Choose a job, not a personality.
Start with a recurring task whose inputs, expected output and responsible owner are known. Classifying incoming requests is a more testable first job than asking an agent to run all customer operations.
Compare three approaches.
First ask whether an existing product already supports the task. If the rules are stable, use conventional automation. Consider an agent when interpreting variable language or choosing among approved tools adds useful value.
Write the acceptance test first.
Collect ordinary inputs and difficult edge cases. Define what a correct result looks like, when the agent must stop and who resolves uncertainty. Measure corrections and escalations as well as handling time.
- Begin read-only or with draft outputs.
- Keep high-impact actions behind approval.
- Set cost limits and record tool actions.
- Assign someone to review behavior after changes.
A small-business example.
An agent could read an inquiry, identify its category and draft a reply from approved product information. A staff member reviews the reply and handles account-specific changes. This is an illustrative design, not a claimed client result.
Plan for the second month.
Knowledge changes, access tokens expire and model behavior can shift. Decide how updates are tested and who responds to failures before launch. If nobody owns the operating process, the pilot is not ready to become critical infrastructure.