The review that cost us nine days did not find a vulnerability. It found a promise. Our privacy notice said users could ask us to delete their data, and our system could delete a row in one table while leaving the same person present in a warehouse export, a search index and a vendor’s retention window. Nobody had lied. Nobody had been careless in the ordinary sense. We had simply hired engineers to build a product and assumed that the data protection part was a document somebody else owned. This is the method I use now, written as the thing I wish had existed before that review.
Step 1: Establish Which Regime Binds You, Before Writing the Description
This sounds like a legal question and it is actually a recruitment question, because the answer determines which five years of a candidate’s experience count. The UAE does not have one data protection regime. It has a federal law, Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data, overseen by the UAE Data Office established under the companion decree-law. Alongside it, the financial free zones run their own regimes: the DIFC Data Protection Law No. 5 of 2020 and the ADGM Data Protection Regulations 2021, each with its own commissioner, its own notification duties and its own transfer conditions. Health and financial services add sector rules on top.
So the first artefact of this hire is not a job description. It is one paragraph, written down, that says which of those applies to your entity, whether you are a controller or a processor in each product line, and where your data physically sits. If you cannot write that paragraph, you are not ready to interview, because every answer a candidate gives will be unfalsifiable. We cover the regime landscape in our guide to UAE data protection, and the overlap with security obligations in the UAE cybersecurity law guide.
One caveat worth stating plainly rather than burying: the detailed implementing regulations under the federal law have taken longer to settle than most of us expected when it was issued, so the practical compliance position is something to verify with current advice rather than assume from an article. That uncertainty is an argument for hiring engineering judgement over regulatory recall, which is the theme of everything below.
Our expert take
The number of job descriptions we see that list a regime the company is not subject to is remarkable, and it is not harmless. Asking for free zone experience when you are a mainland entity, or the reverse, filters out the engineers who would have been productive and attracts the ones who pattern match on keywords. Write the paragraph first. It takes twenty minutes with whoever signed your last customer contract, and it changes the shortlist more than any sourcing channel will.
Step 2: Split the Three Jobs Employers Merge Into One
Almost every requisition I am asked to help with has three jobs inside it. The first is privacy counsel: interpreting obligations, owning the records of processing, handling a regulator. The second is a security engineer: boundaries, encryption, intrusion detection, incident response. The third is the actual data privacy engineer: making the commitments true inside a running system.
They overlap and they are not interchangeable, and the market punishes the merged version twice. You will not find many people who are credible in all three, and the few who are command a salary justified only when personal data is your product. Worse, the merged role has no clear success condition, so the hire spends their first quarter discovering that three different stakeholders expected three different people.
Write the engineering role on its own, and name explicitly what you are buying elsewhere. In most companies that means a data privacy engineer on staff, legal interpretation bought as advisory hours, and security staying with the platform team. If your situation is closer to governance of automated decisions than to data plumbing, the adjacent method is our guide to hiring an AI compliance engineer in Dubai, which is a genuinely different screen.
Step 3: Screen Data Mapping, Not Policy Recall
The interview failure mode here is asking what the law requires. Candidates who prepare can answer that from a summary page, and the answer tells you nothing about whether they can change your system. Ask instead for a map.
The question I use: pick a product you worked on, choose one piece of personal data, an email address or a phone number, and trace everywhere it ended up. Not the design. What was actually true in production when you left.
Strong candidates do something specific here: they go somewhere uncomfortable without being prompted. They mention the application logs that captured request bodies for a fortnight after a debugging session. The analytics warehouse where somebody built a table keyed on email because it was convenient. The customer support tool nobody listed as a processor. The CSV a colleague exported to a laptop in 2024. The staging environment seeded from a production dump.
Weak candidates describe the architecture diagram. The difference is not intelligence, it is whether they have been through an exercise where the diagram was proved wrong. Follow up with the question that confirms it: how did you find the copies you did not know about. Answers involving log sampling, schema scanning, vendor inventories and interviewing engineers are the ones you want. Answers that assume the documentation is correct are a declined offer.
Step 4: Test Whether Deletion Actually Propagates
This is the self marking question, and if you run only one thing from this article, run this. Hand the candidate a small schema on a whiteboard or a shared document: a users table, an events table with a user reference, a nightly export into an analytics warehouse, a search index, a cache, and one third party processor such as a support desk or an email provider. Then ask: a user asks to be deleted, walk me through what happens.
You are listening for a sequence, not a command. The strong answer covers the primary record, the referencing rows and whether they are deleted or anonymised, the warehouse copies and whether history can be rewritten at all, the search index and its reindex lag, cached derivatives with their own expiry, the processor and its contractual retention, backups and the honest statement that backups usually cannot be surgically edited so the policy has to be expiry based, and finally what you are entitled or obliged to keep for accounting or legal reasons, because deletion requests are not unlimited.
A panel with no privacy background can score this. An answer that stops at a SQL statement has never met the problem. An answer that volunteers the backup difficulty before you raise it has met it repeatedly, and that is the signal.
Our expert take
The reason this question works is that it is impossible to answer well from reading. Every element of the strong answer is a scar: the reindex lag because somebody once deleted a user who remained searchable for six hours, the backup caveat because somebody had to explain to a customer why thirty days is the honest answer. We have watched candidates with impressive certifications stop at the first table, and candidates with no certification at all walk the whole chain unprompted. In this discipline, production experience is not merely preferable to credentials, it is the only thing that predicts anything.
