The WhatsApp AI agent
How the AI agent replies on WhatsApp, its knowledge, and what you can tune.
Your WhatsApp AI agent texts customers back from your business number, the way your sharpest human agent would. It answers questions, qualifies the lead, and moves the chat toward a booking. It uses your own business data to do it, and you shape its voice and knowledge in the Agent tab.
This page explains how the agent thinks, what it knows, and exactly what you can tune.
How the agent replies
When a customer messages, the agent reads the chat, looks up anything it is not sure of, and writes a short, natural reply. It always answers in words. It never goes silent on a tool lookup, and it never leaves a customer on read.
The agent is built to text, not lecture. Replies are short, in plain language, and in the same language the customer wrote in. It admits it is an AI if a customer asks outright, and it hands the chat to your team when something is beyond it.
Set up for you
MAZZA builds and tunes the agent's core behaviour during onboarding. You shape the surface: its name, tone, greeting, business instructions, and knowledge. Deeper changes to the underlying rules or model are handled by MAZZA. See Configuring your agent for the voice side of the same idea.
What the agent knows
The agent answers from four sources, in this order of authority:
| Field | Type | Description |
|---|---|---|
| Built-in rulesMAZZA controls | fixed | Core behaviour: how it texts, when it escalates, and a hard rule never to invent cars, prices, or availability. Set by MAZZA. |
| Business profile and catalogSynced | synced | Your fleet, prices, policies, hours, and holidays. These sync automatically, so the agent is always current. |
| Your knowledge entriesYou control | you add | Short questions and answers you write in the Knowledge section. The agent treats these as authoritative. |
| Your instructionsYou control | you add | Business-specific guidance you type in Behaviour, layered on top of everything above. |
The agent cannot invent facts. If a lookup returns nothing, it tells the customer it will check with the team rather than guess. To learn how the synced data flows in, read How your business profile powers the agents.
The model is fixed
The agent runs on Claude Sonnet 4.6, chosen and locked by MAZZA. You cannot change the model, and you do not need to. Your job is to give it good instructions and accurate knowledge.
How a chat moves toward a booking
The agent works like a salesperson, not an FAQ bot. It tracks where each conversation stands through five stages, and the stage shows on your dashboard.
A chat can also end in lost if the customer declines or goes quiet after a no. The agent updates the stage silently as the conversation moves, so you always see real status without lifting a finger.
The Agent tab
You configure everything in the Agent tab, which has three sections.
Behaviour
Set the agent's identity and rules of engagement.
Knowledge
Add the facts and answers it should treat as authoritative.
Playground
Test the agent on your real data before customers do.
Editing behaviour
| Field | Type | Description |
|---|---|---|
| Agent nameYou control | text | What the agent calls itself, for example Sara. It introduces itself when greeting a new customer. |
| Reply languageYou control | choice | Match the customer (recommended), always English, or always Arabic. |
| ToneYou control | choice | Professional, Friendly (the default), or Enthusiastic. This sets how casual or polished it sounds. |
| EmojisYou control | toggle | Allow the occasional emoji, or keep replies clean and text-only. |
| GreetingYou control | text | The first line a brand-new customer hears. The agent adapts it to their language. |
| InstructionsYou control | text | Business-specific guidance, for example offer free delivery on rentals of three days or more. No need to repeat fleet or prices. |
Instructions cannot override the protections
Your instructions shape tone and sales behaviour. They cannot disable the built-in safeguards. The agent will never invent cars or prices, will always be honest about being an AI when asked, and will always escalate complaints and human requests. If an instruction conflicts with these rules, the rules win.
Adding knowledge
The Knowledge section has two layers. The top layer shows what syncs from MAZZA automatically: fleet and prices, policies, hours, and any custom data tables. You edit those where they live. Below that, you add your knowledge entries: short answers to the questions customers actually ask but that are not in your fleet or policy data, such as directions, payment details, or a seasonal offer.
Open Knowledge and add an entry
Give it a clear topic, for example "Do you deliver to Abu Dhabi?", then write the answer the agent should give.
Turn entries on or off
Each entry has a switch. Only enabled entries are given to the agent. Disable one to retire it without deleting it.
Save and it goes live
Changes apply to the next reply. There is no separate publish step.
Knowledge limit
You can store up to 200 knowledge entries. The agent draws on your enabled entries in order. If you have a large library, keep the most important entries near the top, since space in each reply is limited.
Testing in the Playground
The Playground lets you chat with the agent exactly as a customer would, using your real fleet and policy data and your current (even unsaved) settings. Nothing is sent to a customer and nothing is saved. Each reply shows small chips for the lookups the agent ran, so you can see it work. Try a few real questions, in English and Arabic, before you rely on it.
Will the agent ever make up a price or a car?
No. A hard, built-in rule blocks it from inventing vehicles, prices, policies, or availability. If it cannot find an answer with its tools, it says it will check with the team.
Can I change which AI model it uses?
No. The model is fixed to Claude Sonnet 4.6 by MAZZA and is not adjustable.
Do my changes go live straight away?
Yes. Saved behaviour and enabled knowledge apply to the next inbound message. There can be a brief lag of up to a minute as new settings take effect.
What happens when the agent cannot help?
It hands the chat to a human. Read Agent modes and human handoff for how that works.
Last updated June 2026