The most expensive sentence in automation is "wouldn't it be cool if." Plenty of manual work should stay manual, and plenty of automation projects die because they started from enthusiasm instead of arithmetic. Before we agree to build anything, for a client or for ourselves, we run the workflow through four questions.
1. Does it actually repeat?
Not "does it feel constant," but does the same shape of task happen weekly or daily? People overestimate how repetitive their annoying work is. A task that occurs twice a quarter with different inputs each time is a bad automation target no matter how much everyone hates it. The best candidates are boring on purpose: same trigger, same steps, same destination, over and over.
2. Are the inputs and outputs concrete?
"An invoice arrives by email and needs to end up in the accounting system" is automatable, because you can point at the input and the output. "We need to stay on top of vendors" is not a workflow, it's a mood. If you can't describe where the work starts and what done looks like, no amount of AI fixes that. The scoping conversation has to happen first, with people, not software.
3. Where exactly is the judgment?
Most workflows are ninety percent mechanical with a thin layer of judgment somewhere in the middle. That's fine. The trick is knowing precisely where the judgment lives, so the system can handle the mechanical part and route the judgment call to a person. Workflows where every step needs discretion are a poor fit. Workflows where people think every step needs discretion usually just haven't been written down yet.
4. What does an error cost?
An automation that mislabels an internal document is a shrug. One that pays the wrong vendor is an incident. Neither answer disqualifies a project, but the error cost decides the design: how much review to build in, where the system should stop and ask, how loud failures need to be. Teams that skip this question end up either paranoid about a harmless system or casual about a dangerous one.
Then start with the boring one
Teams that pass this filter usually have several candidates, and the temptation is to start with the most impressive. Start with the most boring one instead. The unglamorous workflow that scores well on all four questions will quietly pay for itself, build trust in the approach, and teach you how automation behaves inside your company. The impressive project goes better as the second project.