Arabic AI Chatbot for Customer Service: A Practical Guide

An Arabic AI chatbot for customer service works when it understands how your customers really write, answers only from approved company information, and passes the chat to a person, with the history, when it cannot help. Judge it by the problems it solves, not by the chats that never reached an agent. This guide covers WhatsApp and web assistants, dialects, human handoff and the measures that show whether the project is working.

Where does an Arabic AI chatbot for customer service fit on WhatsApp and the web?

The two channels behave differently. A web chat starts on your site, sometimes with a customer who is logged in and looking at one order or one page. A WhatsApp chat starts in an app the customer already uses all day, and the thread can stay there for weeks.

Build one assistant with one set of approved answers and one set of handoff rules, then adapt the channel details. Our Agentic AI & Chat/Voice Agents service (TT2) covers chat and voice on web, WhatsApp, Teams and phone. The support and voice page describes the use case in more detail.

What does Meta require for a business chatbot on WhatsApp?

These points come from Meta's own pages, read on 18 October 2026. The messaging policy and the platform terms both show a last-updated date of 23 September 2026, so check the live text before you build.

• Opt-in comes first. Meta's WhatsApp Business messaging policy lets a business contact a person only after that person has given opt-in permission, and it recommends separate opt-in for each message category.
• Automation is allowed, with a route to a person. The same policy says you may use automation when responding during the 24-hour window, and requires prompt, clear and direct escalation paths. The options it lists include a transfer to a human agent in the chat, phone, email and web support.
• The 24-hour window sets what you can send. Meta's Cloud API guide says the window starts when a user messages or calls you and resets each time they contact you again. Inside it you can send free-form replies. Once it closes, you can send only pre-approved templates. All messages, inside or outside the window, go only to people who opted in.
• Opt-outs must be honoured. The policy requires you to respect any request, on or off WhatsApp, to block, stop or opt out.
• There is a clause for AI providers. Section 4.7 of the Meta Terms for WhatsApp Business Platform prohibits providers and developers of AI or machine learning technologies from using the platform to make those technologies available when they are the primary, rather than incidental or ancillary, functionality. The clause is written about AI providers distributing AI itself. The clause does not name companies that use an assistant to serve their own customers, but Meta decides how a particular vendor's product is classified. Ask your vendor in writing how its WhatsApp set-up complies, and confirm with counsel.

How should an Arabic AI chatbot handle dialects?

Customers can write in Modern Standard Arabic, Gulf, Levantine, Egyptian or Maghrebi dialects, in Arabic letters or in Arabizi (Arabic typed in Latin letters and numbers), and they can mix in English words in one message. An assistant that expects formal Arabic can misread some of them.

Plan for it in four steps:

1. Collect real past conversations from your own channels, with personal details removed and with approval from your data owner.
2. Build a test set for each variety your customers use, including Arabizi, plus the hard cases: misspellings, voice notes transcribed to text, and messages that mix two languages.
3. Reply in simple, clear Arabic that a customer from any country can follow. Match the customer's register in tone, but keep policy statements, prices, dates and conditions in plain, consistent wording.
4. Repeat the test after every change to the model, the prompt or the knowledge base. Treat a drop in one dialect as a defect, even if the average looks fine.

Containment or resolution: which number should you track?

Containment is the share of chats that end without a human. Resolution is the share where the customer's problem was actually solved. They are different. A bot that says "I cannot help with that" and closes the chat counts as contained and can leave the customer angry. A bot that hides the "talk to a person" option can raise containment while making service worse.

Agree the definitions before launch, set your own targets from a baseline of your current service, and report resolution next to containment every time.

How do you design the handoff to a human?

1. Define the triggers: the customer asks for a person, a complaint, a payment dispute, a legal, medical or safety topic, two failed attempts, or no approved source for the answer.
2. Tell customers early that they are talking to an automated assistant, and how to reach a person.
3. Send the agent a short summary: who the customer is, the issue, what the bot already tried and which sources it used.
4. Let the agent reply in the same thread, and stop the bot from answering while the agent is active.
5. Outside working hours, state when a person will reply and collect a contact detail. Never leave the customer in silence.
6. Return the chat to the bot only when the agent releases it.

Where should answers come from, and what about customer data?

Answers should come from approved documents, with the source shown to your team, not from the model's memory. Our Custom AI Applications & Modernisation service (TT3) covers retrieval platforms and the data foundations behind them. If the assistant can act, for example change a booking or start a refund, set limits and approvals for each action, which is part of TT2.

Chats contain personal data, so check which law applies. SDAIA's data governance portal says the Saudi Personal Data Protection Law (Royal Decree No. M/19 of 9/2/1443H, as amended) applies to processing the personal data of citizens and residents of the Kingdom, including by entities outside it, and lists a separate regulation on transfers outside the Kingdom.

In the UAE, the UAE government portal describes Federal Decree-Law No. 45 of 2021 as covering processing inside or outside the country, and its Article 2 excludes government data, health and banking or credit data that have their own laws, and free zones with their own data protection law, such as the DIFC (DIFC Law No. 5 of 2020) and ADGM. The portal also lists Federal Law No. 2 of 2019 on ICT in health fields. A cloud region does not settle these questions on its own: a UAE region is outside Saudi Arabia, and sector regulators and cybersecurity or cloud rules can add hosting requirements. This is general information, not legal advice, and your obligations depend on your sector and your data.

Launch checklist and how to start

• Opt-in wording and records are in place for each message category.
• The assistant says it is automated and offers a person at any point.
• Handoff triggers, summary format and out-of-hours replies are tested.
• A test set covers every dialect and Arabizi, and runs after every change.
• Answers come from approved documents with named owners.
• Containment and resolution are defined, with a baseline.
• Chat logs are mapped: where stored, who sees them, how long they are kept.
• Your vendor has answered the WhatsApp terms question in writing.

If you are not sure which of these is your gap, start with the free AI Exposure Scan: a free 45-minute conversation that ends in a written next step. If you already know you want a support assistant, the Transformation Lab is where TT2 and TT3 sit.

Related lab: Transformation Lab ↗