A chatbot answers. A copilot drafts. An agent acts. Pick on purpose.

Three products, three risk levels

A chatbot answers questions in a box. It should not update systems. RAG can make it honest about policies.

A copilot sits in a person’s workflow: draft the note, suggest the next step, highlight the clause. A human still decides. This is usually the right first pilot.

An agent can call tools and take multi-step actions. That is where value and disaster both live. You only go here when the band is low-risk, the evals are real, and write actions are gated.

• Customer FAQ on policy → chatbot + RAG.
• Adjuster writes notes all day → copilot.
• Straight-through pay on tiny, clean claims → agent with a cap, then a human on exceptions.

The case test

If handle cost is the prize and leakage is ugly, do not buy an agent to save minutes. If leakage is tiny on a high-volume easy band, an agent can win — after a copilot has proven the retrieval quality.

Say “chatbot” when the user is a customer asking a question. Say “copilot” when the user is an employee. Say “agent” only when you are also saying which tools, which cap, and which kill switch.

Clearclaim Mutual is built on this split. The lesson behind it is twelve minutes.