ai:residency
AI data residency in the UAE: where your prompts are processed.
By Loan Laux · November 9, 2026 · 9 min read
AI data residency in the UAE is usually asked as a storage question: which region is the resource in. For hosted models the sharper question is where the prompt is processed. The text travels to wherever the model runs and sits in its memory for the request, and that location is set by the deployment type on Azure, or by the model ID or inference profile each Bedrock request names. A UAE resource can still send inference out of the country, and logs and retrieval indexes add locations of their own.
This post is the company-side map: which rules apply onshore, in DIFC and ADGM, and to health data; the in-country options; the settings that move processing elsewhere; and a four-question check per workflow. We are not lawyers and this is not legal advice: we describe instruments and settings, and the legal reading stays with counsel. Government entities have a different picture, covered near the end.
Why AI data residency in the UAE is a processing question.
Residency is usually framed as storage: which region holds the database. A prompt travels to wherever the model runs, sits in the model's working memory for the request, and the response comes back. If that compute is outside the UAE, personal data in the prompt has been processed abroad, whatever the storage region says.
The documentation draws the line plainly: on Azure, data at rest stays in the resource's designated geography for every deployment type, and where inference happens is a property of the deployment type. Two side paths follow the prompt as well. Invocation logs record requests and responses, sometimes in full, in whichever region they are kept, and retrieval indexes hold chunks of source documents as embeddings, a second copy of sensitive content in a second system.
Which rules apply: onshore, DIFC, ADGM and health data.
The federal instrument is Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data (PDPL), in force since 2 January 2022. Article 2 excludes three groups that matter here: government data and government entities; personal health data that has its own legislation; and companies in free zones with their own data protection legislation. DIFC and ADGM both have such legislation.
DIFC adds a layer aimed squarely at this topic. Its Data Protection Regulations, in force on 1 September 2023, address personal data processed through autonomous and semi-autonomous systems in Regulation 10, with obligations for the deployers and operators of those systems. ADGM's DPR 2021, enacted 11 February 2021, keeps its transfer rules in Part 5, in a GDPR-aligned structure.
We describe these instruments; we do not interpret them for you. Which regime binds an entity, which Regulation 10 role it holds, and whether a transfer mechanism extends to model inference, belongs with counsel. The instruments in one view:
| Instrument | Who it covers | What it says about transfers |
|---|---|---|
| PDPL, Federal Decree-Law No. 45 of 2021 | Onshore personal data processing | Article 22 permits transfers where a proper protection level exists; Article 23 lists alternatives such as contracts with equivalent safeguards |
| DIFC Law No. 5 of 2020 | Entities in the Dubai International Financial Centre | Article 26 permits transfers to jurisdictions the Commissioner finds adequate; Article 27 sets alternatives including the DIFC standard contractual clauses |
| ADGM DPR 2021 | Entities in the Abu Dhabi Global Market | Part 5 governs transfers out of ADGM: adequacy decisions, appropriate safeguards, binding corporate rules, derogations |
| Federal Law No. 2 of 2019 | ICT in health fields, including free zones | Article 13: health data from services provided in the UAE may not be stored, processed, generated or transferred abroad except under a Health Authority resolution with the Ministry |
The in-country options.
As of November 2026, Azure and AWS both offer hosted models from UAE regions, and this post covers those two. Azure has two UAE regions: UAE North, in Dubai, with availability zones, and UAE Central, in Abu Dhabi, with access restricted to specific customer scenarios such as disaster recovery.
On AWS, Amazon Bedrock has been available in the Middle East (UAE) Region since 29 September 2025. A Bedrock setup there keeps inference in the country for models offered In-Region (each model's Regional availability table says which), with one caveat about inference profiles in the next section.
Neither option settles the question: the region decides where the resource and its data at rest live, and the settings decide where the prompt goes.
The settings that send prompts elsewhere.
Azure model deployments carry a deployment type, and it decides processing location. As of November 2026, data at rest stays in the resource's designated Azure geography for every type. Inference differs: Global types may be processed in any Azure region; Data Zone types process within the US, EU or Asia Pacific zone, and there is no Middle East data zone; Standard and Regional Provisioned types process within the deployment's own geography.
