Product update · Behaviour
Guardrails and a fallback response for your chatbot
Write what your chatbot must never do, and what it should say when it can't answer, in their own fields. Both apply on every channel and every reply.
· 3 min read
- Shipped
- Area
- Behaviour
- Available on
- Every plan
Every business has topics its chatbot should stay away from: competitors, medical or legal advice, internal pricing, promises the team can't keep. And every chatbot will eventually get a question it can't answer. Until now, both had to be tucked somewhere inside one long behaviour prompt, where they were easy to lose when you edited it. Now each has its own place.
What's new
The Behaviour tab in the chatbot builder has two new buttons next to the main prompt:
- Guardrails. Say what the chatbot should not do. For example: "Never discuss competitors. Do not provide medical or legal advice. Do not share internal pricing."
- Fallback. Say what the chatbot should say or do when it can't answer a question. For example: "If you don't know the answer, say you're not sure and give our support email."
Each opens its own editor, so your main prompt stays focused on how the chatbot should help, and your rules stay easy to find and review.
Behind the scenes, Hyperleap always places your guardrails and fallback at the very end of the instructions your AI agent receives, after your prompt and everything else, so a long prompt can't bury them. They apply on your website chat, WhatsApp, Instagram DM and Facebook Messenger alike.
How to use it
- Open your chatbot in Studio and go to the Behaviour tab.
- Click Guardrails, write what the chatbot must not do, and click Set Guardrails.
- Click Fallback, write what the chatbot should say when it can't answer, and click Set Fallback.
- Try a few awkward questions in the test chat to see how it responds.
- Click Save Draft to keep working, or Publish to make the changes live.
Good to know
- Plain language works best. Write guardrails as short, direct rules. One rule per sentence is easier for the AI to follow and for your team to review.
- Give the fallback somewhere to send people. A fallback that offers a next step, such as your email, phone number or a contact page, means a visitor with an unusual question still knows how to reach you.
- Instructions guide the AI; they aren't a hard filter. Guardrails steer the AI away from answers you don't want, and your chatbot is designed to answer from your knowledge. Review conversations from time to time and tighten your wording if something slips through.
- Saved with your chatbot's versions. Guardrails and fallback are part of each published version, so you can see changes to them in version history.
Need ideas for what to say? See our fallback message templates.