An Arabic AI chatbot that understands Gulf dialects the way your customers actually write them

The phrase “supports Arabic” is printed on very different products. Some mean the interface displays right to left. Some mean it understands a carefully written question in standard Arabic. Few mean what your customers need: that it understands a message typed with one thumb in a car.

So we split support into layers and ask about each one separately. A bot that passes the first and second and fails the third will fail with most of your customers, because nobody writes to customer service in standard Arabic.

A customer writes a greeting, then asks the price of the monthly plan and whether there is a discount for taking a year, all in dialect, without a single hamza, with the final ta written as ha and two questions in one message. Another writes half in English: do you have delivery to Dammam, and what are the fees. A third sends a ten-second voice note with the street in the background.

Modern models understand such messages well, and that is what has changed in recent years. But “well” is not “always”. A rare local word, an everyday name for a product, a two-word message with no context: these are the places where the digital employee needs what we add, namely your business’s vocabulary, examples from your customers’ messages, and a rule that tells it when to ask.

Understanding is one thing and replying is another. The digital employee understands colloquial Arabic, but replying in it is not always wise. A dialect imitated imperfectly sounds forced, and more than one dialect can end up mixed in a single sentence.

So we advise most companies to use polite, plain Arabic: close to how people speak, free of stiffness, and understood by Saudis, other Gulf nationals and Arab residents alike. For those who want a tone nearer their audience’s dialect, we write the greeting, apology and sign-off phrases in their own wording and fix them as set text, so the style stays consistent. When a customer writes in English, it replies in English.

Any bot looks clever in a demo where the vendor types the questions. The honest test is to give it your customers’ messages as they arrived. We ask you for a sample from your chat archive, remove the personal data from it and sort it by difficulty.

Then we run the digital employee on it and go through the result with you message by message: what it understood correctly, what it got wrong and what it handed to a person. We fix the errors and test again, and we do not move to launch until you are satisfied with the result. After launch we read samples of real conversations weekly during the first month.

No system understands all of Arabic. Accuracy drops with dialects further from the common training data, with very local vocabulary, with Arabic written in Latin letters and numerals, and with sarcasm and hints. With voice notes it drops with noise, fast speech and overlapping voices.

Names, addresses and order numbers are known weak points, which is why the digital employee repeats them back to the customer for confirmation before relying on them. What it has not understood after one clarifying question, it hands to a person. The aim is not that it understands everything. It is that it knows when it has not understood.

The demo page opens on a fictional sample business. Write to it the way you would write to a friend, not the way you would write a formal letter: no hamzas, English words mid-sentence, two questions in one message. Then call it and speak in your own dialect. Two minutes are enough to see where it stands.

Yours is built on your company’s information and tested on your customers’ messages. Messages are kept with defined access in line with the Saudi Personal Data Protection Law, and follow-up or offer messages on WhatsApp go only to people who agreed, as templates Meta has approved. The number, the accounts and the code belong to your company.