Automating a confused process doesn't fix it. It just makes it faster at being wrong.

The request nearly always arrives in the same shape. There’s something we do by hand, it eats our time, we’d like to automate it. It sounds like a technical question, and for a while it gets treated as one. You look at how it works today, you try to describe it, you start sketching.

Then it stalls, and it never stalls on the technology.

It stalls at the moment you discover that the thing wasn’t being done one way. It was being done three ways. Three people, three different criteria, all three defensible, none ever stated. For years it worked, because each person decided with their own head at the moment it was needed, and took the hard cases to somebody else.

Automating means picking one of the three. And at that point somebody has to say which.

That’s the step where most projects slow down, and it gets reported as a problem with requirements or budget. It isn’t. It’s that you’re asking an organization to make, in a two-hour meeting, a decision it has avoided for five years by keeping it implicit.

A machine can’t hold anything implicit. That’s its one real limitation, and also its greatest use.

The practical consequence is that automation works like an X-ray. Take any process and try writing it down so that someone who has never worked with you could run it. Every point where you stop, every “it depends” you can’t dissolve, is a decision that was never made. They were always there, covered by one person’s judgement.

If you automate them without dissolving them, they don’t disappear. They multiply, because now that confusion runs a thousand times a day instead of ten, and nobody is watching it any more.

I saw this happen with something tiny. A rule about when a case counted as urgent. Nobody had ever written it, and in practice it worked beautifully: whoever received the request understood. Put inside a system, that same rule started flagging things as urgent that weren’t, and letting through things that were, with a consistency that had never appeared when it was done by hand.

The system hadn’t got it wrong. It had applied the rule we gave it, which was the one we believed we’d been following.

The useful part of the work, almost always, happened before a single line was written. It was the conversation where three people put their three criteria on the table and had to choose one. That conversation improved things even in the part that stayed manual.

So when someone asks me to automate something, the first question isn’t about tools. If you had to explain this rule to a new person, in a way that means they never have to check with you, what would you tell them?

If the answer comes in three lines, the rest is execution.

If it doesn’t, we’ve just found the thing worth doing.