The usual way an automation project gets scoped: someone names the task everyone complains about, and that becomes the pilot.
The task everyone complains about is often a bad candidate. It is loud because it is annoying, not because it is expensive, and it is frequently annoying precisely because it needs judgement - which is the thing automation is worst at.
Two questions, in order
How often does it run? A weekly task that takes an hour costs about 50 hours a year. A task that runs forty times a day and takes four minutes costs closer to 700. The second one is almost always the better target, and it is usually the one nobody mentions because it has been absorbed into the shape of the job.
How much judgement does a correct outcome need? Not "could a person do it faster" - whether the rules can be written down. If two competent people would produce different output from the same input, you are not automating a process, you are encoding one person's opinion.
Anything that scores high on frequency and low on judgement is worth doing first, regardless of how boring it sounds.
The cases that look automatable and are not
- Work whose input is unstructured and whose output is contractual. Extracting terms is fine. Acting on them without review is not.
- Processes that only work because one person patches them. Automating them freezes the patches in place. Fix the process first, then automate it.
- Anything with no clear failure signal. If nobody notices when it goes wrong, you will not notice when the automation goes wrong either — and it will go wrong faster and at more scale.
Where AI changes the answer
The judgement test used to be binary. It is now a spectrum: a model can handle input variability that would have broken a rules engine, provided a person still owns the outcome.
That shifts a whole class of work into range - triage, drafting, classification, summarising, first-pass extraction. Not because the model is right every time, but because being right most of the time with a review step is a genuine improvement over a person doing all of it from scratch.
The design question becomes where the human sits, not whether they are there.
Start with the boring one
The most valuable first automation is usually unglamorous, runs constantly, and nobody had it on the list. Find it by counting, not by asking.