Most people's first automation is the one they find interesting: an AI agent that does something clever, a scraper for a site they like. It is a reasonable way to learn and a poor way to get value, because interesting jobs are rare and complicated jobs break.
The job worth automating first is the one you are quietly annoyed by every week.
Four questions
Score any candidate job on these:
- How often? Daily beats weekly beats monthly. Frequency is what turns a small saving into a real one, and it is also how you find out quickly whether the thing works.
- How mechanical? If doing it requires judgement, automation will encode your judgement badly. Copying fields between two systems is perfect. Deciding which customer gets a discount is not.
- How often do you get it wrong? Tasks people fumble — transcription errors, missed steps, forgotten follow-ups — pay back twice: the time saved and the mistakes avoided.
- What happens if it misfires? A wrong row in a spreadsheet is recoverable. A wrong email to two thousand customers is not. Start where the blast radius is small.
The sweet spot is high frequency, fully mechanical, error-prone, and low consequence when wrong. That is not a glamorous quadrant, which is exactly why it is unoccupied.
Three that are almost always worth it
1. Getting data out of email and into a system
Invoices, order confirmations, form notifications. Somebody is opening these and retyping the numbers. A parser plus a validation step removes a daily chore, and this job is the archetype: the same shape appears in a hundred businesses.
2. The report somebody builds by hand
Find the recurring spreadsheet that gets assembled every Monday from three sources. It is mechanical, it is frequent, and the person doing it will be your loudest advocate for the next one.
3. The follow-up nobody remembers
Quote sent and never chased. Trial started and never checked on. These jobs are pure loss when forgotten, and a scheduled check with a clear rule is often a couple of hours of work.
Then do the unglamorous part
Once it works, spend the extra hour: add the retry, add the error branch that messages you, add the check that notices if it stops running. First automations become load-bearing faster than anyone expects, and the ones that quietly stop are worse than the ones that never existed.
How to know you picked right
Two weeks in, nobody mentions it. Nobody thanks you for it either — the task simply stopped being a task. That is what a good first automation looks like.
If you would rather be shown than told, the courses here start from exactly these jobs, with the workflow files included so you can import and adapt rather than build from a blank canvas.