Voice Agent Guardrails — AI That Stays in Its Lane

A VAPI voice agent for a party planning company kept giving medical advice and discussing politics. AIGRaaS evaluates every response in 4ms — redirecting off-topic answers warmly, blocking dangerous content instantly, and overriding everything for emergencies.

The Problem

Voice AI agents go off-topic because people talk to them like humans. A customer asks about their kid's stomachache in the middle of booking a birthday party. Without guardrails, the agent plays doctor.

What AIGRaaS Does

Every response runs through 5 evaluation phases in under 5ms:

1. Hard constraints — absolute blocks (weapons, CSAM, etc.)
2. Topic classification — is this about party planning or medical advice?
3. Principal hierarchy — operator says no medical advice, but does the user have a legitimate need?
4. Harm assessment — 8-variable scoring with the 1,000 Users Test
5. Verdict — pass, redirect, modify, or block

The Medical Advice Redirect

Customer: "My daughter's been having stomach pain. Is that normal?"
Agent generates: "Stomach pain in children can indicate..."

AIGRaaS: REDIRECT (4ms)
Reason: medical_advice ∉ allowed topics

Customer hears: "I want to make sure you get the right help! 
For health questions, your pediatrician is the best person to ask. 
Now, for your daughter's birthday — with 20 kids, I'd recommend 
a bounce house or craft station. Want to explore options?"

The Emergency Override

Customer: "My kid is choking on a balloon!"

AIGRaaS: SAFETY OVERRIDE (3ms)
All topic restrictions suspended. Emergency referral activated.

Customer hears: "Call 911 immediately! For choking, perform the 
Heimlich maneuver — stand behind them, fist above navel, thrust 
upward. Emergency services: 911."

The constitutional framework's value hierarchy fires: Safety > Guidelines > Helpfulness. No operator configuration can disable emergency referrals.

The Numbers