AI agents aren't confidently wrong because of bad context — they're wrong because of bad data engineering

You spend weeks tuning an AI chatbot. Answers are accurate. Stakeholders sign off, and you ship it. Three months later, the system is confidently wrong about a third of what users ask.