What should you automate first?

The question comes up in every organisation that has decided to take AI seriously, and it is usually answered badly. Either the most visible process is picked, because it would look good, or the most painful one, because everyone complains about it. Neither is the right test.

Three questions that sort the list

Does it repeat? Something done twice a year is not worth automating no matter how tedious it is. Weekly is a floor. Daily is better. Frequency is what turns a small saving into a real one.

Is there a right answer? Not one correct output word for word, but a result a competent person could look at and say yes or no to. Tasks that are pure judgement, negotiation or relationship do not qualify, and dressing them up as automatable wastes everyone's time.

Are the inputs actually reachable? The information the task needs has to exist somewhere a system can get to it. If half the context lives in someone's memory or in a shared mailbox nobody has structured, that is the project, and it is a different project.

A task that passes all three is a candidate. Most organisations have between three and ten, and are usually surprised by which ones they are.

What to weigh after that

Among the candidates, pick on two axes: how much time it takes today, and what it costs when it goes wrong.

Start where time is meaningful and the cost of error is low. Internal work beats customer-facing work for a first attempt, not because the technology is unreliable but because your own processes for catching errors are not yet built. The first workflow is also how you learn what your review process needs to look like.

Where the line between manual and automatic goes

Whole-task automation is rarer than the pitch suggests. The useful pattern is usually a split: the language model does the gathering, the reading, the cross-checking and the first draft, and a person does the judgement and the sign-off.

That split is worth being explicit about, because it changes the job rather than removing it. Someone who spent three hours assembling a report and twenty minutes thinking about it now spends five minutes checking and forty minutes thinking. That is a better job and a better report, and it is a more honest promise than a headcount saving.

The trap of the big first project

The instinct is to pick something substantial so the effort is justified. It is the wrong instinct. A small, boring, genuinely finished workflow teaches you what your organisation's real obstacles are: the system without an API, the approval nobody documented, the data that turns out to be wrong. Those obstacles apply to everything you do next, and finding them on a small task is much cheaper.

Next step

Listing the candidates is usually a single workshop with the people who do the work. The ranking tends to be obvious once the three questions are on the table.