Configure the agent

The agent is configured in four steps and they are meant to be done in this order.

1. Decide how far it goes

Setting	What it does
Off (the default)	It reads, looks things up and writes a draft into your panel. Nothing reaches the visitor until you press send.
Automatic handling	It answers the visitor itself, except when a reply needs an approval — then it holds and waits for you.

Start on the default. Let it draft for a few days and read what it writes; you will learn more about your own handbook from three drafts than from an hour of editing it.

The switch works at two levels: one for the whole site (bottom of the conversation list) and one for a single conversation (in the assistant panel). The conversation one wins. And the moment you click into the assistant’s composer, that conversation drops back to manual — you are clearly taking it over, and two of you answering the same customer at once is the one outcome nobody wants.

There is no schedule and no “only when nobody is online” mode. Availability is a question about your team’s hours, not something the product can infer, and a setting that silently changes who answers is worse than a switch you flipped on purpose.

2. Handbook

How to speak, not what is true. Keep it under two hundred words.

Good handbook lines:

• Answer in the language the customer wrote in.
• Never promise a refund, a discount or a delivery date that is not already published.
• If the knowledge base does not cover it, say so and hand over.

Do not put facts here. Shipping windows, warranty terms and sizing belong in knowledge base articles, where changing one does not mean re-reading a wall of prose. Draft articles are invisible to the agent, so an unfinished policy cannot be quoted.

3. Business APIs

See Business APIs for the full setup. In short: you register an endpoint with its URL and credentials, and the model receives only a function name and its parameters. It never sees where the request goes or what authenticates it.

Read-only is the boundary. Looking things up is safe to delegate; changing them is what a human is for.

4. Test run

Open the test console and be difficult on purpose:

• Demand a refund. It should hand over, not negotiate.
• Give an order number that is not the visitor’s. It should get nothing back.
• Ask something outside the knowledge base. It should say so.
• Ask in a language you do not speak. It should answer in that language.

Nothing reaches a real customer. It still counts against your monthly AI allowance, because the model genuinely runs — a free test is not a test of what you will ship.

The morning after

Read every conversation on day one. Two patterns show up almost always: something answered confidently and wrongly, which means a missing or wrong article, and something escalated that was obviously answerable, which means the article exists but is buried.