Most teams pick the wrong first automation. They start with the process everyone complains about, because the noise feels like a signal. It is not. The loudest process is rarely the one that, once fixed, actually moves the numbers.
After doing this across a bunch of companies, I use a simple four-step funnel: surface the processes, filter them with a few signals, find the real constraint, then score the finalists. Here it is.
1. Surface the processes (the good ones are usually invisible)
The best candidates are often the ones nobody mentions, because people do them semi-consciously: the copy-paste between two tools, the manual re-formatting, the "I just check this every morning." Annoyance is loud. These are quiet. So do not brainstorm "what annoys us." Shadow a normal day instead and write down every repeated, rules-based action, especially the ones people do without thinking. You are looking for recurring, not painful.
2. Filter with the signals
A process is a strong candidate when it is:
- Recurring: it happens daily or weekly, not once a quarter.
- Rules-based: a human could write down how to decide, even if they never have.
- High volume or high stakes: it eats real hours, or a mistake is expensive.
- Structured in, structured out: data in, data out, not "use your judgment."
- Sitting on a bottleneck: more on that next.
If it fails most of these, it is a bad first automation, no matter how loud it is.
3. Find the constraint
Between two equally good candidates, automate the one on your bottleneck. Output is governed by the slowest step, so automating anything else just makes work pile up faster in front of the constraint. Ask where work waits. Where do things queue? That waiting is your map. Fixing a non-constraint feels productive and changes nothing.
4. Score the finalists
For the last two or three candidates, score each on impact (hours saved or risk removed), frequency, and feasibility (how clean the rules and data are). Multiply, do not average. A high-impact process with messy, undocumented rules is a trap, not a win. The winner is usually the boring, recurring, high-frequency one that sits on the constraint and has clean inputs.
The trap to avoid
Starting with the visible, emotional process, the one in the all-hands complaint. It is visible precisely because it is irregular and messy, which is exactly what makes it a bad first automation. Start where it is quiet and repetitive. The wins that compound are unglamorous.
One thing that took me a while to learn: the hard part is rarely the automation itself, it is keeping the context. The process changes, the rules drift, and six months later nobody remembers why it was built that way. That is a separate problem, memory and documentation, and I have been open-sourcing how I handle it, but that is another post.
If you are picking your first one: ignore the noise, follow the queue, and automate the quiet repetitive thing on your bottleneck.
I write about AI agents and automation in production at Yempik.com. What was your first automation, the one that screamed loudest or the quiet one?
