WhatsApp AI agents: qualify leads 24/7

A WhatsApp AI agent handles the chat like a well-trained junior salesperson: it asks, filters, books and, when it should, closes. It is not the 2018 bot with three buttons and “I didn’t understand that”. It is a model with instructions, your business data and a rule for handing over to a human.

If your team loses leads after 19:00, this is not a marketing extra. It is coverage.

How it differs from a classic chatbot

The classic chatbot routes: press 1 for sales, 2 for support. It breaks as soon as the customer writes “I want last year’s one but in white”.

The AI agent:

• Reads natural language.
• Asks qualification questions (budget, timing, area, model, intake).
• Can check stock, hours or prices if you connect it.
• Hands the chat to a human with a summary, not 40 loose messages.

ManyContacts has three profiles: Victor (qualification), Ana (support) and Dani (sales). You decide how far each one goes.

What it should qualify (and what it should not)

Train the agent with a short script. Example for an academy:

1. Course or intake of interest.
2. Whether they already have a prior qualification.
3. Timing (this month / this quarter).
4. Budget, or whether they need finance.

Example for a dealership:

1. Model and budget.
2. Cash or finance.
3. Part-exchange.
4. When they can come for a test drive.

What it must not improvise: unauthorised discounts, medical data, legal promises, stock it has not checked. There, it transfers.

Where the agent lives: in the inbox, not on an island

If the AI is at another vendor and the team is somewhere else, the salesperson does not trust it and turns it off. The agent has to write in the same shared inbox. The human sees the transcript, takes control and continues.

That also avoids the classic “the bot said 12,900 and I said 13,400”.

24-hour window and AI

AI can reply in free text while the customer wrote less than 24 hours ago. If the thread is cold, you need a template. Design the agent to:

• Answer instantly when the lead comes in.
• Use a re-engagement template if the lead vanished for three days.
• Not fire marketing at anyone who did not opt in.

More detail in 24-hour window.

Metrics that matter (not “number of bots”)

• % of chats resolved without a human (appointments, FAQs).
• % of qualified leads passed to sales.
• Time to first response (should be seconds).
• Transfer rate for “the customer asked for a person” (if it spikes, the script is weak).
• Conversions (enrolment, visit, order), not only messages.

An agent that chats a lot and books nothing is a cost. At €0.20/conversation, ROI shows in a week if you measure appointments.

How to launch it without a three-month project

1. Pick one flow: qualify Meta Ads leads, or recover carts, or book an appointment.
2. Write the goal in one sentence: “Book a visit with name, phone and time slot”.
3. List 10 questions and the answers it cannot invent.
4. Connect data (stock, LMS, calendar) if the flow needs it.
5. Start in night hours or on 20% of traffic.
6. Review 20 real conversations and adjust.

ManyContacts includes an agent configurator so you do not start from a blank page.

Usual fears

“It will sound like a robot.” Tone is editable. What sounds like a robot is the numeric menu, not a short, concrete message.

“It will invent the price.” If it does not have the source, ban it in the prompt and force the handoff.

“Salespeople will complain.” They complain if you take the hot lead away. They do not complain if it arrives qualified at 09:00 with the appointment already set.

“Meta does not allow bots.” Meta allows automation on the API. What it does not allow is spam. The agent replies to people who wrote to you, or who have a template and permission.

AI + humans: a clean split

• AI: greeting, FAQ, qualification, reminder, cart, appointment confirmation.
• Human: negotiation, exceptions, VIP customers, serious complaints, high-ticket close.

That split is what we use in sales teams and in automotive. The kanban still belongs to people; AI fills the first column.

A minimum script that actually works

Do not start with a 20-page manual. Start with this, pasted into the configurator:

• You are the first contact for [brand]. Goal: [appointment / enrolment / qualify].
• Ask only what is on the list. If a data point is missing, insist once.
• If they ask for a price that is not in the source, transfer.
• If there are insults or a threat, transfer and do not argue.
• Close with a three-line summary for the salesperson.

Then paste 15 good real conversations (from your best advisors) as examples. The AI copies the team’s tone, not an American playbook.

When to turn AI off (and that is fine)

• The number has low quality after a campaign: better humans and utility.
• Launch of a product that is not yet in the knowledge base.
• A crisis (mass delay, recall, leak). AI is not an emergency community team.

Turning it off is not failure. It is not letting a model improvise at an expensive moment.

Next step

Do not design “the bot for the whole company”. Design one agent for one flow. Create the account or book a demo and we have it qualifying this week.