AI in Human Resources

In December 2022 the Pew Research Center asked more than 11,000 American adults how they felt about AI in hiring. The answers split along a line that every HR leader should notice. Using AI to review job applications drew opposition from 41 percent. Using AI to make the final hiring decision drew opposition from 71 percent, with only 7 percent in favor. Two in three said they would not want to apply to an employer that used AI to help make hiring decisions¹.

Employers are moving faster than that public mood. In SHRM’s surveys of US HR professionals, the share of organizations using AI for HR tasks rose from 26 percent in 2024 to 43 percent in 2025². Much of that use is sensible: drafting job descriptions, answering employee questions, summarizing applications. Some of it drifts, often without anyone choosing it, from reviewing toward deciding.

The public drew the line in the right place. The distance between helping a person decide and deciding for them is where HR AI succeeds or fails, and it is a distance the executive, not the vendor, has to set.

Service work and people decisions

HR runs a wide range of work: employee questions, service requests, recruiting, onboarding, learning, talent reviews, workforce planning, pay and performance. AI can help in all of it. But the work comes in two kinds, and they need different designs.

Service work answers, routes and drafts. An employee asks how parental leave works; a manager needs a job ad; a request needs to reach the right team. The goal is speed and consistency, and the main risk is showing the wrong person the wrong data. People decisions are different in kind. Who is hired, who is promoted, who is paid more, who gets which shifts, who is asked to leave. The goal there is a decision that is fair, that someone owns and that the person affected can have explained to them.

The test that separates the two is one question: does the output shape someone’s job, pay or opportunity? Leaders who skip it make two opposite mistakes. They under-design the consequential cases, treating a candidate ranking as just another assistant. And they over-fear the safe ones, holding back a policy assistant that could save many hours because “AI in HR” sounds dangerous.

Start with the questions employees already ask

The highest volume and the lowest risk sit in the same place: the questions employees ask every day. Picture an illustrative case. Today, an employee who wants the parental leave rules searches the intranet, opens three documents that disagree, and then emails HR anyway. Someone in HR reads the email, finds the policy for that country and contract type, and writes the same answer for the fortieth time that month.

With a knowledge assistant, the flow changes. The assistant recognizes the kind of question. It retrieves the approved policy, and the right version of it, because policies differ by country, contract type, business unit and effective date. It answers and shows the source, so the employee can check it. And it hands over to a person when the case is unusual, sensitive or unclear: a bereavement, a grievance, a health question.

Two design choices decide whether this works. The first is the source. The answer must come from your policy, not from what a general model happens to know about leave law in general. The second is access. An assistant connected to HR systems can surface salaries, health notes or performance records to the wrong person if its permissions are loose; Privacy and Confidential Data covers how permission-aware retrieval prevents that. Design what the assistant may see before you design what it says. Then measure what matters to the employee: whether the answer matched the policy, whether sensitive cases reached a person, and whether HR still receives the email. Take that baseline yourself: independent evidence of what HR assistants deliver in real deployments is still thin, and most published figures come from vendors, so the time saved is a capability to prove, not a result to assume.

In recruiting, AI assists and a person decides

Recruiting is where many HR AI proposals arrive, because volume is the recruiter’s daily problem. It is also where the line between helping and deciding matters most.

AI does useful work on the left. It can extract and organize what is in an application, so a recruiter reads a clean summary instead of a messy file. It can draft job ads and interview guides. It can suggest internal candidates whose skills fit an open role, for a manager and the employee to review together. These are capabilities; there is little independent evidence yet that they improve who is hired, as opposed to how fast recruiters read. What it should not do alone is reject a candidate, promote or dismiss someone, or set pay. “Best candidate” is not a field in a database. It is a judgment about what the organization needs now, and it changes with strategy. A model trained on past choices knows only what you used to want.

A person at the end of the process is not, by itself, enough. In its first ruling on the GDPR’s rule on automated decisions, the EU Court of Justice held that an automated score that plays a determining role in someone else’s decision counts as a decision in its own right³. The case was about credit scoring, but the logic travels: if recruiters only ever see the candidates a ranking puts on top, the ranking has decided for everyone below the line.

The risk is not only subtle. In 2023 iTutorGroup, an online tutoring business, paid 365,000 dollars to settle the US Equal Employment Opportunity Commission’s first lawsuit over automated hiring, which alleged that its application software automatically rejected women aged 55 or older and men aged 60 or older⁴. No machine learning was needed; a filter nobody questioned was enough.

The same caution applies to attrition prediction. A model can find patterns linked to people who left before. That is a signal for a conversation at team level, not a verdict about an individual.

Yesterday’s decisions become tomorrow’s rules

HR data needs more care than most business data, because it is made of past decisions about people.

The loop starts with who was hired, promoted and rated highly over many years. Those decisions become the training data and, quietly, the definition of a good candidate. The model learns the pattern, including the parts that were habit rather than strategy. Its recommendations then shape the next round of decisions, and the loop closes.

Two shortcuts do not break it. Removing gender or age fields is not enough, because other fields carry the same signal; Bias and Fairness shows how these proxies work. Nor do newer models escape the problem. When researchers ran more than 3 million résumé-to-job comparisons through three language models, résumés with white-associated names were preferred in 85.1 percent of cases⁵. And a recruiter glancing over a ranked list does not remove a systematic pattern, because people tend to accept a ranking they are shown, as Operational and Workforce Risk explains.

