The market uses “AI agent” the way it used “synergy.” A copilot drafts, suggests, and stays in a human loop. An agent is allowed to take an action — file, send, buy, update — inside a policy. Small businesses should almost always ship the copilot first.
Start with a copilot when
- The job is drafting, summarizing, or classifying.
- A wrong answer is annoying but reversible (a human hits send).
- You do not yet have clean tools or a system of record.
Graduate to an agent when
- The same action happens dozens of times a week and the policy is crisp.
- You can evaluate success automatically (ticket closed, field updated, citation present).
- You can log, replay, and shut the thing off.
The first surface that gets used
Pick a workflow someone already lives in: inbox, CRM, intake form, weekly research packet. Put the model next to that work. Do not open with a greenfield “company GPT” nobody has a reason to visit. Taste in the empty states matters more than the model card.
Internal automation lives in agentic workflows. Customer-facing products live in AI product design. The ten-week version is written up in how to ship an AI product.
Continue in the practice: related service.