Security in enterprise AI is a shared responsibility: the platform secures the platform, and the configuration is yours. That distinction decides who is accountable when a model is given access to data it should not have seen. This pillar covers access control and data handling when deploying on managed services, and what UAE businesses should consider when governing AI systems and personal data. It is operational guidance, not legal advice.
Managed services can support a secure deployment, and several document enterprise-grade controls for exactly that purpose. What none of them do is make an organisation compliant on its own: security on such platforms is a shared responsibility, and the controls inside your own implementation stay yours to set. Whether your configuration has a lawful basis, whether your retention is defensible and whether your cross-border transfers have a basis are assessed against what you built.
Treat an unqualified claim that a product is compliant as a reason for caution rather than comfort. No vendor can offer that, and the ones who do are describing a certification of their platform, not of your deployment.
A group with entities in a free zone and on the mainland can be subject to several data protection regimes simultaneously: DIFC and ADGM each operate their own law with their own regulator, separate from the federal framework. Sectoral rules sit on top — health data relating to services provided in the UAE carries its own restriction on being processed outside the country.
The practical first step is a document most organisations have never produced: every legal entity, the regime that governs it, and where its data physically sits. Every later decision about where inference runs and what is logged depends on it. None of this is legal advice; validate it with qualified advisers.
Security is a property of how a deployment is configured rather than something the technology supplies. Managed services describe enterprise-grade access to foundation models and configurable controls, but security on them is a shared responsibility and your own implementation decisions remain yours. The questions that decide the answer are concrete: where inference runs, what data leaves your environment, whether your data trains anyone's model, how long prompt and output logs are kept, and who can read them.
Wherever the deployment was configured to process it, which is why this is an architecture question rather than a detail to settle before launch. If processing happens outside the UAE and touches personal data, that is a cross-border transfer requiring a lawful basis. The reach is wider than the database: it covers model inference, prompt and output logs, backups, sub-processors, and whether a vendor's support staff abroad can access any of it.
If the conversation contained personal data, the logs do too. Prompt and output logs are the most frequently overlooked personal data store in an AI deployment: they are routinely retained longer than the source system, stored in a different region, readable by more people, and rarely covered by anyone's retention policy. Set a deliberate retention period, enforce deletion, and establish who has access — including at your vendor.
No. A platform's certifications describe the platform. Whether your configuration has a lawful basis, whether your retention is defensible and whether your transfers have a basis are assessed against your implementation. Managed services shift work; they do not transfer responsibility. Ask what the certification actually covers, and have your own privacy, security and legal stakeholders assess the design against your obligations.
A list of entities and the regime governing each; a decision on where processing happens; access scoped per action rather than per system, so an agent cannot reach data it has no business reading; logging of what the system did and on whose request, in a form an auditor can read; a retention period for prompt and output logs; and a human approval point on anything that sends, pays, deletes or commits. These are cheap to decide at design time and expensive to retrofit.