An AI agent reads inbound WhatsApp messages and answers them. It has a persona you write, a model you choose, and — optionally — tools it can call and documents it can search. It works the hours nobody is online, and gets out of the way the moment a person steps in.
When an agent replies #
Four conditions all have to hold. If a customer is not getting an AI reply, it is almost always one of these:
-
The agent is active
An inactive agent is ignored entirely. -
It is bound to the number the message arrived on
Agents are attached to specific WhatsApp accounts. A message on an unbound number gets no AI reply. -
No human has taken the thread over
A human reply takes the conversation off the bot. See Working a conversation. -
The contact is not excluded
The per-contact bot switch is a standing instruction that beats everything above.
What an agent is made of #
A persona
The system prompt: who it is, what it may say, what it must refuse. The single biggest lever on quality.
Tools
REST and database calls it can make mid-conversation to look something up or write something down.
Knowledge
Documents it searches before answering, so replies come from your material rather than the model's guesses.
Memory
How many previous turns it carries into each reply.
A sensible way to roll one out #
-
Connect a provider and register a model
Nothing works until there is a model to call. See Providers & models. -
Write a narrow prompt
Start with one job done well, not everything done vaguely. -
Test it in the playground
Push the awkward questions at it before a customer does. See Playground. -
Bind it to one number and watch
Read the audit log for the first day. Take over anything it handles badly and adjust the prompt. -
Then widen it
Add tools and knowledge once the basic behaviour is right — not before.