Guardrails are the rules that keep an AI agent useful and safe: what it may say, what it may do and when it must step aside. They matter more than how clever the model is. Here's a checklist to work through before your agent talks to customers.
1. Ground every answer
- Turn on grounded answers only, so replies come from approved sources.
- Write a fallback reply for when nothing relevant is found.
- Review sources for contradictions. See grounding with a knowledge base.
2. Block topics you shouldn't discuss
List subjects the agent must not answer, and what it should do instead:
- Medical, legal and financial advice: hand off or use approved wording
- Competitor claims: decline politely
- Pricing not on your site: hand off
- Personal opinions, politics, unrelated chat: steer back
Industries such as clinics and financial services need stricter lists.
3. Verify before sharing personal data
Anything tied to an order or account should require verification, for example an order number plus the email used at checkout. Never let the agent reveal an address, phone number or order contents on a name alone.
4. Confirm before acting
Any action that changes something (cancelling, exchanging, updating an address) needs an explicit "Yes, go ahead" from the customer. Show what will happen before it happens.
5. Set limits
- Value limits: refunds or credits above a threshold go to a person.
- Execution limits: cap how many steps or API calls a single conversation can trigger.
- Rate limits: protect against loops and abuse.
6. Use reviewed replies for sensitive questions
For questions where exact wording matters, like warranty terms or allergy statements, write an approved answer once and have the agent use it verbatim. LiveCart calls these reviewed replies.
7. Disclose that it's AI
Label the agent clearly in chat, and say it out loud at the start of voice and avatar calls. Don't give it a human name and photo that imply otherwise.
8. Always offer a person
"Talk to a person" should always work. Make it visible in the widget and honour it immediately. See handoff best practices.
9. Version and review changes
Publish agent changes as versions, test before publishing, and keep the ability to roll back. Review a sample of conversations every week, especially after a change.
10. Log everything
Keep transcripts, sources used and actions taken, so you can explain any answer after the fact.
None of these slow a good agent down. They're what let you trust it with more over time.
