Start with repetition, not novelty
The best first automation is usually not the most impressive task in the business. It is the task that happens often, follows recognizable rules, and creates a clear finished result.
A workflow that takes twenty minutes every day is often a better candidate than a complicated project that happens twice a year.
Use the five-part fit test
Before building an agent, write down the workflow and score it against five questions.
- Frequency: does this happen at least weekly?
- Consistency: are most steps repeated in the same order?
- Inputs: can the agent reliably find the information it needs?
- Verification: can a person quickly tell whether the result is correct?
- Consequence: can high-risk actions stop for approval?
Measure time to first useful result
Do not begin with a vague target such as “use more AI.” Track the current minutes per run, runs per month, correction time, and the number of completed outputs.
The automation is valuable only if saved time remains after review and correction. A fast but unreliable agent simply moves work into a different queue.
Expand only after the first workflow is boring
A dependable workflow should become routine before a second one is added. Stable inputs, predictable exceptions, and a clear approval process are signs that the first agent is ready to expand.