Skip to content
CaseworkPublished · Updated · 4 min read

The Article 27 FRIA for hiring AI: what to include, what to leave out

Article 27 of the EU AI Act introduced the Fundamental Rights Impact Assessment. It is mandatory for public bodies and public-service providers, and voluntary for most private employers, who use it to evidence their Article 26 duties. This post walks through what a FRIA for an AI hiring system should contain, and where deployers tend to over- or under-write.

Key takeaways

  • Article 27(1) requires a FRIA only from bodies governed by public law, private entities providing public services, and deployers of credit-scoring and life- and health-insurance systems. Most private employers are not in scope.
  • A public hospital, university or council using hiring AI is in scope. Private operators in public-service sectors such as education, healthcare, social services and housing should check (Recital 96).
  • For other employers a FRIA is voluntary, but Article 27(1)(a)–(f) is the best available structure for the rights analysis behind the Article 26 duties: human oversight, monitoring, and worker and candidate notice.
  • The structure we use for a hiring FRIA has seven parts: the deployer's context, the system and the candidate's journey, who is affected, the risks to fundamental rights, mitigations, oversight, and complaint and redress.
  • Since the Digital Omnibus, Article 27(4) lets a FRIA cross-reference the GDPR DPIA instead of repeating it.
  • Leave out speculative harms the deployer has no realistic ability to address. Four concrete risks per right beat twenty abstract ones.

The DPIA was a culture shock when GDPR introduced it in 2018. Privacy teams had to learn a new format. Engineering had to provide inputs they did not previously document. Legal had to sign off on something that looked like a checklist but felt like a position paper.

The FRIA is the same shock, with two differences. First, the substance is broader — fundamental rights, not just privacy. Second, no one has done this before. There is no settled industry template (Article 27(5) asks the AI Office to produce a questionnaire), and no list of "30 questions you must answer." The operative list in Article 27(1) is six short points.

Who has to do one

Article 27(1) requires a FRIA from three kinds of deployer: bodies governed by public law, private entities providing public services, and deployers of credit-scoring and life- and health-insurance systems (Annex III, points 5(b) and (c)). A public hospital, university or council using hiring AI is in scope. Recital 96 names education, healthcare, social services, housing and the administration of justice as public services private entities may provide, so private operators in those sectors should check. Most private employers are not in scope.

We write one for them anyway. Article 27(1)(a)–(f) is the best available structure for the rights analysis behind the Article 26 duties every hiring deployer has: human oversight (26(2)), monitoring (26(5)), worker notice (26(7)) and candidate notice (26(11)). It also sits alongside the GDPR DPIA: since the Digital Omnibus, Article 27(4) lets a FRIA cross-reference the DPIA instead of repeating it.

This is the structure we use for the FRIA on a hiring AI deployment.

§1. The deployer’s context

Who is the deployer. What is its size and sector. What is its overall hiring posture. This is short — half a page — but it bounds the analysis. A FRIA for a 600-person SaaS company is not the same document as a FRIA for a national retailer hiring 5,000 associates per year.

§2. The system’s purpose and the candidate’s journey

What does the system do. Where does it sit in the hiring workflow. What does a candidate experience. This section is where you describe the human-in-the-loop posture: who reviews the system’s outputs, with what criteria, at what point in the funnel. If the system’s output is final — that is, no human reviews it — you note this and you cross- reference the higher level of justification required.

§3. Categories of natural persons likely to be affected

Article 27 calls this out explicitly. For hiring, the set is wider than people sometimes assume:

  • Candidates who interact with the system.
  • Candidates who do not interact with the system because of how it filters at earlier stages.
  • Current employees whose data trained the system or whose data is referenced for fit.
  • Workers’ representatives who must be informed.

We include each category. The "candidates who do not interact" line is often the one that lands the FRIA at the most useful place — the system shapes the pipeline before any candidate sees a job posting.

§4. Risks to fundamental rights

This is the substance. We organize it around the Charter of Fundamental Rights of the European Union categories that are most likely to be at risk in a hiring deployment:

  • Article 21 — non-discrimination.
  • Article 8 — protection of personal data.
  • Article 31 — fair and just working conditions (for employees affected by the system).
  • Article 47 — right to an effective remedy (does the candidate have recourse).

