Why 60% of an AI project is knowledge extraction

60% of every AI agent project I run is spent on knowledge extraction. Not model tuning. Not technical integration.

It's the tacit knowledge problem. The expertise that lives in someone's head, in a shared drive nobody reads, on sticky notes stuck to a screen.

Companies come in with a simple idea: 'We want a chatbot that answers our customers' questions.' Great. But where are the answers? Scattered across 400 pages of documentation in 3 different systems. In the memory of Marie in accounting who's been there for 15 years.

The real work in an AI project is transforming this implicit knowledge into something structured, searchable, and reliable. Before even talking about models or embeddings.

My approach: I always start with a knowledge map. Who knows what? Where is it documented? What only exists in people's heads? This diagnostic takes 2-3 weeks. It saves 3 months of wasted development.

Projects that skip this step invariably end up building a brilliant AI agent... that gives wrong answers because it was fed incomplete data.