Half of my client conversations in the UAE start with the same question: "what does the PDPL actually require for our AI plans?" The honest answer in 2026 is that the principles are clear and in force, while a lot of the operational detail is still waiting. That is an awkward place for an architect, but it is a workable one.
The usual caveat applies. I'm an architect, not a lawyer, and the regulatory position here moves. Use this to shape engineering decisions, and confirm the legal position with UAE counsel.
Three regimes, one company
The federal Personal Data Protection Law, Federal Decree-Law No. 45 of 2021, has applied since 2 January 202212. It does not cover everything. Free zones with their own data protection laws, notably DIFC and ADGM, are carved out, and some categories such as health data are also shaped by sector laws2.
So a group with an onshore entity, a DIFC entity and a healthcare line of business can sit under several regimes at once. That matters more for platform design than any single rule.
| Regime | Where it applies | What it means for AI design |
|---|---|---|
| Federal PDPL | Onshore UAE, and data about people in the UAE | Principles, rights and automated-decision protections apply now; transfer detail reportedly still pending |
| DIFC Data Protection Law | Entities in the DIFC free zone | Regulation 10 adds specific duties for autonomous and semi-autonomous systems |
| ADGM Data Protection Regulations | Entities in Abu Dhabi Global Market | Separate regime; map obligations per entity rather than assuming federal rules |
The gap: principles now, operational rules later
The PDPL anticipated Executive Regulations to fill in much of the operational detail. Several 2026 commentaries report that, as of this year, those regulations had still not been published34. The same commentaries note that items such as the list of adequate destinations and the approval routes for cross-border transfers depend on them4.
For an architect, "the rules are not finished" is not a reason to wait. It is a reason to build in ways that can absorb whichever rules arrive, without a rebuild.
DIFC already has rules aimed at AI
DIFC moved earlier. Regulation 10 of its data protection regulations covers personal data processed through autonomous and semi-autonomous systems, which includes AI6. Commentary on it highlights a few points that translate directly into engineering work56:
- Most duties sit with the deployer, the organisation using the system, rather than the vendor operating it.
- Clear notice about the technology and how it may affect people's privacy rights.
- Human-defined purposes. Systems used for high-risk processing should process personal data only for purposes humans have defined or approved.
- An accountable person. For high-risk processing, commentators describe an autonomous systems officer role, similar in spirit to a data protection officer.
Even outside DIFC, I treat Regulation 10 as a useful design brief. It describes the controls that most AI regulations are converging on, including the EU AI Act, which I cover in the EU AI Act after the Digital Omnibus.
The architecture I use
The aim is one platform that can satisfy the strictest regime you operate under, with configuration rather than code changes when rules land.
- Tag data by jurisdiction and category at ingestion. Every record carries which entity collected it, under which regime, and whether it is personal, sensitive or health data. Everything downstream depends on this.
- Region-pinned storage by default. Keep personal data in UAE regions unless a transfer has been approved. Make cross-border movement an explicit, logged operation, not a side effect of a backup or an API call.
- An AI gateway in front of every model. Authentication, PII redaction, routing by data sensitivity and a full audit log in one place. Sensitive workloads go to models running locally; the rest can go to hosted models. I describe the pattern in my notes on private LLMs.
- Purpose binding. Each AI use case is registered with its purpose, data categories and owner. The gateway refuses calls that don't match a registered purpose.
- Human review for consequential outputs. Where a model's output affects a person, record the decision, the inputs and the reviewer.
- Policy as code. Encode residency and access rules in versioned policy, so that new Executive Regulations become a policy change and a review, not a project.
Where agents raise the stakes
Agents that call tools can move data in ways a chatbot cannot: reading from one system and writing to another, possibly across borders. The controls I set out in securing MCP servers apply directly here. Per-tool scopes, audience-bound tokens and a gateway log of every tool call give you the evidence a regulator, or an auditor in DIFC, will ask for.
What I would do this quarter
- Map each legal entity to its regime and list where personal data physically lives today.
- Inventory every AI use case with its purpose, data categories and model location.
- Put the gateway and audit logging in before the next AI feature ships, not after.
- Agree with counsel which interim transfer safeguards you will rely on while the federal detail is pending.
How I can help
I work with teams across the UAE and the US on exactly this: AI and data platforms that hold up under several regimes at once. See my UAE fractional CTO and AI advisory page, or analyse your challenge and bring the result to a call.
- 1Federal Decree by Law No. 45 of 2021 Concerning the Protection of Personal DataUAE Legislation portal
- 2New data protection law introduced in the UAE: 10 key takeawaysHerbert Smith Freehills · 2021
- 3
- 4
- 5
- 6