WhatsApp AI chatbot: how to build support that qualifies leads

Qualifying is not collecting data. It is finding out three things: whether the person has the problem you solve, whether they can afford to buy and whether the timing makes sense. A good reference framework is an adapted BANT: budget, authority, need and timeline. You do not have to extract everything at once, but you do need to leave the conversation knowing enough for the salesperson to decide where to invest time.

The classic mistake is treating WhatsApp as a disguised form. The bot asks for name, email, phone and service of interest in sequence, the lead answers on autopilot or drops off, and the team receives a dead record. Good qualification feels like conversation. The AI asks one question at a time, reacts to what the person said and only moves on once it has understood the previous answer.

Before programming anything, write down on paper the profile of the ideal lead and of the lead you want to discard quickly. A high-ticket aesthetics studio, for example, wants to know the region, the procedure of interest and whether the person has already requested a quote. A law firm wants the type of case and the urgency. Without this definition, the AI will chat nicely and qualify nothing, which is the worst of both worlds: it costs money and generates no sales.

Start with the opening. The AI's first message should confirm the context the lead came from (ad, link in bio, referral) and ask an open question that reveals intent. Something like "what brought you here today?" yields far more than a numbered menu, because the answer already classifies the lead before the second message.

Then comes the discovery phase, and this is the core of qualification. Chain two to four questions that measure fit and urgency, always reacting to the previous one. If the lead says they want facial harmonization, the AI does not jump to price: it asks whether it is their first time, what the person would like to improve and by when. Each answer becomes a data point the salesperson reads later. The secret is for the AI to know when to stop asking, because a long interrogation scares people off just like a form does.

Build this flow in a visual builder instead of tying everything into code. With the Flow Builder you draw the paths, define the triggers for each answer and let the AI fill the natural language gaps, keeping control over what is a fixed rule (business hours, pricing policy) and what is open-ended conversation. This separates what needs to be deterministic from what benefits from the AI's flexibility.

Always keep a polite way to disqualify. Not every lead is a fit, and that is fine. If the person is outside your service area or is looking for something you do not offer, the AI should close respectfully and maybe point to an alternative, without taking up space in the salesperson's queue. Qualifying also means saying no quickly.

The quality of qualification lives in the AI's system prompt. That is where you describe the business, the tone of voice, the ideal lead profile, the mandatory questions and the things it must never do, such as promising a delivery date or closing a price outside the table. Treat this prompt as a training document for a new agent: the more specific it is and the more real examples it has, the better the result.

Give the AI context, not just instructions. It needs access to lead information that already exists (conversation history, which campaign they came from, whether they are already a customer) so it does not repeat questions or treat an existing customer as a stranger. In practice, this means passing the thread history and some metadata in each call to the model. An agent who asks your name for the third time loses the sale, and AI is no different.

Define clear decision limits. The AI qualifies, collects and schedules, but it should not give discounts, confirm real calendar availability without checking the system, or give opinions on sensitive cases. Make these boundaries explicit and create an escalation trigger for when the lead asks for something outside the scope. It is better for the AI to say "I will bring in a specialist for this" than to invent an answer that causes problems later.

Each instance can have its own AI, with its own prompt, tone and rules. If you serve a clinic and a real estate agency in the same dashboard, each number runs a different agent without mixing context. If you want to go deeper on configuring the agent per instance, see the guide on WhatsApp chatbots with AI, which details the options for behavior and humanization of the conversation.

For the bot to work, you need three pieces: a WhatsApp number reachable by API, a webhook that receives each new message and a service that decides the reply with the AI. When someone writes, WhatsApp delivers the event to your webhook, your backend builds the context, calls the language model and sends the reply back through the API. It is a simple cycle of receive, think, reply, repeated on every message.

With ZapZap API, which works through unofficial WhatsApp instances, you provision a number, connect it by scanning the QR code and start receiving events at your endpoint. The cost is R$15 per active number per month, prepaid, and there is a welcome credit when you create your account, subject to the current campaign. There is no volume contract and no template approval bureaucracy.

A webhook handler in Node with Express is enough to get started. The skeleton is straightforward: you read the event body, ignore messages sent by the number itself to avoid a loop, build the conversation history and call the AI. Something like the snippet below, illustrative, gives you the idea. You can check the exact route and field names at api.zapzapapi.com/llms.txt, which is the source of truth for the integration.

app.post('/webhook', async (req, res) => { const msg = req.body?.message; if (!msg || msg.fromMe) return res.sendStatus(200); const resposta = await gerarRespostaIA(msg.chatId, msg.text); await enviarTexto(msg.chatId, resposta); res.sendStatus(200); }); Note the care with fromMe: without this filter the AI replies to its own messages and enters a loop, one of the most common bugs in WhatsApp bots.

Qualification without a well-done handoff wastes the work. Define the exact handoff trigger: it can be the lead reaching a fit score, explicitly asking to talk to someone, or reaching the end of the discovery questions. At the moment of handoff, the AI should tell the lead that a human will take over and, at the same time, notify the team with a summary of what it found out.

The summary is the bot's most valuable deliverable. Instead of dumping the raw conversation on the salesperson, the AI condenses the essentials: who they are, what they want, signaled budget, urgency and suggested next step. The salesperson opens WhatsApp already knowing who they are talking to and where to start. This reduces human response time and raises the chance of conversion, because the contact arrives warm and with context.

Store the structured data somewhere you can query, whether your CRM, a spreadsheet or your own database. Every qualified lead should become a record with the information collected, so you can later measure which campaigns bring good leads and which just fill the queue. Without this data layer, you have a bot that chats well but produces no learning or sales predictability.

Use webhooks also to trigger automations at the end of the flow: schedule a reminder, fire a follow-up if the lead goes quiet, or push the contact into a nurturing sequence. The point is that the chatbot is not the end of the journey, it is the front door that decides where each lead goes, with the human team receiving only what is worth their time.

A new number that starts firing dozens of conversations per hour attracts the attention of WhatsApp's anti-spam systems. That is why warm-up matters: raising volume gradually, keeping two-way message exchanges and avoiding identical mass messages reduces the risk of the number being restricted. Treat the qualification number as an asset that needs a healthy routine, not as a blasting cannon.

ZapZap API offers community warm-up and anti-ban mechanisms precisely to bring the number closer to human behavior before scaling. This is not a promise of immunity, and no serious tool promises that. It is risk management: the more natural the usage pattern, the lower the chance of a problem. Spreading the operation across more than one number as volume grows also dilutes risk instead of concentrating everything on one line.

Pay attention to the AI's own cadence. Instant replies in milliseconds, with no variation, give automation away. Small delays, simulated typing and text variation make the conversation more believable to the lead and to the platform's filters. Humanizing is not just friendliness, it is also safer operation.

Finally, measure. Track reply rate, how many leads reach the handoff, how many convert and whether any number showed signs of restriction, then adjust questions and prompt with that data. Start with one number and a small flow: in ZapZap API the AI is configured per instance, so you can test your qualifier on one number, see what converts and only then replicate it. Scale what has proven it qualifies.