Most people learning automation start by studying tools and never ship anything. Reverse it. Choose one real, boring, repeated task, build the smallest thing that handles it, and let the tools be whatever that required.
Pick the right first task
A good first candidate has four properties. Miss any of them and your first project becomes the one you abandon.
| Property | Why it matters |
|---|---|
| Repeated | At least weekly, or the payback never arrives. |
| Rule-shaped | You can describe the steps to another person in under two minutes. |
| Low stakes | A wrong output is embarrassing, not expensive. Keep the first one away from money and customers. |
| Text in, text out | Language models are strongest here. Leave images, voice and spreadsheets for the second project. |
Good first projects: summarising a week of support tickets into themes, turning meeting notes into task lists, drafting first-pass replies to routine enquiries, tagging incoming feedback by topic.
Write the steps down before touching a tool
Every workflow is a trigger, some steps, and a destination. Writing them as plain sentences first is what stops you building a complicated thing that solves the wrong problem.
Trigger: a new row lands in the feedback sheet
Step 1: pull the comment text
Step 2: classify it as bug, request or praise
Step 3: if bug, draft a one-line summary
Destination: write the class and summary back to the row
Escalate: if the model is unsure, flag for a humanBuild the smallest version that works
- Do one item by hand, with the prompt you plan to use. Confirm the output is genuinely what you want.
- Run ten items and read every one. This is where you find the cases your prompt did not anticipate.
- Only then connect it to the trigger and let it run on its own.
Have the workflow draft and stage the output rather than send it. Move to fully automatic only once you have read a few weeks of results and stopped finding surprises.
Handle the cases that will break it
Automations fail quietly, which is why they lose trust. Decide in advance what happens when the input is empty, when the model returns something unparseable, and when the source system is down.
- Log every run with its input and output, so you can audit later.
- Give the model an explicit escape: "if you are not confident, reply UNSURE".
- Route anything marked unsure to a person rather than guessing.
- Cap the run rate so a loop cannot burn your budget overnight.
Measure whether it was worth it
Note the time the task took before, and the time it takes now including the checking. If the honest saving is small, that is a useful result: you have learned the shape of the work, and the next candidate will be a better one.
Key takeaways
- Choose a task that is repeated, rule-shaped, low stakes and text-based.
- Write the trigger, steps and destination as sentences before opening any tool.
- Run ten by hand and read every output before automating the trigger.
- Give the model a way to say it is unsure, and route that to a person.
Ready to automate more than one task?
The AI Productivity & Career Accelerator builds the workflows, the judgement and the portfolio together.
Explore the course