WhatsApp AI agents that do more than auto-reply

In most of the markets we work in — North Africa, the Gulf, Southern Europe — WhatsApp is not a channel, it is the channel. Customers message businesses there first and expect an answer in minutes. A basic auto-reply bot handles none of that. A real WhatsApp AI agent qualifies leads, answers from your actual product data, takes actions in your systems, and knows when to hand off to a human. Here is what building one involves.

Start with the Business API, not the app

The WhatsApp Business app on a phone is fine for a one-person shop. Anything automated needs the WhatsApp Business Platform (Cloud API), accessed directly through Meta or via a Business Solution Provider like Twilio, 360dialog, or Vonage. You get a verified business number, webhooks for incoming messages, and programmatic sending. Two hard constraints shape every design:

• The 24-hour session window: you can reply freely within 24 hours of the user's last message. Outside that window, you can only send pre-approved template messages.
• Template approval: outbound templates (order updates, re-engagement) must be approved by Meta, which takes from minutes to a couple of days, and utility/marketing templates are billed per conversation.

Pricing is per conversation, not per message, and varies by country and category. Budget roughly 0.01-0.08 USD per conversation depending on market and template category — meaningful at scale, trivial next to the cost of a missed lead.

Lead qualification that feels like a conversation

The highest-value use case we deploy is inbound lead qualification. A prospect messages 'how much for the 3-bedroom units?' and the agent does not dump a price list — it answers the question, then asks the two or three questions your sales team would: budget range, timeline, location preference. The LLM keeps this natural instead of form-like, and every answer lands in structured fields. We typically see qualification rates of 40-60% on inbound conversations, versus under 10% for 'leave your number and we will call you back' flows.

CRM handoff is where the money is

An agent that chats but never writes anything down is a toy. Every qualified lead should land in your CRM — Salesforce, HubSpot, Pipedrive — with the full transcript, the extracted fields, a one-paragraph LLM summary, and a lead score. The handoff code is boring webhook work, and it is what turns a chatbot into a sales channel. The same integration closes the loop: when a deal stage changes in the CRM, the agent can nudge the prospect on WhatsApp with a template message.

Human escalation: design it on day one

Every production agent needs explicit escalation paths. Ours trigger on three signals: the user asks for a human, the model's confidence drops (detected via repeated clarification loops or negative sentiment), or the conversation touches excluded topics — complaints, legal, refunds. Escalation routes the thread to a shared inbox with the full context and the LLM's summary, so the human starts informed. Aim for the agent to resolve 70-80% of conversations end-to-end; the goal is not zero humans, it is humans spending time only where they matter.

Multilingual is a feature, not an afterthought

Our clients' customers write in Arabic, French, English — often mixed, often in dialect. Modern LLMs handle this well, but your retrieval layer must too: if a customer asks in Tunisian Arabic and your product docs are in French, the agent needs embeddings that match across languages (multilingual-e5 and similar models do this reliably). Test with real dialect messages, not textbook translations — that is where naive setups break.

Limits to plan around

• Voice notes: customers send them constantly. Transcribe with Whisper before the LLM step, or you lose a third of your inbound messages.
• Rate limits and quality ratings: spammy outbound behavior gets your number restricted or banned. Outbound volume must be earned with engagement quality.
• Group chats and channels are largely off-limits to bots — design for 1:1 conversations.

We build these agents end to end at SLASH AI: Business API setup, the LLM layer grounded in your data, CRM integration, escalation workflows, and multilingual support out of the box. If WhatsApp is where your customers are, talk to us — we will scope what an agent can realistically take off your team's plate.