Most AI agent projects should not start with a general-purpose chatbot. They should start with a boring workflow that already happens every week and already has a human owner.
Start where the rules are visible
Good early agent work has clear inputs, a predictable decision path and an obvious handoff. Intake triage, quote preparation, customer follow-up, invoice chasing, report drafting and internal knowledge lookup are better candidates than “replace our operations manager”.
Keep the human in the loop
The first useful version should prepare the work, not pretend to be accountable for it. Draft the reply, assemble the quote context, summarise the job history, flag the missing information, then let a person approve it.
Measure saved friction
The right metric is usually not “AI usage”. It is fewer missed enquiries, faster response times, less re-keying, cleaner handovers and fewer tiny decisions stealing attention from the people who actually know the business.