Using AI with customer data in the UAE, and who says yes
The federal data protection law has been in force since January 2022. Its implementing regulations have still not been issued, and in June 2026 the office that was meant to enforce it was folded into a new authority. That gap is the whole engineering decision.
A client wants to put a model in front of their support queue, or their CRM, or three years of order history. The first question they ask is whether that is allowed under UAE law, and they expect a yes or a no.
The accurate answer is more useful than either. There is a federal law, it has applied for nearly four years, and the document that was supposed to tell you how to comply with it does not exist yet. That is not a technicality you can wait out, because you have to pick an architecture now and the cost of guessing wrong is a migration rather than a fine.
What is actually in force
The UAE government's own page on data protection laws names the instrument: Federal Decree Law No. 45 of 2021 Regarding the Protection of Personal Data, known as the Personal Data Protection Law. It came into force on 2 January 2022.
Two sentences on that page do most of the work for an engineering decision. The first is scope: the law applies to the processing of personal data, whether in full or part through electronic systems, inside or outside the country. Hosting abroad does not put you outside it. The second is the default basis: the page says the law prohibits the processing of personal data without the consent of its owner, with exceptions where processing is needed for a public interest or to carry out legal procedures and rights.
Consent as the default is a stronger starting position than many teams assume, and it is the sentence to design against. A model that reads a customer record is processing it. If your consent language covers service delivery and nothing else, sending that record to a third party model is not obviously covered by it.
The part that is missing, and why it is the whole problem
The same government page says the law sets out the requirements for the cross border transfer and sharing of personal data for processing purposes, and then says nothing further. That is the sentence everybody needs and it is a pointer to a document that has not been published.
A June 2026 note from Morgan Lewis on the new Federal Authority for Artificial Intelligence and Data states it directly: the implementing regulations to the law have still not been issued, and no supervisory authority has been unambiguously designated for private sector oversight.
The same note explains why that is unlikely to resolve itself quickly. On 14 June 2026 the UAE announced the Federal Authority for Artificial Intelligence and Data, consolidating the Artificial Intelligence Office, the Information and Digital Government Sector within the telecoms regulator, and the Emirates Data Office, which the note says never became fully operational in practice. The authority reports to the Cabinet and its mandate includes setting unified national AI and data policy and establishing standards for data and AI management. The market expects it to finish the regulations. The note is explicit that it is unclear whether that will happen quickly or whether AI and digital government work will take priority first.
What that means for where you put the data
Normally a cross border transfer is a compliance exercise: you check whether the destination is on an adequacy list, or you sign the approved contractual clauses, or you file for a regulator's approval. In the UAE right now, the adequacy list, the approved clauses and the filing procedure are all in the document that has not been published, and the body that would process a filing has just been reorganised.
So the practical rule is not a legal opinion, it is a risk ordering. The further your customer data travels from a region you control, the more you are relying on a mechanism whose procedure nobody can show you.
| Where the model runs | What you are relying on | What changes if regulations land |
|---|---|---|
| In your own UAE region, self hosted | No transfer to justify | Very little |
| Vendor region inside the UAE | Your contract plus the vendor's residency claim | You may need to re-paper the contract |
| Vendor region abroad, personal data sent | An unpublished transfer mechanism | Potentially a migration |
| Vendor abroad, data de-identified first | Your de-identification holding up | Depends on whether it is personal data at all |
The fourth row is where most of the useful engineering sits, and it is also the row people are most confident about when they should not be. Stripping a name and an email is not de-identification if the remaining fields single a person out, and a support transcript routinely singles a person out in the body of the message. If the plan is to send abroad only what is no longer personal data, that claim has to survive someone reading the payload, not just the schema.
Sector rules sit on top, and they are not optional
Two carve outs catch teams out. The federal law is not the whole map: the government page lists the DIFC's own Data Protection Law, DIFC Law No. 5 of 2020, as a separate instrument, and the ADGM has its own regime too. If your client is licensed in a financial free zone, the federal analysis above is the wrong analysis. Which licence they hold is not always obvious from the outside either, and BDG Labs has the entity side of that in local agent vs branch office in the UAE.
Health data is the other one. The government page notes that the health ICT law regulates the use of information technology in health care, including its free zones. If there is patient data in the pipeline, treat the question as closed until someone with a health regulator mandate opens it.
A decision order that works without the regulations
- Classify first. Decide which fields are personal data, which are special category, and which are genuinely not about a person. Do this on real sample payloads, not on the table definition.
- Check the licence. Mainland, DIFC or ADGM changes which law you are reading before anything else.
- Read your own consent text. If it describes service delivery only, a third party model is a new purpose and your copy needs to say so.
- Default to a region you control for anything personal, and treat an abroad default as a decision requiring a reason.
- If you de-identify in order to send abroad, write down the test you applied and have someone try to re-identify a sample.
- Keep a transfer register even though nobody has asked for one. When the procedure is published, the question will be what you were already doing.
The honest limits of this note
This is an engineering read, not legal advice, and the gap it describes is exactly the kind of gap that a lawyer is paid to take a position on. Two things in particular are worth saying plainly.
First, secondary commentary on whether the implementing regulations exist is genuinely inconsistent. Several published guides name specific cabinet decisions as the regulations. The government's own page does not mention executive regulations at all, and the June 2026 legal note says they have not been issued. We have followed the government page and the dated legal note, and if you need certainty on that point, it is a question for counsel in the UAE rather than for a vendor.
Second, nothing here says do not use a model on customer data. It says the transfer question is the one with no published answer, so architecture that avoids needing an answer is cheaper than architecture that depends on one. For the neighbouring jurisdiction, where the confident answer people quote was removed from the document they quote it from, the same pattern plays out in data residency requirements in Saudi Arabia.
And if the actual goal is a support assistant rather than a data platform, the scope question comes before the jurisdiction question. What a model can be trusted to answer is covered in how to stop an AI chatbot giving wrong answers, the operational version is in what AI agents can do for business operations, and whether the thing is worth building at all is measuring the ROI of AI automation.
Describe it. We build it.
Seven or twelve days, pay on delivery, a year of maintenance included. Bring the problem, not a spec.
Book a meetingRead next
How to load test a web app before launch, in five runs
How to load test a web app before launch: the five test types k6 documents, which one to run first, and the threshold that decides whether you passed.
ReadA phased software rollout plan, and the order that holds
A phased software rollout plan borrowed from reliability practice: why one branch goes first, the few signals worth watching, and keeping a way back.
Read