Answering questions over a company's own documents with AI works in a demo almost every time, and in production considerably less often. This book is the inventory of why, written by people who watched it fall over.
What I did with it
I corrected my own page on AI over a company’s own data in twelve places. The most important one: it said to start with simple search and add complexity later. The book shows the opposite, and that searching by keyword and by meaning at the same time from day one saves redoing the work.
I adopted a measurable threshold for faithfulness — what share of an answer is actually supported by the documents — with a number below which the system does not go live. I used to talk about quality without a number, which amounts to saying nothing.
And I added what few people mention before invoicing: tables and scanned documents are the real problem, not the model. A crookedly scanned contract is not fixed by a larger model.
What I left out
The parts tied to the authors' own platform. It is understandable that they write from their product; the decisions I carried over are the ones that hold with any tool.