Appointment automation

Booking by message is how a great deal of business already works in Colombia. The friction is not the customer — it is what happens after they press send. Someone has to read the message, work out what was being asked, open a calendar, check whether the slot is free, book it, and write back. At 11pm on a Sunday, nobody does.

The harder part is the format. People do not type carefully worded requests; they send a forty-second voice note while driving. Any system that only accepts clean text has already lost most of its input.

1. 01

   Take the message as it actually arrives

   A WhatsApp trigger fires on the inbound message and the audio is pulled down before anything else happens. Voice is a first-class input here, not a fallback path bolted on later.

2. 02

   Transcribe before reasoning

   The audio goes through Whisper and becomes text. Everything downstream — planning, tool selection, the reply — works from that transcript, which keeps a single, inspectable representation of what the customer said.

3. 03

   Normalize and resolve

   The transcript is shaped into a predictable payload and the loose parts of a human request — “next Tuesday afternoon”, “same as last time” — are resolved into concrete fields the rest of the system can act on.

4. 04

   Let the agent plan, with tools

   An agent runs with a chat model and conversation memory, and is given real tools rather than instructions to imagine an outcome. It can check availability and it can create a booking. Those are the two things it can do, which is the point — the boundary is the design.

5. 05

   Close the loop

   A confirmation goes back on the same thread the customer started. From their side it is one conversation, and no part of it required a person.