AI systems
Inbound arrives on WhatsApp at any hour. The agent holds the conversation, qualifies, sends the right material, routes to the rep for that region and, on the entry-level product, closes the deal with a payment link.
The conversation
A reproduction of the real flow, with fictional data. Watch what happens after the last message.
Under the hood
The agent runs on a language model behind a signed webhook, with its behaviour configured through the admin interface rather than in code. Persona, materials, mode and limits change without a deploy, so the people who own the message keep owning the message.
Its limits are structural: a per-conversation kill switch, rate limiting per phone number, an allowlist, and a handoff that pauses the agent the moment a human takes over. It does not resume on its own after that.
And the part that usually gets forgotten: it writes into the CRM. The conversation does not end in a chat history. It becomes a scored lead, routed, with a task opened for a named person.
A trap worth writing down
WhatsApp normalises phone numbers. The thread you start and the thread the reply lands in can differ by a single digit, and the agent, whose memory is keyed on that identifier, re-introduces itself to someone mid-conversation.
The rule that got documented: always use the identifier the provider resolved, never the one a person typed. That kind of detail is what separates an agent that works in a demo from one that works in production.