How to turn a messy process into an automation

Most processes that look like automation candidates are not processes. They are habits with exceptions.

The automation fails not because the tooling is wrong but because nobody wrote down what actually happens — including the cases that go sideways, which nobody counts because handling them is just called Tuesday. What gets built is the version in the process document, which nobody has run in years.

The fix is not more analysis. It is three hours in front of a wall, before anybody opens a tool.

Draw it before you build it

Get the people who do the work in a room with a wall.

One card per step, on physical cards or a shared board. The rule is that a card describes something a person or system does, not a stage it is “in”. “Awaiting approval” is not a card. “The manager approves it” is. Status names hide the actor, and the actor is the whole question.

Two things happen every single time:

• The steps on the wall do not match the documented process
• Two people who do the same job lay out different cards

That disagreement is the finding. You cannot automate a process two practitioners describe differently — you would just be picking one arbitrarily, and the other person would spend six months correcting a system built to do their job wrong.

Do not settle it by consensus. Usually neither version is right: one describes the process as designed, the other as adapted. Follow ten real cases through the logs and count how many matched the wall exactly. Fewer than seven, and the wall is a fiction you were about to encode.

Find the exceptions before they find you

Now walk the wall and ask one question at every card: what makes this step go wrong?

You are looking for the branches nobody mentions in the happy path:

• The customer replies with something that does not fit the form
• The upstream system is down and somebody keys it manually
• A manager overrides the rule for a particular account
• The record already exists, and now there are two

The happy path is maybe eighty percent of volume and twenty percent of the work. Automate it alone and you will have built something that handles most of the cases and none of the effort.

Ask at every card, including the ones where people say that step never fails. That is where the undocumented workaround lives, absorbed so smoothly nobody experiences it as an exception. The follow-up that works: when did you last have to do something unusual here? Memory of an incident is reliable. Estimates of frequency are not.

That is the opening. The rest of the lesson — the remaining sections, the exercise and the check-your-understanding questions — is behind the gate below.