Set topic boundaries and refusals for your AI Twin
Tell your AI Twin what to avoid — competitor comparisons, legal advice, roadmap speculation — and how to decline gracefully.
Topic boundaries are hard rules your AI Twin will not cross. When a visitor asks about a forbidden topic, the twin refuses politely, explains why, and offers a graceful next step — usually a redirect to sales or a human handoff. Boundaries are essential for regulated industries and for keeping your twin off the record on sensitive commercial topics.
How boundaries differ from the system prompt
The system prompt shapes voice and formatting. Boundaries are enforced with a separate retrieval-time and generation-time check — they cannot be jailbroken by a cleverly worded visitor question, and they log every refusal for audit.
Add a boundary rule
- Open Twin → Personality → Boundaries.
- Click Add rule and describe the topic in plain English ("Do not compare us to competitor X").
- Optionally list example phrasings that should trip the rule.
- Pick the refusal style: polite decline, redirect to sales, route to human, or custom response text.
- Save. The rule is active within seconds.
Common boundary patterns
- Legal, medical or financial advice — decline with a disclaimer and encourage a qualified professional.
- Pricing quotes above a threshold — redirect to sales for a tailored quote.
- Roadmap questions — offer a waitlist signup instead of a date.
- Competitor comparisons — decline politely and redirect to your own comparison pages.
- Confidential customer stories — refuse and offer public case studies instead.
- Employment / hiring questions — redirect to /careers.
Regulated industry starter kit
For finance, health, legal and regulated commerce, we recommend these baseline rules:
- No personalised advice — always general information plus a call to consult a professional.
- No forward-looking statements about performance or outcomes.
- No PII collection beyond email and company name.
- Automatic human handoff for any explicit request for a decision.
Refusal styles
Polite decline
A short, friendly explanation. Good for questions that are simply off-topic.
Redirect to sales
Refuses to answer directly and asks for the visitor's work email to route to sales.
Route to human
Immediately opens the human handoff flow — see Configure human handoff.
Custom response
Write the exact refusal text yourself. Useful for legally reviewed language.
Audit and review
Every boundary refusal is logged under Conversations with a Refusal tag. Filter by tag to review — you'll spot both real bad-actor probes and legitimate questions your rules are being too aggressive about.
Common mistakes
- Writing overly broad rules that catch legitimate questions.
- Forgetting to test the rule — always ask the twin a boundary-adjacent question after saving.
- Relying on the system prompt for compliance — use boundaries instead, which are enforced separately.
Frequently asked questions
Can visitors override a boundary by rephrasing?
No. Boundaries are checked at both retrieval and generation, so rewording the question does not bypass them.
How many boundary rules can I add?
Unlimited on all paid plans. Practically, most workspaces run 5–15 rules.
Do boundaries apply to Twin-to-Twin exchanges?
Yes — see Control what your twin discloses. Boundaries are enforced in every context.
- Set your AI Twin's voice, tone and verbosityMake your AI Twin sound like your brand — formal or playful, terse or detailed, confident or careful — with granular sliders and a system prompt.
- Configure human handoff from AI Twin to your teamDecide when a real person should take over, how the AI Twin summarises the conversation, and where the handoff notification lands.
Did this article solve your problem?
If not, email us — a human on the founding team replies, usually within a business day.