Cascade of Errors: How to Align Business Metrics and Machine Learning Without Anyone Getting Hurt

"Build an AI assistant that answers well" — the most popular and the most useless requirement. In this talk, we'll share how we turned abstract "well" into concrete business metrics on a project for a financial department.

We'll walk through the journey from business questions: "Is it allowed to lie?", "Where is it acceptable to make mistakes?", "What happens if it doesn't understand the query?" — all the way to honest, measurable quality criteria.

And then comes the most interesting part: how we tried to align these business metrics with ML metrics — and why the honest numbers turned out to be much more modest than our ambitions. Spoiler: if OCR gives you 80% and NER gives you 85%, the realistic quality ceiling for the system is around 40%, not 95%.