A task is worth automating when it happens often, follows rules that can be written down, currently involves a person copying between systems, and has an agreed definition. Fail any of the four and the automation will cost more than it saves. The fourth is the one that catches people out.
Does it happen often enough?
Multiply the time it takes by how often it happens. A ten-minute task done daily is forty hours a year; the same task done monthly is two. The first is worth automating and the second is worth leaving alone, and the difference is entirely in the frequency rather than the tedium.
Can the rules be written down?
Not "can a person do it consistently" but "can you write the rule on paper such that somebody new could follow it". If the answer involves "it depends" more than twice, you are looking at judgement, and judgement automates into confident wrong answers.
Is a person currently moving data between screens?
This is the strongest single signal. Reading an order from an email and typing it into a system, copying a figure from a spreadsheet into a form, checking one list against another: that work is pure transfer, it is error-prone precisely because it is boring, and it is what automation is genuinely good at.
Does everyone agree what the task is?
The test people skip, and the one that decides whether the project works. Ask three people who do the task to describe step four separately. If they disagree, automating it will pick one of their answers and make that disagreement permanent and invisible.
This is why we map a process before automating it, and it is why the map is often worth more than the automation. Quite often the mapping alone removes two steps that existed because somebody left in 2021 and nobody revisited why.
The two that are never worth it
Anything that runs once. And anything where the person doing it is also the person checking it, because automating that removes the check without removing the risk.
AI Agents Automation n8n