Most enterprise AI projects don't stall on technology. They stall when legal, risk and IT can't agree on who is accountable for what an AI system recommends.

This guide sets out a practical governance approach for AI agents that process personal data, based on the principles of the UAE Personal Data Protection Law (PDPL, Federal Decree-Law No. 45 of 2021). Free zones such as DIFC and ADGM have their own data protection rules, and sector regulators may add more. It is not legal advice; work through it with your data protection officer and legal advisers.

Why governance comes first

Treat governance as the design brief, not the final hurdle. When the controls are clear before a pilot starts, teams move faster because every reviewer knows what good looks like and every approver knows what they are signing.

The question isn't whether AI can decide. It's who signs off, and what they saw when they did.

What data protection law asks of AI

Data protection law applies to AI the same way it applies to any other processing of personal data. Its principles are familiar: a lawful basis, a clear purpose, only the data you need, accuracy, limited retention, security, and the rights of the people the data is about.

What changes with AI is scale and traceability. An agent may read thousands of documents containing personal data in a day, and combine information from several systems. That makes three questions harder to answer:

  • Was this use of personal data covered by the notice we gave?
  • Who could see the data, including through the agent's outputs?
  • How long is the data kept, including in logs and memory?

Good governance makes those questions easy to answer for every agent.

Five controls to put in place

1. An inventory of AI use. List every agent and AI tool in use, what it does, which systems and data it touches, and who owns it. You cannot govern what you cannot see.

2. A named owner for each agent. A business owner, not only an IT owner, who is accountable for the agent's purpose, performance and the decisions it supports.

3. A review step for consequential outputs. Decide which outputs affect individuals or the business enough to need human approval, and route them to people with the right authority. Make the evidence visible to reviewers.

4. Logging that captures inputs and decisions. Record what data the agent used, what it produced, which model and prompt versions were involved, who reviewed it and what they decided. Apply your retention rules to these logs too.

5. A regular performance review with the business. Compare agent outputs with reviewer decisions, track where they disagree, and decide whether the agent, the procedure or the training needs to change.

Human review that holds up

A review step only protects you if the reviewer can genuinely challenge the output. That means three things in practice: they can see the sources behind each finding, they have time to check, and they can change or reject the output without friction. Record their reasons. Over time, those reasons show you where the agent needs work.

Data residency and access

Decide early where each agent runs and where its data and logs are stored. Some organisations require their own cloud tenancy or in-country hosting for agents that process personal data. Apply role-based access so the agent can only retrieve what the user asking is allowed to see, and make sure outputs don't leak data across those boundaries.

Check that your privacy notices cover the purposes for which agents process personal data. Where agents analyse customer feedback or communications, mask personal identifiers that aren't needed. Where you link a person's enquiry to their website activity, say so clearly and give them a choice.

Your 90-day plan

  • Days 1–30: build the inventory, name owners, and agree which outputs need human review.
  • Days 31–60: set up logging and retention for one pilot agent, and run it in shadow mode.
  • Days 61–90: hold the first performance review with the business, report to your risk committee, and decide whether to go live.

Download the full checklist for the control set, policy templates and a board reporting pack.

GovernanceData protectionRisk