistant on top of your internal documentation? What's the failure mode you see most often?" - "When the AI returns a wrong answer about your business, how does anyone catch it?" - "What would have to be true about your knowledge base for an agent to use it without hallucinating?" ^discovery-questions ## Examples - Air Canada chatbot case: Quebec tribunal held Air Canada liable after its website chatbot gave misleading information (Moffatt v Air Canada).[^1] - Mata v Avianca: U.S. federal judge sanctioned lawyers after ChatGPT invented non-existent case citations filed in court.[^2] - Michael Cohen-affiliated lawyer case: ChatGPT fabricated citations, prompting judicial warning about AI hallucinations.[^3] - Coveo case studies: customer support and e-commerce chatbots that previously hallucinated or surfaced wrong answers due to fragmented knowledge.[^4] [^1]: https://www.cbc.ca/news/business/air-canada-chatbot-court-case-1.7113932 [^2]: https://www.nytimes.com/2023/06/22/nyregion/chatgpt-lawyer-sanctions.html [^3]: https://www.reuters.com/legal/government/another-lawyer-sanctioned-using-chatgpt-without-checking-work-2023-12-14 [^4]: https://www.coveo.com/en/customers