Two resources, both in UAE North, can sit one menu choice apart: a Global Standard deployment and a Standard deployment of the same model give different residency answers. The type shows in the deployment's SKU name, such as GlobalStandard.
Bedrock has the same shape of trap. Invoking a model by its in-Region ID in the Middle East (UAE) Region keeps the request there. Cross-Region inference profiles can route requests to other AWS Regions by design. As of November 2026, geographic profiles route within a geography such as the US or EU, and global profiles may route to any supported commercial Region worldwide. The destinations depend on the Region you call from, and AWS tells teams with residency requirements to check every destination Region with GetInferenceProfile rather than trust the profile's prefix.
The side paths have settings of their own. Bedrock's model invocation logging is off by default and, once on, writes only to CloudWatch Logs or S3 in the same account and Region. Application logs, tracing tools and prompt history are set up separately and can sit anywhere. Retrieval indexes inherit the location of the vector store.
A four-question residency check for each AI workflow.
The check runs per workflow rather than per company: a summarizer on published data and a drafting assistant on patient queries get different answers.
01:
Which data does the workflow touch
Personal data, health data, government data, or published material. The class decides which instrument applies; health data changes the conversation entirely.
02:
Which regime does the entity sit in
Onshore, DIFC or ADGM. Where the entity is registered tells counsel which instrument's transfer rules to read.
03:
Which deployment setting is the model on
On Azure, the deployment type: Standard keeps processing in the geography, Global and Data Zone can process outside it. On Bedrock, an in-Region model ID or a cross-Region inference profile.
04:
Where do logs and indexes live
Invocation logs, prompt history, vector indexes and their backups. Each has a region and gets the same question as the prompt.
When government data is involved.
Everything above is the company picture. Article 2 of the PDPL excludes government data and government entities from the federal law, so public-sector governance comes from emirate-level frameworks and each entity's classification scheme.
We have covered that side at length: public sector AI governance for the residency-versus-sovereignty distinction, AI adoption in the UAE public sector for the data classification gates, and AI for UAE government for day-to-day operation. Workflows mixing company and government data need both classifications settled before anything reaches a model.
Do it yourself, or bring us in.
The four-question check is doable in-house: an afternoon with the consoles open, a list of workflows, and honest answers about what each one touches. Look in particular for Global deployments left over from a proof of concept, and for logs or traces kept in a region nobody chose.
Where we come in is the part after the check: moving the settings without breaking live workloads, matching deployment types to data classes, and writing the result down so the next review is a reading exercise. As in our AI consulting in the UAE engagements, we leave the legal reading to your counsel and build the systems to match it. An audit takes two weeks and ends in a written, sequenced roadmap that fits your security review, procurement and data-residency rules.
Frequently asked questions.
Does the PDPL apply to companies in DIFC and ADGM?
Article 2 of Federal Decree-Law No. 45 of 2021 excludes companies and establishments in free zones that have their own personal data protection legislation. DIFC has DIFC Law No. 5 of 2020 and its Data Protection Regulations; ADGM has the ADGM DPR 2021. Whether a given entity falls within that exclusion is for counsel to confirm.
Is a resource in a UAE region enough?
It fixes where data at rest lives and leaves processing open. On Azure, Global and Data Zone types can process prompts outside the UAE; Standard and Regional Provisioned types process within the geography. On Bedrock, an in-Region model ID keeps the request in the Region, while cross-Region inference profiles can route it to other Regions.
Can health data go to a model hosted abroad?
Article 13 of Federal Law No. 2 of 2019 provides that health data from services provided inside the UAE may not be stored, processed, generated or transferred outside the country, except where the Health Authority issues a resolution in coordination with the Ministry. Whether a resolution covers a given workflow is a question for health-sector counsel.
What should we ask a vendor about residency?
Four things, in writing: where inference runs for our deployment, named by type or profile; where invocation logs and prompt history land by default and how to change that; which regions and subprocessors can see the data under the standard contract; and what happens to the data when the contract ends.