Setting up your AI agent
Give the agent a persona, choose an engine, teach it to send files, and decide when it hands over to a human.
The AI agent replies to inbound WhatsApp messages on your behalf, in your voice, around the clock. It is off by default — you configure it, then switch it on.
Go to Settings → AI.
Choose an engine
You can run on Google Gemini or OpenAI, configured per workspace. Both handle multi-turn conversation and understand voice notes — the OpenAI path transcribes them, the Gemini path understands the audio natively.
Write the instructions
This is the part that matters most. The agent's behaviour comes almost entirely from what you write here. Be specific about:
- Who it is — "You are Chanda, a sales assistant for Kabo Motors in Lusaka."
- What you sell, with the details customers actually ask about: prices, opening hours, delivery areas, payment methods.
- Tone — formal or casual, how much English versus local phrasing.
- What it must never do — invent prices, promise delivery dates, discuss competitors, give legal or medical advice.
- When to hand over — "If the customer asks for a refund, or is angry, stop and hand over to a human."
There is a creativity setting: lower keeps it factual and repetitive, higher makes it more conversational and more likely to improvise. For sales, stay low.
The agent has no document knowledge base
It knows what you write in the instructions, what is in the current conversation, and the skills you give it. You cannot upload a folder of PDFs and have it answer from them. If a fact matters, put it in the instructions.
Teach it to send things
Custom skills let the agent deliver a file when the moment is right — "when someone asks for the menu, send the menu PDF", "share the price list". Each skill is a pre-set image, PDF, audio or video with an optional caption, and the agent decides when to use it.
Skills can also send an approved message template, with the variables derived from the template itself.
A skill can also start an automation flow, for when a question is better answered by a fixed sequence of steps than by the model — booking intake, a returns process, a qualification questionnaire. Build the flow under Automation, set its Start node trigger to AI agent, turn it on, then add a skill of type Automation and bind it. Once the flow starts, the agent stops replying until the flow ends, so the two never talk over each other.
Skills send assets, they do not call your systems
A skill delivers a file or a template. It does not query live inventory, look up an order in your ERP, or book a slot in a calendar.
Handover to humans
The agent pauses itself when it decides a human is needed, and the conversation appears in the inbox as normal. You can also toggle the AI off for a single conversation from the chat header — useful when a rep takes over mid-thread.
Automations and the AI work together: a flow can hand a conversation to the agent, and the agent can hand it back to a person.
Orders
If you use the product catalog, the agent's behaviour when an order arrives is a separate setting:
- Acknowledge (default) — send the customer a receipt, pause the AI, and drop a note in the thread for your team.
- Silent — pause the AI, no receipt.
- AI — let the agent handle the order conversation itself.
Switching it on
Set the agent active in Settings → AI. It only responds to inbound messages, and it does not reply to your own outbound ones.
If it is not responding, the usual causes are: the config is inactive, the AI is paused for that specific conversation, an automation flow has precedence, or the connection's token is invalid.