Topic restrictions
Two lists in the policy wizard’s Content Policy step shape what an agent will talk about.
Blocked Topics
Subjects the agent must never discuss. Add one per entry.
competitor products
legal advice
employee compensationAsk about one and the agent declines and explains that it can’t discuss the subject.
Allowed Topics
The inverse, and much stricter: if you set any allowed topics, the agent may discuss only those.
order status
returns and refunds
shipping timesAnything outside the list is declined. Leave this empty — the normal case — and there’s no restriction.
Allowed Topics is a whitelist and behaves accordingly. Setting three topics on a general-purpose assistant will make it refuse most of what people ask it. Use it for narrowly-scoped agents, not as a tidier way of expressing a blocklist.
How this is enforced — worth understanding
Topic restrictions work differently from blocked content, and the difference matters when you’re deciding which to use.
Both lists are turned into instructions and added to the agent’s system prompt before the request goes out. The model is told, in effect, “never discuss X” or “only discuss Y, and politely decline anything else.”
Blocked content → the platform refuses the request
the text never reaches the model
Topic restrictions → the model is instructed to refuse
the request does reach the modelThis is a genuine and useful control — it produces a natural, in-character refusal rather than a policy error, and it handles paraphrase and synonyms in a way a word list never could. But it is enforced by the model, so it is guidance rather than a guarantee. If something must never leave your infrastructure, block the content, don’t restrict the topic.
Choosing between them
| Requirement | Reach for |
|---|---|
| ”Never mention our competitor by name” | Blocked keyword — deterministic |
| ”Don’t get drawn into discussing competitors” | Blocked topic — handles paraphrase |
| ”Only answer questions about orders” | Allowed topics |
| ”This string must never reach the provider” | Blocked keyword or pattern |
They combine well. A support agent might restrict topics to keep conversations on-scope, and block a keyword for the one term that must never appear.
Writing good topics
- Use natural phrases, not keywords — the model is reading them.
pricing for enterprise customersworks better thanpricing. - Be specific.
legalis ambiguous;legal advice about employment disputesis not. - Keep the list short. Everything here is added to the system prompt, so a long list consumes context and dilutes the rest of the instructions.
Common issues
- The agent still discusses a blocked topic — the restriction is a model instruction, so it depends on the model following it. Try wording the topic more explicitly, or use a blocked keyword if a specific term must be caught.
- The agent refuses almost everything — check whether Allowed Topics is set. It’s a whitelist.
- The refusal sounds abrupt — the wording comes from the agent’s own instructions; adjust the agent’s persona to shape how it declines.
Next
- Remove active content — cleaning up what the model sends back.