Step 5: Check That Consent Is Modelled as Data
Consent is where most systems are quietly indefensible, because it is usually implemented as a boolean on a user row that is overwritten whenever preferences change. That design cannot answer the only question that ever gets asked in anger: what exactly did this person agree to, on what date, under which version of the notice, and through which interface.
The screen is to ask for that directly. How would you prove what a specific user consented to eighteen months ago. A candidate who has done this reaches for an append only record with a timestamp, a scope, a notice version and a provenance, because the current state is a projection of history rather than the truth. They will also mention withdrawal as a first class event rather than a flag flip, and they will raise the awkward downstream point: when consent is withdrawn, the processing that already happened does not unhappen, so your pipeline needs to know which derived datasets are now tainted.
Ask one more: what about data collected on a basis other than consent. Someone who treats consent as the only lawful basis will over-collect consent and under-document everything else, which produces a system that is both annoying to users and weaker in a review.
Step 6: Set a Take-Home Shaped Like a Real Request
Keep it to two hours, pay for it if your policy allows, and make it a document rather than code. The brief: here is a simplified schema and a list of our processors. A user has submitted a subject access request. Produce the plan for responding inside the statutory window, list what you would need from whom, and mark every place where you believe this system cannot currently comply.
The last clause is the whole exercise. You are not assessing whether they can produce a complete answer from a schema they saw this morning. You are assessing whether they will tell you what is broken. Candidates who return a confident plan with no gaps have either not understood the system or will not raise problems once hired, and both of those are the same risk wearing different clothes.
Score on three axes: did they find the copies that are not in the obvious tables, did they distinguish between what the law requires and what is merely good practice, and did they flag the thing that would have embarrassed you. In our experience the strongest submissions are shorter than the weakest ones and contain more question marks.
| Step | What you are testing | Strong signal | Decline signal |
|---|---|---|---|
| 1. Regime | Your readiness, not theirs | You can state it in a paragraph | Description lists every regime |
| 2. Role split | Scope discipline | One job, named trade-offs | Counsel plus security plus engineer |
| 3. Data mapping | Production reality | Volunteers logs, exports, vendors | Recites the architecture diagram |
| 4. Deletion | Propagation thinking | Raises backups unprompted | Stops at a SQL delete |
| 5. Consent | Modelling as history | Append only, versioned notice | Boolean on the user row |
| 6. Take-home | Willingness to flag gaps | Short, with marked unknowns | Confident and gapless |
| 7. Audit | Whether the role has authority | An interview answer changed | Nothing moved in 90 days |
Not sure this needs a full time hire
Send us the shape of your data and the regime you operate under, and we will tell you honestly whether this is a specialist role or a senior engineer with advisory support.
Start the conversationStep 7: Audit the First 90 Days Against Your Own Screen
Re-run the interview. Literally: take the deletion question you asked and ask it again about your own system, ninety days after the hire started. If the answer has not changed, one of two things is true and they need different fixes.
Either the screen was wrong, and you hired someone who interviews better than they build, which is unpleasant and recoverable if you act at ninety days rather than at the annual review. Or, far more commonly, the role has no authority. The hire has written an excellent assessment that nobody has scheduled, because deletion propagation competes with a feature release and loses every sprint. That is not a hiring failure, it is an organisational one, and the person you just hired cannot fix it from where they sit.
The preventative measure is cheap and almost nobody does it: before the hire starts, name the first two things they are expected to change and put them in a sprint. Not a policy, not an audit. Two concrete system changes with a date. If you cannot name two, that is useful information about whether this role is real at your company yet.
Where This Method Stops
Three limits, stated plainly, because a method that claims to cover everything is a sales document.
- This is not legal advice and the screen is not a compliance assessment. Every question above tests engineering capability. Whether your processing is lawful is a question for a qualified adviser who has read your contracts, and no hire removes that need.
- Most companies do not need this role yet. The threshold where a full time specialist pays for itself is roughly: personal data is central to the product, or you operate under more than one regime, or a regulated customer audits you. Below it, a senior backend engineer with audit trail experience and a few advisory days a month is the better answer, and it is easier to retain because a specialist with eight weeks of real work will leave.
- The screen tests the system you have, not the one you are building. If your data model is still changing weekly, a privacy engineer will spend their quarter re-mapping. Sequence this hire after the data model settles, which is the same reason we argue against hiring a model specialist before a data layer is trustworthy.
If the compliance surface rather than the hire is what you are sizing, our guide to building a compliance platform in the UAE covers the build side of the same problem, and the UAE data protection guide covers the ground rules a regulated buyer will ask you to evidence.
Frequently Asked Questions
Is a data privacy engineer the same thing as a data protection officer?
Do I need someone with specific UAE experience, or will a GDPR background transfer?
What does this role cost compared with a senior backend engineer?
How do I test this skill if nobody on my interview panel has it?
Run the screen on real candidates
We help UAE employers scope this role honestly and screen for the constraints a data protection review actually tests, including telling you when a specialist is not the right hire.
Get startedRelated reading: UAE data protection · Building a compliance platform · UAE cybersecurity law
