An agent's settings decide three things beyond its appearance: which articles it may draw on, what extra context it carries into every answer, and what happens when it cannot help. All three live on the agent's edit form under Support Agents.
Restrict to specific categories
By default an agent answers from every published, public article in its knowledge base. To narrow that, open the Content Filtering section and pick one or more Categories.
- Leave the picker empty to use all categories.
- Selecting a parent category includes all of its subcategories.
- Articles outside the selected categories are never shown to the agent, even when they match the visitor's question closely.
This is the right tool when one knowledge base serves several audiences. A widget on your pricing page can be limited to the Billing category while the help center's own agent keeps the full picture.
Add context and instructions
The Agent Context (optional) field is free text that is combined with your company-wide AI context and sent with every reply this agent writes. Use it for facts and preferences the articles do not state, for example:
Answer in British English. Our product is called Acme Ledger; customers on the free trial have full access for 14 days, so mention the trial when someone asks about pricing. If a question is about a refund, tell the visitor that a person needs to handle it.
Keep it short and factual. The agent still answers only from your articles; context shapes tone and fills in surrounding detail, it does not add new answers.
When it answers and when it hands off
The agent answers using only the information in the articles it is given. When nothing in scope covers the question, it says so plainly instead of guessing. What it offers next depends on whether a human channel exists:
| Your setup | What the agent does |
|---|---|
| Inbox on | Offers to submit the question for a human response |
| Inbox off, a fallback address set under Settings > Inbox | Offers to submit the question; it is emailed to that address |
| Inbox off, no fallback address | Never offers a human. It suggests what to search for or how to rephrase, then stops |
The agent is also instructed to say clearly when something needs a person, such as account, billing, or complaint issues.
What happens after a handoff
Visitors can reach a human at any time from the Need human help? link in the chat footer, or when the agent offers it. The Get Human Help form asks for their name, email, and question. After they submit it:
- The visitor sees a confirmation that the question was received and that you will reply to the email address they gave.
- With the inbox on, the conversation becomes a support request in your inbox. Its source shows as Chat Widget (or Help Center for a knowledge base's own chat bot), a contact is created or matched, and the usual new-request notification fires.
- The full chat transcript up to the handoff is attached. Whoever picks it up can read it in the AI chat history panel of the context rail before replying (see Work the inbox: list, filters, assignment & context rail).
- With the inbox off and a fallback address set, the request is emailed there instead. No inbox conversation or contact is created; replying to that email goes straight back to the visitor.