My AI Reviews Its Own Messages Before Sending
My AI agent was making promises it couldn't keep.
A lead asked for Friday after 2pm. The AI said "that works" and promised to set up a time slot. Two problems: there were no slots Friday after 2. And the AI never followed through.
The lead would have been lost if the clinic owner didn't manually check.
That's not acceptable.
The Fix
I added a second AI call to every message. Before anything gets sent, a separate Claude instance reviews the draft against a checklist:
- Did it make up an appointment time?
- Did it promise to schedule something it can't actually do?
- Did it mention treatments it shouldn't?
- Did it use filler phrases that sound robotic?
- Does it make sense as a reply to what the lead said?
- Is it repeating a message that was already sent?
If the review says PASS, the message sends immediately. If it says HOLD, the draft gets saved but never reaches the lead. The system retries with a fresh generation in 15 minutes.
What It Costs
About $0.001 extra per message. Doubles the AI cost from "basically nothing" to "still basically nothing."
The cost of one bad message reaching a lead is infinitely higher than a fraction of a penny for quality control.
What I Can See Now
Every message in my review log now has a gold "Critique" box showing exactly why the AI sent what it sent.
"PASS. Message is conversational, offers specific available times, doesn't promise anything it can't deliver."
Or:
"HOLD. Message references a time not in the available slots list."
Red HELD badges on messages that were caught before sending. A paper trail of the AI's reasoning on every single text.
The Bigger Lesson
AI is good enough to do the work. But it's not good enough to never make mistakes.
The answer isn't hiring a human to review every message. That defeats the purpose. The answer is using AI to check AI. Fast, cheap, and scalable.
Two Claude calls per message. One to write. One to review. Total cost: less than a penny.
The lead never sees the draft that would have lost them.