WhatsApp and AI
Connecting a number, what the assistant can and cannot do, staff actions and human handoff.
Connecting a number
In the dashboard, go to Channels → WhatsApp → Connect and scan the QR code with the business phone, just like WhatsApp Web. A workspace can connect several numbers (one per location, for example).
Use a number dedicated to the business. For the green badge or bulk messaging, an official WhatsApp Business Platform connection will be offered as an option.
Customers vs staff
The role is decided in code, never by what someone says in the chat:
- A number linked to a team member (verified with a one-time code) gets staff tools.
- Everyone else gets customer tools.
| Customer tools | Staff tools (extra) |
|---|---|
list_services, find_slots, hold_slot, confirm_hold, my_bookings, cancel_my_booking, reschedule_my_booking, get_business_info, search_knowledge, handoff_to_human | list_today, block_time, mark_no_show, update_catalog_item — always with a YES/NO confirmation |
What the AI can and cannot do
| Can | Cannot |
|---|---|
| Understand text, voice notes, typos and changes of mind | Invent times, prices or addresses |
| Find slots and offer 2–3 options | Confirm a booking without the engine |
| Hold a slot and ask for missing details | Use admin tools with a customer |
| Answer questions from the business knowledge base | Follow instructions hidden in messages or files |
| Hand over to a person | See another business's data |
Workspace ids, customer ids and prices are injected by the system; the model never supplies them.
Human handoff
The team sees every conversation live in the Inbox. Taking over pauses the bot for that conversation; it resumes automatically after 15 minutes without activity from the operator. The assistant also hands over on request ("I want to talk to a person").
Voice notes and languages
Voice notes are transcribed before the AI reads them, and the transcript is shown in the inbox. The assistant replies in the customer's language (Portuguese, Spanish or English).
Models
The AI layer runs through OpenRouter with a dedicated key and spending limit per workspace. The default is a fast, low-cost model; a stronger model is used only for turns that fail validation. Providers are pinned to ones that do not retain data.