← sources

The booking bug that taught us to never let the model name the service

source: reddit (reddit:AI_Agents)published 2026-10-26 12:55ingested 2026-10-27 02:35

retained text

We run WhatsApp and phone agents for small businesses (gyms, spas, clinics, retail) that book customers into MindBody, Google Calendar or Shopify, in English and Spanish.  
  
Early on, the agent would understand the customer perfectly and still book the wrong thing. A customer asks for "the intro class", the model passes "intro class" to the booking system, and the system matches whatever it matches. In a catalogue with hundreds of items, including retail products, that is a coin flip.  
  
What fixed it was taking that decision away from the model:  
  
1. The agent qualifies what the person needs, in conversation.  
2. A resolver maps it to one real catalogue id. If it can't, the agent asks, or hands the conversation to a person with the context attached.  
3. Availability comes from the provider, never from the model.  
4. Only then does it book, and it confirms in plain prose.  
  
A free-text service name reaching the booking API is now a hard block in code, not a prompt instruction.  
  
The general lesson: let the model own the conversation, and let deterministic code own every fact the business is held to: ids, prices, times.  
  
Curious how others draw that line. Where do you stop trusting the model's output and start validating it?  
  
Disclosure: I build HIVRA, AI voice and text Agents that empower your team

document fetched 2026-10-27 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xfw123/the_booking_bug_that_taught_us_to_never_let_the