We wrote the process down expecting a compliance exercise. What we really got was a diagnosis.

This was repair operations, in regulated wagering. If there is a repeating issue, an examiner doesn't ask you to describe your process. He asks you to show it.

We couldn't show it. There was no process document — only a series of emails that not everyone had access to, or remembered. There was a way things were done, understood well enough by some of the people doing the work, and in all the years the operation had been running, it had never really been documented.

So we documented it from start to finish. Not specifically to improve it — just to have it. Every step, in order, with the person responsible for each one named and with reasonable timeframes for accountability.

That document turned out to be the most useful thing we produced that year, and not because any regulator was satisfied.

Once the process was on paper, you could see it. Steps in an order nobody would have chosen on purpose. Waiting built in where nothing was actually being waited for. Approvals that had made sense during some crisis years earlier and had quietly outlived it. None of this was hidden. It had just never been in one place where a person could look at all of it at once.

Then we redesigned it, and trained all parties on the process. No technology was purchased.

Repair cycle turnaround times dropped 75%.

Here's the part I keep coming back to. An undocumented process isn't only a continuity risk. It's an unexamined one. Nobody is concealing anything — there's simply no artifact for anyone to analyze. Bad steps survive not because someone defended them, but because nobody could see them well enough to argue with them.

Which is what worries me about putting AI on top of one. You have to train the AI, and AI runs on documented knowledge — procedures, context, examples, prompts. Automate a process nobody has written down and you haven't just automated it. You've industrialized something that was never inspected. It goes from invisible to permanent without ever passing through examined.

Last week, I said automation amplifies whatever is actually there. This is the harder version of that: sometimes nobody knows what's there.

Most operations have a process like this one. Not the one that's failing loudly — the one that just quietly takes longer than it should, for reasons nobody has ever laid out end to end.

The first AI project in a lot of businesses isn't an AI project at all. It's writing down what you already do. And then reading it. And then, once it's right, teaching it.