AI chatbots that answer in Arabic and English.

Whichever one they wrote in, unless you decide otherwise, and that decision belongs in the scope rather than in a model default. Gulf dialect, Modern Standard Arabic and Arabic typed in Latin letters are three different registers, and customers move between them and English inside a single sentence. We agree which varieties the assistant reads, which one it writes back in, and what it does with a mixed sentence, before anything is built. Modern Standard Arabic is the correct answer for published policy and government-facing wording, and the wrong one for a support thread.

By making every answer come out of a document you own, and by letting it refuse. The assistant retrieves from your policies, price lists and records, and each reply carries the source it was built from, so an answer you disagree with is corrected at the document rather than argued with. Where nothing covers the question, refusing is the correct output, and that path is tested as hard as the easy ones. A model with no retrieval behind it will answer anyway, fluently, which is the failure that ends a pilot.

It says so, and hands over with the thread intact. The person who picks it up sees what was asked, in which register, and what the assistant already tried, so the customer is never asked to start again. The escalation names who receives it and inside what window, agreed at scope, and unclear messages reach that person rather than a general queue, the same rule we apply to bilingual escalation across the rest of a Dubai operation.

Yes, from one build, answering from the same documents on both. What changes in this market is the escalation: WhatsApp voice and video calling is blocked under the TDRA policy while messages, documents and voice notes work normally, so a handoff inside the thread cannot end in a call from the thread. The written handoff therefore carries more weight here than it does elsewhere, which is why we test the escalation wording as carefully as the answers. On the website the same assistant runs as a widget, with the transcript held under the same consent.

A support transcript is personal data, so the federal PDPL applies unless your entity sits in the DIFC or the ADGM, which run their own regimes. The DIFC operates under Data Protection Law No. 5 of 2020 as amended in 2025, the ADGM under its 2021 regulations, and mainland UAE is on neither of their adequacy lists, so a group holding entities in two of them has two answers rather than one. Sending a transcript to a hosted model is a cross-border transfer, and the federal data protection regulator has published no adequacy list of its own, so it is answered on safeguards rather than looked up. We name the processing location in the contract and agree the retention window before launch.

What you were sold was probably a decision tree with a chat window on it, which understands only the words it was scripted for. The difference is not the model, it is what sits behind the model: retrieval over your own documents, a written rule for which register it replies in, and a refusal path that reaches a person. Where the bot you already have can be fixed, we say so and price the fix rather than a rebuild, the same posture we take on any automation purchase.

When the answers do not exist yet. If the questions your team fields are answered from judgement rather than from a document, the first deliverable is the document and not a bot, and we say so on the scoping call. We also decline when the volume is genuinely low and every query is one-off, when a product you already pay for covers it, and when nobody internally will own the knowledge base after handover. What the build costs follows the same three things: how many documents have to be trusted, how many surfaces it runs on, and whether it only answers or also acts. The tests a process has to pass before it is worth automating at all are set out on automate business workflows.