What to automate first, and what to leave alone

Automation projects fail in two predictable ways. Either they automate a process nobody had agreed on, which makes the disagreement permanent and harder to see. Or they automate a decision that genuinely needed a person, and the business slowly stops noticing the cases where the rule is wrong.

Both are avoidable with one question asked honestly before anything is built.

The question

Does this step follow a rule you can write down completely?

Not mostly. Completely, including what happens in the cases you would rather not think about. If you can write the rule and the exceptions, the step is a candidate. If writing it down produces "it depends on the customer" or "you get a feel for it", that step is judgement wearing the costume of a process, and automating it will remove the judgement without removing the need for it.

What is worth automating first

Moving information between places. The highest-value and lowest-risk automation is almost always the one that stops a human retyping something that already exists. It follows a perfect rule, it is pure cost, and nobody enjoys it.

Things that must happen on time regardless of who is busy. Acknowledgements, reminders, scheduled reports, renewals. These fail because people are busy, not because people are careless — which makes them exactly the right shape for a machine.

Checks that should never depend on mood. Did this order get its deposit. Is this document missing a field. Has this job been open for more than a week with no update. A person doing these consistently is a person being wasted; a person doing them inconsistently is a risk.

Making things visible. Not a decision at all, but usually the highest return. A view that shows what is stuck, updated automatically, changes behaviour more than most automation does — because you cannot fix what you cannot see.

What to leave alone

Anything where the exception is the value. If your customers stay because you handle their awkward cases well, do not automate the awkward cases. Automate the paperwork around them so there is more time for the case itself.

Pricing that depends on context. Rules-based pricing works where the rules are real. Where a quote depends on how difficult this client is to work with, or how much you want this project, a formula will be confidently wrong.

The first human response. Acknowledgement can be automated honestly. The actual reply should be from someone who read the message. People can tell, and it costs more trust than it saves time.

Anything you have not done manually enough times to understand. If a process is three months old, automating it freezes a version you have not finished learning yet.

Automate the step that follows a rule. Keep the step that needs a judgement. The mistake is not knowing which is which.

Where AI fits, honestly

AI has genuinely widened what can be automated. It handles unstructured input — a message written in a person's own words, a document in an unpredictable format — in a way that rule-based automation could not.

It has also made it easy to automate things that should not be, because the demo is always convincing. The useful test is the same one as before: can you describe what correct looks like, and would you notice if it were wrong? Summarising an enquiry so a person can triage it faster passes that test — a human reads the output immediately. Deciding which enquiries are worth answering does not, because nobody sees the ones it discarded.

Use it where a person checks the output as part of the normal flow. Be careful where it acts and nobody looks.

A practical order

Write down the process as it actually happens, including the parts people work around. Mark each step as rule or judgement, honestly. Take the rule-following steps and sort them by how often they happen and how much they cost when missed. Automate the top one. Then look again, because removing one step usually changes the shape of the rest.

One step at a time sounds slow and is not. The alternative — automating the whole process at once — means you cannot tell which part delivered the value and which part quietly broke something.

The measure that matters

After a month, ask whether anyone would notice if the automation stopped. If it would be reported within a day, it was worth building. If nobody would notice for a week, it was automating something that did not need doing at all — which is a useful thing to learn, and cheaper to learn small.