How AI Is Used in HR: The Tools and the Legal Line

Most professions adopting AI face a quality question. HR faces a legal one first, and the difference shapes everything.

Three things make it distinctive.

The decisions are about people, and they determine access to work, pay and advancement. Employment law across most jurisdictions already regulates them heavily, independently of any AI rule.

Employment and worker management is one of the eight areas listed in Annex III of the EU AI Act, so systems used for recruitment, selection, promotion, termination, task allocation based on behaviour or traits, and performance monitoring fall in the high-risk tier.

And the affected person is usually outside the organisation, or in a position of dependency inside it. A rejected applicant is not a customer, has no relationship to protect, and often never learns why. This is the structural asymmetry the responsible AI cursus describes, in its sharpest form.

The consequence is that the sequence differs from other professions. Elsewhere you ask whether a tool works and then whether it is compliant. In HR the legal analysis comes first, because a tool that works well and sits in a prohibited category cannot be used at all, and one in the high-risk tier carries obligations that change the deployment entirely.

This lesson covers that map. The next two cover the workflows and the practice.

Before any risk analysis, two practices are banned outright under Article 5, and both reach ordinary employers who would not consider themselves to be doing anything exotic.

Emotion inference in the workplace. AI systems to infer emotions of a natural person in the areas of workplace and education institutions are prohibited, outside narrow medical or safety purposes. This catches more than it appears to: sentiment analysis applied to employee communications, tools claiming to read engagement or stress from video during interviews or meetings, and voice analysis inferring mood or confidence. Vendors do not always describe these as emotion recognition, so the question to ask is what the system claims to infer rather than what it is called.

Biometric categorisation to infer sensitive attributes. Systems categorising people on biometric data to deduce race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation are prohibited.

Two notes on what this means practically.

These have applied since 2 February 2025, so this is not preparation for a future deadline. An organisation running emotion inference on staff today is already outside the law.

And the prohibited tier carries the highest penalty band, up to 35 million euros or 7 percent of worldwide annual turnover, which is the figure people misattribute to everything else in the Act.

Check the estate against these two before anything else in this cursus.

Mapping HR uses onto the tiers, which determines the obligations for each.

Prohibited: emotion inference on staff or candidates, and biometric categorisation for sensitive attributes.

High-risk under Annex III: recruitment and selection, including targeted job advertising and filtering applications; decisions on promotion and termination; task allocation based on behaviour or personal traits; and monitoring and evaluating performance.

Not high-risk, generally: scheduling on availability, payroll processing, benefits administration, drafting non-decisional communications, and answering policy questions from documentation.

The boundaries that get argued. Targeted job advertising is explicitly inside, which surprises marketing teams who assumed recruitment risk began at application. Task allocation on availability and skills is outside; the same tool allocating on behaviour or traits is inside. And a tool summarising an interview is different from one scoring it.

Annex III high-risk obligations apply from 2 December 2027 after the 2026 Omnibus deferral, but the prohibitions apply now and Article 50 transparency lands on 2 August 2026.

Across the employee lifecycle, the current capability, stated without the marketing.

Sourcing and advertising. Drafting job descriptions, and targeting where adverts appear. The drafting is genuinely useful and low-risk. The targeting is inside Annex III and is where advertising platforms have already produced documented discrimination problems.

Screening. Parsing applications into structured fields, and ranking or scoring candidates. The parsing is mechanical and defensible. The ranking is the high-risk core of the function and the subject of most of this cursus.

Interviewing. Scheduling and transcription are ordinary automation. Scoring interview responses, or analysing video or voice, is where the prohibitions and the high-risk tier both bite.

Onboarding. Answering policy questions from documentation, generating personalised checklists, drafting communications. Mostly low-risk, and a good early deployment.

Performance. Summarising feedback and drafting reviews is assistive; scoring or ranking employees is high-risk under monitoring and evaluating performance.

And people analytics. Attrition prediction is the case people underestimate, because a model predicting who will leave is used to make decisions about those people, which puts it close to or inside the high-risk boundary depending on what happens with the output.

The pattern: description and drafting are safe, and anything that scores, ranks or predicts about an individual is not.

Article 6(3) lets a listed system out of the high-risk tier where it does not pose a significant risk of harm, including by not materially influencing the outcome of decision-making. Vendors invoke it frequently for HR tools, and it usually does not hold.

