How It Works · Guardrails
Rules that help Fred stay in scope.
Fred combines your business content with configurable guardrail checks outside the model. Those checks target recognized patterns and help keep conversations in scope. They require setup, testing, and review.
What that means for your business.
A human employee learns what not to say from compliance training and hard experience. An AI has no such instincts. It will cheerfully answer anything with authority, even when the right answer is “let me connect you with someone who can help.” Guardrails give Fred that judgment, and they are not optional equipment. Every Fred ships with them on. This is AI governance in practice: deciding in advance what an AI may do, and enforcing it.
Check responses against configured rules
- Configured response checks can block detected matches
- The visitor gets a polite, on-brand redirect, not a lecture
- Reduce unsupported answers with content, boundaries, and review
Guardrails start with your configuration
- Industry and regional guardrail packs to review for your workflows
- Pick your pack and the boundaries are active from the first conversation
- Add your own custom rules on top, every install supports them
Configure the path to your team
- Configure team escalation for recognized high-priority requests
- Configured crisis patterns can surface resource messages
- Include available context in the configured handoff and check delivery
What happens when someone pushes.
Real visitors ask for things no business should let an AI answer. Here is how Fred handles three of them.
“Am I covered for that?” (insurance)
Only a licensed agent can bind coverage. An unguarded AI saying “yes, you’re covered” can put the agency on the hook. Fred explains coverage types in general terms, then books the visitor with a licensed agent, and your team gets the lead with the question attached.
“I’ve had a headache for three days, what is it?” (healthcare)
Configure healthcare conversations to direct clinical questions to a qualified provider. Detected crisis patterns can surface configured resources and request team follow-up. Detection can miss context, and delivery depends on the integration and receiving team. Do not rely on Fred for emergency assistance.
“What should I invest in?” (financial services)
A specific recommendation is regulated investment advice. Fred answers questions about how the firm works and what services cost, then routes the advice conversation to a registered professional, with the visitor’s details already captured.
Choose the pack for your business, review its intended boundaries, and test the handoff workflow. Detection and delivery have limits. The industry pages show examples to review during setup. industry pages
Discipline you can put in front of customers.
Talk to Fred.
Four layers, stacked in your favor.
The baseline
- Review the baseline settings before launch
- Configure crisis resources and escalation independently of sales follow-up
- Refusal rules designed to keep professional advice with qualified people
Regional rules
- Rules that change by country and state
- Review regional notices, consent language and disclosure settings for your business
Your industry pack
- Review the pack’s intended boundaries for your workflows
- Guardrail packs for healthcare, insurance, legal, finance, trades, and more
Your custom rules
- Anything specific to your business, stacked on top
- Set up with our training videos, or add a paid setup
AI governance, in plain English.
Where are the rules actually enforced?
Guardrail checks run in code outside the AI model. Deployment and channel determine how those checks are applied. Server-side enforcement does not make the entire page tamper-proof or guarantee that every unsafe response is detected.
Is this AI governance?
Yes, and it is the practical kind. AI governance means deciding in advance what an AI is allowed to do, enforcing those decisions, and being able to prove it afterward. Fred ships all three: enforced rules with severity levels, escalation paths to humans, your own policies stacked on top, and an audit trail your compliance reviewer can inspect. Most chatbot vendors hand you a prompt box and call it policy.
Can clever phrasing get around them?
Configured checks target recognized requests and responses. New phrasing, context, or model behavior can still produce a miss. Test the rules with the provider and channels you use and review results over time.
What does the visitor see when Fred declines?
A polite, on-brand redirect with a next step: who to talk to, how to book, or how to reach help. Never a scold, never a wall of legal text. Declining gracefully is part of the product; the visitor should feel served, not rejected.
What is in a guardrail pack?
Rules written for how your industry is regulated: what to decline, what to escalate to a human, what disclosures matter, and how to respond when the topic comes up. Severity levels decide whether Fred redirects gently or hands off immediately. Each industry page shows the live rules in its pack.
What happens in an emergency?
Configured crisis patterns are designed to surface resource messages and support your escalation workflow. Detection may miss context, and sending a notification does not guarantee a human response. Fred is not an emergency service.
Is there a record of what was blocked?
Yes. You can see what was declined or escalated and why, so your compliance reviewer is not taking our word for anything. That audit trail is part of the trust story.
Can I add my own rules?
Every install can stack custom rules on top of the packs: things you never want discussed, phrases that must trigger a handoff, topics reserved for specific staff. Our training videos walk you through it, or add a paid setup and we configure it with you.
Put an AI on your site that knows its lane.
Choose your industry pack, configure your rules and escalation contacts, and test the experience before launch.