When every software product mentions AI, the hardest question is often where to begin. A useful starting point is a recurring task that has clear inputs, a recognizable good result, and someone who can review the output.
Start with a week of ordinary work
Ask your team to list tasks they repeat: sorting inquiries, copying details between systems, preparing updates, looking for documents, or drafting the same kind of response. Record frequency, approximate time per task, who does it, and where work gets stuck. Keep the first pass simple.
Look for three qualities
- The task repeats. A rare task can be interesting, but a recurring one gives you opportunities to learn and measure.
- The inputs are available. Identify which documents, messages, or system records the work requires and whether they can be used appropriately.
- A person can check the result. Write down what a correct result looks like and how the reviewer will catch mistakes.
Separate the task from the decision
“Prepare a draft reply using our approved policy” is a clearer pilot than “automate customer service.” Drafting has a defined source and a reviewer. Customer service includes judgment, exceptions, promises, and account changes that need separate rules.
Write a one-page pilot brief
- Input: what starts the task and where the information comes from.
- Output: the exact draft, category, summary, or prepared record you want.
- Owner: the person who reviews and approves the work.
- Exceptions: the situations that go straight to a person.
- Measures: preparation time, review time, correction rate, and useful completions.
- Stop rule: the result or error that pauses the pilot.
Run the smallest useful comparison
Use a small, representative sample and compare the current process with the assisted one. Include difficult cases. Count review and rework time, not just how quickly the AI produces a draft. Expand only when the result is useful to the people doing the work.