Generative AI and LLM Integration

Every Gulf company now has an AI chatbot story, and most of them end the same way. The bot launched on the website, it handled the easy questions, and then it told a customer that a discontinued package was still available at last year's price. Someone screenshotted it. The bot was switched off "for maintenance" and never came back. The problem was not the model. It was a knowledge base nobody cleaned, no way to measure accuracy before launch, and no rule for what the bot should do when it did not know.

Language is the second trap. Most off-the-shelf assistants are built in English and translated, and it shows. They answer a relaxed Gulf Arabic message in stiff, formal Arabic, or misread a mixed Arabic and English voice note completely. In Saudi Arabia in particular, buyers expect an assistant that sounds like a capable local operator, and one that does not is not deployable, however good its English. Choosing the wrong model here is expensive, because you usually discover it after the integration work is done.

The third cost is regulatory, and it arrives late. UAE PDPL is in force, and Saudi PDPL has been fully enforceable since September 2024 with fines of up to SAR 5 million (GRC Vantage). An assistant that sends customer records to a model hosted outside the region, with no agreement on retention or training, is precisely the kind of processing those laws ask about. Answering that question in the design phase costs a few decisions. Answering it after launch costs a rebuild.

There is also the quieter cost of the assistant nobody uses. Staff try it twice, get a vague answer that cites nothing, and go back to asking the colleague who has been there longest. Adoption is not a training problem. It is an accuracy and trust problem, and it is solved in the build, not in the launch email.