For each, we describe the specific risk this system creates, the evidence we have for it, and the severity. We do not pad the list with abstract risks. Better to have four concrete risks per right than 20 abstract ones.

§5. Mitigation measures

For each risk, the mitigation. Some are technical (input filtering, output thresholding, monitoring). Some are organizational (oversight roles, training, escalation paths). Some are contractual (what your agreement with the vendor lets you do, and what it requires the vendor to do).

§6. Oversight measures

Article 14 obliges the provider to design for human oversight. Article 26(2) obliges every deployer to staff it with people who have the competence, training and authority to act, and Article 27(1)(e) asks the FRIA to describe how that oversight is implemented. Document the oversight that actually happens in deployment. Who oversees. What they look at. How often. What they are trained on. What they can do when they see a problem.

§7. Complaint and redress mechanisms

How does an affected person — a candidate — surface a complaint. To whom. With what response SLA. With what right of appeal. With what record-keeping.

The Colorado AI Act requires this explicitly. The EU AI Act expects it. A FRIA without it reads as incomplete.


What to leave out: speculative harms that the deployer has no realistic ability to address. The FRIA is not an academic exercise; it is a record that the deployer thought about the risks that apply to this deployment and put measures in place. Throwing in every theoretical risk dilutes the actually-load-bearing risks.

The FRIA sits alongside the technical documentation file, which we walk through section by section in a companion post on Annex IV. The Pack produces both on the same four-week methodology; the SaaS EU-readiness study shows what that looked like on a real engagement shape.

The hardest part of writing a first FRIA is reading the Charter. The hardest part of writing the second is realizing that the first one was too long.

Frequently asked questions

Is a FRIA mandatory for private employers using AI in hiring?

Usually not. Article 27(1) of the EU AI Act requires a fundamental rights impact assessment from bodies governed by public law, private entities providing public services, and deployers of credit-scoring and life- and health-insurance systems. Most private employers are outside those categories, though private operators in public-service sectors such as education, healthcare, social services or housing should check.

Why write a FRIA if it is not required?

Article 27(1)(a)–(f) is the best available structure for the rights analysis behind the Article 26 duties every hiring deployer has: human oversight (26(2)), monitoring (26(5)), worker notice (26(7)) and candidate notice (26(11)). It also sits alongside the GDPR DPIA.

What should a FRIA for an AI hiring system include?

The deployer's context; the system's purpose and the candidate's journey, including who reviews its outputs; the categories of people likely to be affected; the risks to their fundamental rights; mitigation measures; the human oversight that actually happens in deployment; and complaint and redress mechanisms for candidates.

Which fundamental rights does a hiring FRIA usually assess?

The Charter of Fundamental Rights articles most likely to be at risk in a hiring deployment: Article 21 (non-discrimination), Article 8 (protection of personal data), Article 31 (fair and just working conditions, for employees affected by the system) and Article 47 (the right to an effective remedy).

Who counts as affected by an AI hiring system?

Candidates who interact with the system; candidates who never do, because it filters them out at an earlier stage; current employees whose data trained the system or is referenced for fit; and workers' representatives who must be informed.

Does a FRIA replace the GDPR DPIA?

No, the two sit alongside each other. Since the Digital Omnibus, Article 27(4) lets a FRIA cross-reference the DPIA instead of repeating it.

Begin engagement

Need this documented for your own hiring AI?

The Pack turns the analysis into a deployer-side file — conformity documentation, FRIA, risk management, and monitoring plan — for one tool, in four weeks.

  1. Today

    You send the intake

    Four short sections, about fifteen minutes. A questionnaire, not a sales call — your draft saves as you go.

  2. Within one business day

    We reply personally

    With a scoped quote, a question or two, or a candid pointer elsewhere if we are not the right fit.

  3. Four weeks after kickoff

    The file is in your hands

    Discovery, drafting, review, handoff. Fixed scope and price from $35,000, with one round of revisions within 30 days.