WhatsApp AI works in five steps: a customer's message arrives; the AI reads and understands it; it checks the business information it was given (services, prices, policies); it decides what to do based on rules you set; and it replies in your tone — or escalates to a human. No magic, and no knowledge it wasn't given.
Step 1: The message arrives
Your business number is connected to WhatsApp through the official Business API. When a customer messages, the conversation is delivered to the software that hosts your agent — securely, the same way any WhatsApp Business tool receives messages.
Step 2: The AI understands it
A language model reads the message. It doesn't look up keywords; it understands meaning. “Hey, do you guys maybe still have that special on tyres?” and “Tyre special still available?” are recognised as the same question. This is the part that separates an agent from a script — and it's also the part vendors overhype.
Step 3: It checks what the business knows
This is the step most people miss, and it's the one that makes AI useful instead of dangerous. The agent doesn't answer from general knowledge — it answers from your business information: the services, prices, hours, policies and tone that were captured when the agent was built. If your price list says R1,200, the agent quotes R1,200, not a guess.
This is called grounding. An ungrounded AI makes things up (hallucination); a grounded one is limited to what's true about your business. Which would you rather have answering customers?
Step 4: It follows your rules
The agent also has rules — a policy layer. What it may answer, what it must ask before answering, what it must never promise, when it must escalate, how it hands over. These are written with you during setup. The model provides the language ability; your rules provide the judgment boundaries.
Step 5: It replies or escalates
Inside its rules, the agent replies in your tone, in the customer's language, with the correct information. Outside its rules — an angry customer, a negotiation, an emergency — it hands the conversation to a human with a summary attached. The customer never has to repeat themselves.
What the AI model is (and isn't)
The language model is a general-purpose engine — like an engine in a car. It doesn't know your business any more than a new employee does. The agent's value comes from what's built around it: your information, your rules, your tone, your channel setup. That's why two businesses can use the same model and get completely different results.
It's also why WhatsMyBiz never claims to have created the AI models. We build the intelligence layer around your business — the part that actually determines whether the agent helps or harms.
The South African angle
The same five steps work in any language, which matters in SA's 11-language market, and the whole chain runs in the cloud, which matters during load shedding. The rules layer is where POPIA gets applied (what data is captured, how it's used), and the knowledge layer is where ZAR prices, local suburbs and SA service reality live.
Frequently asked questions
Does WhatsApp AI understand Afrikaans or isiZulu?
Modern language models understand and reply in the language of the message. The knowledge and rules work the same in any language.
Can the AI make up answers?
Only if it's not grounded. A properly configured agent answers from your business information and escalates when it doesn't know — it's built not to invent.
Who owns the AI model behind my agent?
The model provider does (the same way a phone company owns the network). What your business owns is the configuration: your information, rules and tone — which is where the value is.
Is my customer data seen by the AI company?
Data handling depends on the provider and model. A POPIA-aware setup minimises what's stored, uses it for the customer's expected purpose, and is documented — ask your provider directly.