What to automate first, and what to leave alone

When people decide what to hand over to AI, they usually start with the task that annoys them most. That is a good instinct, but the most annoying task is often the hardest one to automate. Before you pick, it helps to ask what kind of work it is.

Four kinds of work

I use a framework called Cynefin for this. Dave Snowden developed it to help people match how they act to the situation they are in. It describes four kinds of situations. The Cynefin Company: about the framework

There is also an area in the middle for confusion: the times when you don’t yet know which kind of situation you are in. Clear and complicated work are both ordered: cause and effect can be known. Complex and chaotic situations don’t, at least not in advance. Work does not move neatly from one domain to the next, and not everything can or should become clear. Cynefin wiki: the domains

One job, several kinds of work

Most jobs are a mix. Take a customer request that arrives by email. Logging it, attaching the original message, and checking that the required details are there is clear work: you can write the rules down. Deciding whether the request is technically possible is complicated: it needs someone who knows the system. Finding out whether the change will actually help customers is complex: you learn it by trying.

An agent can take over the first part completely, prepare the second, and support the third. If you treat the whole job as one task, you either automate too little or hand the agent decisions it should not make.

Busy is not the same as chaotic

A company that is always firefighting feels chaotic. In the Cynefin sense it often isn’t. There are patterns; nobody has had the time to look for them. Snowden makes this point about the everyday use of the word. “The chaos word”

So look at recent cases before you label anything. Where did each request arrive? What was missing? Who decided? Where did it wait, and why did it come back? The people who do the work know the exceptions, so ask them. Their workarounds often show what the official process missed.

What to automate first

Automate the clear parts first. Let the agent prepare the complicated parts for a person to decide. Keep people in charge of the complex parts, and use AI there to explore, not to decide. Give the agent more room only when its results show it can handle it.

Anthropic’s engineering guidance makes a similar distinction. Fixed workflows suit well-defined tasks, agents that choose their own steps suit open-ended ones, and you add complexity only when it improves the result. Anthropic: Building effective agents

Measure, but don’t over-read it

Track something that matters for the job: how long requests wait for missing information, how often work comes back, and how much time people spend checking it. Check that the result still meets the need. Faster intake is no help if it creates rework later.

A quieter week is welcome, but on its own it doesn’t prove that your change caused it. Look at several weeks, at the spread, and at the exceptions. In complex work, numbers help you notice patterns and test ideas. They won’t give you a precise forecast, and they don’t have to.

Try it on one job

Pick one job that keeps coming back and answer five questions about a few recent examples:

1. Start and finish. What starts the job, and what should be true when it’s done?
2. What repeats. What stays the same across recent examples, and what changes?
3. What kind of work. Which parts follow a rule, which need an expert, and which need an experiment?
4. The agent’s part. What could an agent do on its own, what should it prepare, and who handles the exceptions?
5. The check. How will you know it worked, and how will you notice if it causes harm?

The answer might be an automation. It might also be a clearer handoff or a better question at the start. Both make the next round of work easier.

If you would like a second pair of eyes on a job you are thinking of automating, write to me.