The model has never been the hard part

The model is the easiest part. It always has been.

The hard part is the undocumented schema that no one maintains, where the person who understood it left years ago and is not answering calls. Columns that mean different things depending on who entered the data. Business logic that lives in someone’s head, except they have moved on.

I build AI systems on enterprise data for a living. The databases I work with have hundreds of tables, misspelt column names that downstream joins depend on, and business rules frozen into query expressions that nobody documented. One system I replaced stored user passwords in plaintext alongside bank account numbers. Another had a column overloaded to mean four different things depending on which department entered the record.

The AI model, whether it is Gemma, Claude, GPT, or anything else, screws in at the end. You can swap it for a different one on Tuesday and nobody notices. The months of orchestration, schema mapping, business logic extraction, and evidence governance that make it safe to point an AI at production data: that is where every dollar of value lives.

Enterprise AI is not the brain. It is the entire body, the nervous system, the skeleton, and the immune system that stops the whole thing catching fire every time someone enters a date in the wrong format. Building the brain is the easy part. The hard part is everything else.

And everything else is what I do.