Arabic AI Customer Service: Dialect In, an Approved Answer, a Human Handoff
31 Oct 2026
An Arabic reply helps when it understands an Egyptian dialect question, answers only from a list the company approved, and hands the thread to a person when the question leaves that list. That is separate from the WhatsApp bot article, which explains the channel and the platform.
A customer in Egypt writes the way they speak: short, colloquial, sometimes a voice note or a two-word question. A system that answers with a generic formal paragraph, or that invents a price and a date, turns service into a new problem. Arabic AI customer service, in this article, is three constraints rather than a slogan: understand the question as it arrived, answer only from approved text, and hand the thread to a person together with what was already said.
The approved list is a document the company owns, not a document the model owns. It holds services, branches, opening hours, what may be said about price when the price is fixed and written down, and what must not be said. Any question with no line on that list is not answered by a guess. The customer is told a colleague will continue, and a task opens for the person on duty. That is the gap between a tool that protects a promise and a tool that promises what management cannot perform.
Dialect in the question, approved wording in the answer
Do not ask the customer to rewrite the question in formal Arabic before the system understands it. Two casual phrasings of "when do you open" are one question if opening hours are on the list. The answer leaves in the wording a responsible person approved, not in a style the model picks on each turn. Changing tone is an edit on the list, not a surprise in the thread. If you want a colloquial answer, write it yourselves on the list. Do not leave wording as a daily experiment on customers.
- Collect the questions that repeated this month, in the customer wording.
- Put one approved answer on each question, with the name of who approved it.
- Mark what is forbidden: a discount, a clinical time, a delivery promise.
- Name who receives the thread when no answer exists.
- Review each week what was handed over because it left the list.
Handoff is part of the design, not a decorative safety net
Handoff means the staff member sees the question, the answer already given if any, the channel, and the time. They start where the system stopped, not with a fresh greeting that forces the customer to retell the story. If earlier messages vanish, the handoff is cosmetic. Set the human reply window you will actually keep, and put it in the approved answer: a colleague will continue during working hours. Do not write a minute you will not meet.
Outside working hours the system either confirms receipt and promises a follow-up on the next shift, or it stays quiet if you do not own that sentence. Silence is better than a night promise with nobody behind it. The sentence itself is approved with the rest of the list, so the night policy and the afternoon policy do not drift apart.
This is not the WhatsApp bot article
The WhatsApp bot article explains the gap between the business app and the official platform, and what it takes to run a bot on that channel. This article does not repeat that. The subject here is language and operating policy: dialect on the way in, an approved answer on the way out, and a person in the middle when the list runs out. The channel may be a website, a mailbox, or WhatsApp. Channel rules live in the other article. If your project is specifically a bot on WhatsApp, read that after the list is stable, not before.
The only AI delivery this page will cite
The AI work ENS points to is menu import inside ENSMENU. A photo or a PDF goes in, dish lines come out, and a person checks them before a guest sees the menu. That is a menu-reading task. It is not a customer-service bot, and this article does not claim a reply bot was delivered for a client. It is mentioned so the boundary stays honest: a closed task with a human check is one thing, and an open conversation with no approved list is another.
What fails in the first week
Launch fails when an "about us" page is renamed into an answer list. The customer asks about a branch, a delay, and an invoice, and the page talks about vision. It fails when the model is allowed to improvise a price in order to close the thread. It fails when no staff name sits on the handoff queue, so threads pile up and the model is blamed for a missing shift. Repair those three before any talk of accuracy.
What we ask for before a scope is estimated
- Twenty real questions from the last two weeks, verbatim.
- The approved answer for each, or a mark that it goes to a person.
- Working hours and who covers handoff inside them.
- The channel the question arrives on today.
- What must not be said even if the customer insists.
Questions before a build
Will the system handle every dialect?
Start with the Egyptian colloquial your customers actually write, and test with the lines you collected rather than with laboratory sentences. Widening dialect is a later decision after the list holds, not a promise on day one.
May it offer a discount?
Only when the discount is an approved line with its conditions. Otherwise the thread is handed over. A commercial discount is a human decision on a permission, not a sentence the model produces to please the asker.
Who updates the list after launch?
A person you name, with a review time. A list with no owner lags behind reality, and the system then states an old fact with confidence. The name of the list owner is a line in the agreement, not a detail for later.
The WhatsApp bot rules themselves are in the WhatsApp chatbot guide. Building this kind of reply starts from AI solutions.
Send a sample of questions and one approved answer through start your project. The WhatsApp button at the foot of the page can carry the customer wording together with this article name.
