AI in HR: what is safe to use and what to avoid
AI can take hours of drafting and repeated questions off an HR team. It can also quietly screen out good candidates or make a pay mistake nobody notices. The line is simple: let it prepare, and let people decide.
The short answer
Use AI first for drafting, translating and answering routine policy questions from approved documents. Treat anything that affects someone’s job, pay or career as higher risk: a person must make that decision and be able to explain it.
Do not use a language model to calculate payroll. Payroll needs fixed rules that give the same correct answer every time.
Where each use sits
| Risk | Use | How to handle it |
|---|---|---|
| Lower | Drafting job posts, letters and policies | A person edits and approves it |
| Lower | Translating between Arabic and English | Have someone check legal and pay terms |
| Lower | Answering employees from the approved handbook | Link the source, and hand off personal cases |
| Higher | Summarising or ranking CVs | Only as help. Someone reviews rejected candidates too |
| Higher | Summarising performance data | The manager decides, and the employee can respond |
| Higher | Flagging attendance or payroll anomalies | A flag starts a check. It is never a penalty on its own |
| Avoid | Reading emotions from video interviews | Do not use it. It is already banned at work in the EU |
| Avoid | Automatic dismissal, discipline or pay decisions | Keep these with people |
Why hiring needs extra care
A model learns from patterns, including unfair ones. It may favour certain universities, names or career paths, or mark down career breaks. Arabic CVs and transliterated names can be read less accurately than English ones, so test in both languages.
Regularly compare who the tool passes and rejects across groups, and read a sample of rejected CVs yourself. If nobody ever overrides the tool, your human review may not be real.
What the rules say
Employee and candidate records are personal data, so data protection law applies in the UAE, Saudi Arabia and Egypt. Under Article 18 of the UAE’s PDPL, people can object to automated decisions with legal or serious effects and ask for human review.
If you hire or manage people in the EU, most AI for recruitment, performance, promotion and dismissal counts as high-risk under the EU AI Act. Those duties now start on 2 December 2027, and the ban on emotion recognition at work already applies.
Questions to ask any HR AI vendor
- What exactly does the AI decide, and what does a person decide?
- Can we see why a candidate or employee received a given result?
- Has it been tested for bias, and on Arabic data?
- Where is employee data stored, and is it used for training?
- Can we turn the AI off and still use the product?
- What logs do we get if someone challenges a decision?
A safe way to start
- Clean up records first: AI cannot fix wrong hire dates or missing contracts
- Pick one low-risk task, such as answering handbook questions
- Tell employees what is being used, and why
- Name an HR owner who checks a sample each month
- Write down what AI may never decide in your company
Sources
- Morgan Lewis: EU AI Act deadline changes for employers and HR technology (June 2026)
- Lewis Silkin: the EU Digital Omnibus on AI enters into force (July 2026)
- Lewis Silkin: prohibited practices under the EU AI Act, including emotion recognition at work
- OB Gulf: AI law in the UAE, including PDPL Article 18
Rules change. Check the official text, or ask your accountant or lawyer, before you rely on a figure.
Get your HR data in order first
keyt HR keeps employee records, attendance, leave and Egyptian payroll in one place, in Arabic and English, with approvals and locked payroll runs.
keyt HR: HR, attendance and payrollSee it working