The reason is the profiling limit. A system that performs profiling of natural persons is always high-risk regardless of which condition it might otherwise satisfy. Profiling means evaluating personal aspects of an individual, which is precisely what a candidate-fit score does.

The second obstacle is material influence. A ranking system determining which twenty of four hundred applications a recruiter reads has materially influenced the outcome, because the three hundred and eighty never seen were decided upon by the model. The human decided among survivors.

What can genuinely rely on the derogation: a parser extracting structured fields from a CV so a human reads them consistently, provided it does not score, rank or filter. That is a narrow procedural task and it does not profile.

What cannot: anything producing a fit score, a shortlist, a ranking, or a flag that changes how an application is treated.

So when a vendor says their tool is not high-risk, ask which condition they rely on and how they address the profiling limit. A vendor who cannot answer has not done the analysis, and under the procurement cursus that is diagnostic well beyond compliance.

The AI Act's high-risk obligations arrive in December 2027. Non-discrimination law applies now, and it applied before anyone deployed a model.

The framing that matters: in most jurisdictions, a selection process producing disparate outcomes for a protected group can be unlawful regardless of intent and regardless of what produced it. A model is not a defence, and in several jurisdictions the employer remains liable for a vendor's tool because the employer made the decision.

Two doctrines to know by name.

Direct discrimination is treating someone less favourably because of a protected characteristic. A model using a protected attribute as a feature is the obvious case, and it is rare because it is obvious.

Indirect discrimination, or disparate impact, is a neutral practice that disadvantages a protected group and cannot be objectively justified. This is the one that applies to almost every AI screening tool, because proxies reconstruct protected attributes as the bias cursus explains, and the disparity appears in outcomes without any protected attribute in the features.

The practical consequence is that the fairness measurement in the bias cursus is not a compliance nicety for 2027. It is how you find out whether your current process is lawful today.

And the burden in many jurisdictions shifts once a claimant shows disparate outcomes: the employer must then justify the practice. Without measurement you cannot, because you do not know what your own outcomes are.

The most important technical point in this cursus, and it applies to every hiring model regardless of vendor.

A hiring model does not predict who would perform well. It cannot, because that data does not exist for people who were never hired. It predicts a proxy, and the proxies available are all records of past human decisions.

Trained on who was hired, it learns to reproduce past hiring decisions. Trained on performance ratings, it learns to reproduce past rating behaviour, including whatever bias those ratings carried. Trained on tenure, it learns who stayed, which reflects who was made welcome.

So the model's accuracy is a measure of how faithfully it reproduces the past. A model with excellent accuracy on historical hiring data is, by construction, a good imitator of the decisions your organisation used to make. If those decisions were even mildly uneven, that unevenness is the training signal.

This is why hiring is different from most prediction problems, and why fairness interventions at the model layer often disappoint here. The bias cursus makes the general point; in hiring it is close to universal.

The practical implications. Ask any vendor what their model is trained to predict, and what generated that label. Treat historical-decision labels as a serious limitation rather than a detail. And be sceptical of claims that a tool removes human bias, since a model trained on human decisions has encoded it rather than removed it, and made it faster and more consistent.

Given all of the above, a sensible adoption order for an HR function.

First, audit for the prohibitions. Emotion inference and biometric categorisation, across every tool including features inside platforms you did not buy as AI. This is a check for an existing exposure rather than preparation.

Second, deploy the low-risk uses, where the value is real and the analysis is short. Drafting job descriptions, answering policy questions from documentation, summarising notes, generating onboarding checklists, scheduling. These improve the function's capacity without touching a decision about a person.

Third, measure your current process before automating any part of it. Selection rates and outcomes by group, on the existing human process. This does two things: it tells you whether you have a problem today, and it gives you the baseline against which any tool must be judged. Without it, a vendor's claim to reduce bias is unfalsifiable.

Fourth, and only then, consider decision-influencing tools, with the diligence from the procurement cursus and the measurement from the bias cursus.

The ordering is deliberate. Most of the available value sits in the second group, most of the risk sits in the fourth, and the third is what makes the fourth defensible. Organisations that start at the fourth acquire the risk before the capability to manage it.