There's an idea going around in AI right now: get inside a business with one product, instrument everything, and you'll learn how the company actually works.

It's a reasonable idea. It misses the part that matters.

The gap between how a company says it works and how it actually works lives in the exceptions. The call someone makes before an order ships. The spreadsheet that reconciles two systems nobody else reconciles. The workaround that has been holding things up for three years and appears in no system of record.

Exceptions are exactly what doesn't leave a trace. Instrument everything and you've captured the sanctioned version of the operation, which is the one they already told you about.

In August, I wrote about a repair process in regulated wagering that had never been written down, and how documenting it turned out to be the diagnosis.

What I didn't say then is why it had never been written down. Nobody was hiding anything. Nobody was careless. Whole pieces of the operation lived in one or two people's heads and in email threads, because documenting them had never been anyone's job.

That's what the instrumentation idea misses. The exceptions aren't undocumented because they're hard to see. They're undocumented because nobody owns writing them down. It's an ownership problem, not a knowledge problem, and no amount of observation fixes an ownership problem.

Writing them down did more than speed the work up. The work stopped depending on those one or two people being available, or on someone finding a three-year-old email.

And it's the precondition for everything after. You can't reproduce it if it's not documented. You can't scale it if it's not reproducible. And you can't train a model on it either.

So before anyone offers to learn your operation by watching it, ask one question: who in your company owns writing down the exceptions?

If the answer is nobody, no tool is going to find them for you. Someone has to be given the job.