A WhatsApp chatbot in Darija: how AI can understand Moroccan customers, even voice notes

Moroccan customers do not write like a textbook. On WhatsApp, one message mixes Darija (Moroccan Arabic), French and Arabic, is written in Arabic script or in Latin letters with numbers ("3", "7", "9"), or arrives as a voice note. A chatbot built for standard French or English answers off the mark, and the customer ends up calling or giving up. This guide explains how to design a WhatsApp chatbot in Darija that genuinely understands your customers, and how to check that it does.

Why Darija is hard for AI

• No standard spelling: the same word is written several ways, in Arabic script and in Latin letters.
• Arabizi: Latin letters with numbers for certain sounds, such as "3endkom" (you have) or "7it" (because).
• Mixed languages in one sentence: "salam, wach livraison gratuite l Casa?".
• Regional variation in vocabulary.
• Voice notes, widely used, which need reliable transcription before anything can be understood.

Recent large language models often understand written Darija well, including in Latin letters, but quality varies with phrasing. It has to be measured on your own messages, not assumed.

How to build an agent that understands Darija

1. Start from your real messages

Collect a few hundred real (anonymised) messages from your WhatsApp: price questions, availability, delivery, complaints. They serve both to design the agent and to test it. This is the foundation of a serious project.

2. Choose and test the model

Test several models on your message set: intent understanding, extracting details (product, city, quantity), answer quality. The best model in English is not necessarily the best on your Darija.

3. Handle voice notes

A voice note must be transcribed by a speech recognition engine that handles dialectal Arabic, then treated as text. Test transcription on real customer voice notes (street noise, fast speech, mixed languages). If the transcription is uncertain, the agent should rephrase and ask for confirmation rather than guess.

4. Decide the reply language

Replying in the customer's language is usually most natural: Darija in Latin letters for an Arabizi message, Darija or Arabic in Arabic script, French for a French message. Set these rules with your team and keep a tone consistent with your brand.

5. Connect your information

The agent answers from your catalogue, prices, delivery zones and terms through retrieval (see RAG explained). It must never invent a price or a delivery time: if it cannot find the information, it hands off to a person.

6. Plan the human handoff

Complaints, negotiation, special cases, unhappy customers: the agent passes them to your team with a summary of the conversation, so whoever takes over does not have to reread everything.

WhatsApp: doing it by the rules

A chatbot must run through the WhatsApp Business Platform (the official API), not the standard WhatsApp Business app driven by an unauthorised tool. Unauthorised automation leads to restrictions and then bans, covered in WhatsApp Business banned. Also secure your account.

Use cases that work best

• Retail and e-commerce: prices, availability, sizes, delivery, order tracking, cash-on-delivery confirmation.
• Local services: bookings for salons, clinics and garages. See AI agents for clinics.
• Real estate: qualifying enquiries (city, budget, property type) before passing them to an agent.
• Customer service: common questions and complaint follow-up. See deploying an AI customer service agent.

Common mistakes

• Testing only with correct French sentences written by the team.
• Letting the agent invent prices or delivery times when information is missing.
• Ignoring voice notes, which a large share of customers use.
• Replying in very formal Classical Arabic to a customer writing casual Darija.
• No visible human fallback.

We build agents that speak Darija

Our AI agents service designs WhatsApp agents in French, Arabic and Darija, connected to your catalogue and tools, tested on your real messages and deployed through the official API. See also AI WhatsApp chatbot for business. Send us a few sample customer messages and we will show you how an agent would answer them.