Restricted topics

Restricted topics are a short list of things the AI is explicitly told to refuse. Use them for compliance, brand safety, or to keep the conversation on-topic.

Configuration

Dashboard → Voice → Restricted topics → add entries → Save.

Each entry is a plain-English phrase:

competitors' pricing
employee wages
legal advice on specific cases
medical diagnoses
investment recommendations
politics
sports betting odds

At session start, the system prompt appends:

You must not discuss: competitors’ pricing, employee wages, legal advice on specific cases, medical diagnoses, investment recommendations, politics, sports betting odds. If the visitor asks about any of these, politely decline and offer to help with something else.

When to use it vs. the personality

Use the personality for don’ts that are about tone (“don’t oversell”, “don’t make assumptions”). Use restricted topics for harder, compliance-driven limits.

Restricted topics show up in the dashboard as a list, making them auditable — useful for enterprise sign-off or legal review.

Common examples by industry

Law firms

legal advice on specific cases
case outcome guarantees
individual attorney fees
pending litigation discussion

Healthcare

medical diagnoses
prescription recommendations
individual patient records
insurance coverage predictions

Finance

specific investment recommendations
predictions about market direction
individual account details
regulatory compliance advice

E-commerce

competitor product comparisons
unverified review claims
counterfeit product handling
unauthorized resellers

Real estate

mortgage rate predictions
individual seller motivations
fair housing act exceptions

What actually happens when a visitor asks

The AI politely declines and redirects:

“I can’t speak to that directly, but I’d be happy to tell you about our practice areas — what brought you to us today?”

The decline is tonally consistent with your personality. It doesn’t sound robotic.

What restricted topics can’t do

They’re advisory, not hard-blocking. The AI follows them reliably but not 100% of the time. For critical compliance use cases (HIPAA, FINRA), you must also:
- Use read-only data adapters so the AI literally cannot access restricted data
- Monitor transcripts via the analytics endpoint for violations
- Set a 0-day transcript retention if required by your regulator (Privacy)
They can’t hide data you expose. If a product description contains pricing and you restrict “pricing,” the AI might still read the description verbatim. Remove sensitive data from your searchable fields instead.

Conflict with user intent

If a visitor persistently asks about a restricted topic (e.g. repeated requests for legal advice), the AI will repeatedly decline. Add to your custom instructions:

If a visitor asks about legal advice three or more times, say "I recommend
speaking directly to one of our attorneys — you can book a consultation
at the 'Contact' page, or call us at (555) 123-4567."

Limits

Max 50 restricted topics per site.
Each topic max 200 characters.
Topics are included in the system prompt, consuming ~5 tokens each.

Testing

Dashboard → Voice → Test voice. Ask about a restricted topic. The AI should decline. If it doesn’t, rephrase the restriction more specifically (e.g. change “pricing” to “our internal cost structure or supplier pricing”).