What works is measurement. Compare outcomes by group, such as selection rates for women and men or by age band, before launch and at fixed intervals afterward, and assign someone to act on what the comparison shows. Fairness in HR has to be measured; it cannot be assumed.

Risk rises with what the system decides about a person

Put the two kinds of work and the bias loop together and you get a risk map. At each tier the question is the same: how much could this change someone’s working life?

At the bottom sit policy answers, drafting and routing; the main control is access. Next come talent matching and workforce analytics, which shape opportunities indirectly, so people need to be able to see a suggestion and why it was made. Then the high-risk tier. The EU AI Act lists as high-risk any AI used to recruit or select people, “to analyse and filter job applications”, and AI used to decide promotion or termination, to allocate tasks based on behavior or personal traits, or to monitor and evaluate performance⁶. An Annex III system that profiles people is always treated as high-risk⁷.

At the top is a tier many leaders do not expect. Since 2 February 2025, the EU has prohibited AI systems that infer the emotions of people at work, with narrow medical and safety exceptions. The Commission’s guidelines make the line concrete: a call center using webcams and voice analysis to track its employees’ emotions is prohibited, while analyzing customers’ emotions is outside this ban, and detecting physical states such as a driver’s fatigue is not emotion recognition at all⁸. Engagement scores read from faces on video calls fall on the wrong side of that line. The public is not far behind the law: 70 percent of Americans in Pew’s survey opposed using face recognition to analyze employees’ facial expressions¹.

The rules that already apply

HR AI is one of the most regulated uses of AI, and the rules sit in several places at once.

Two patterns run through the list. The first is timing. Several of these duties apply before the system is switched on: the consultation, the candidate notice, the bias audit, the impact assessment. A team that discovers them after a pilot has already broken some of them. The second is that a rule on paper is a floor, not proof of safety. Bias and Fairness shows how weakly New York’s audit law has been enforced¹⁴; a tool can pass a narrow legal test and still screen out the people you most wanted to hire.

Before approving, ask who owns the decision

When an HR AI proposal reaches your desk, one question sorts it.

If the answer is no, it is service work. Approve it with an approved source, access controls and a clear handover to a person, and measure accuracy and speed. If the answer is yes, the bar rises, and the proposal should answer five questions before it reaches you. Who makes the final decision, by name, and do they see every case or only the ones the system passes through? What personal data does the system use, and is any of it more than the task requires? How will outcomes be compared across groups, before launch and after? Who is told, and when: candidates, employees, their representatives? And what is the value once the cost of oversight, testing and consultation is counted? Sometimes the assistance is worth it and the automation is not.

Story: the program that agreed with its selectors

One of the clearest post-mortems of an automated applicant screen is almost forty years old, and it happened at a university.

In the 1970s, a doctor on the staff of a London medical school, part of the University of London, began writing a program to take over the first sift of applications. The school received about 2,000 a year, and reading them all was a heavy load for the selection panel. He built the program to do what the panel did. By 1979 it agreed with the panel’s gradings 90 to 95 percent of the time, and from 1982 every initial application passed through it before any person decided who would be interviewed¹⁵.

In December 1986 two senior lecturers at the school told the UK’s Commission for Racial Equality that the program discriminated. The commission’s investigation found that it deducted points from applicants with non-European names, about 15, and from women, about 3¹⁵. In 1988 the commission found the school guilty of racial and sexual discrimination. As many as 60 applicants a year may have been refused an interview because of their sex or racial origin¹⁶.

The post-mortem holds four lessons, and none of them is about software. First, the program was validated against the wrong standard. Agreement with past selectors was treated as proof of quality, but as the British Medical Journal noted, the program had not introduced new bias; it had faithfully reproduced the bias already in the panel’s decisions¹⁶. Second, nobody compared outcomes by group for years. The problem was found by two colleagues who looked, not by a control that was designed to look. Third, taking people out of the first sift removed the one place where someone might have noticed a strong application being turned away. Fourth, the school remained responsible. The program was its own, and so was the finding; three applicants who had been rejected were later offered places¹⁵.

Replace the medical school with any employer and the panel with ten years of hiring decisions, and the story describes a modern résumé screen almost exactly. The EU AI Act’s description of high-risk recruiting AI, filtering applications and evaluating candidates, would fit it word for word.

What this means for leaders

The lessons for an executive are practical. Start with service work, where volume is high and the consequence for any one person is low, and invest in the source and the access rules. In hiring, promotion, pay, task allocation and exit, let AI assist people and keep a named person accountable for every decision, including every rejection. Treat agreement with past decisions as a warning sign as much as a quality mark, and require outcome tests by group before launch and on a schedule. Find out which duties apply before the pilot, not after: consultation, notice, audits and the EU ban on emotion recognition at work. And judge success by the people outcome, not only by recruiter hours saved.

Check yourself

1. Removing gender and age fields makes a hiring model fair.MythFact
2. In the EU, using AI to infer employees’ emotions at work is prohibited, not just high-risk.MythFact
3. If a recruiter signs off at the end, a ranking tool cannot count as an automated decision.MythFact
4. High agreement with past hiring decisions proves a screening model is good.MythFact
5. In New York City, an automated tool that substantially assists hiring needs a yearly independent bias audit and candidate notice.MythFact
6. An employee policy assistant carries no real risk.MythFact

Reflection: where does your line sit?

What comes next

In HR, the cost of a confident mistake falls on a candidate or an employee. The next chapter, AI in Sales and Marketing, turns to the functions closest to revenue, where AI shapes what customers see, are offered and buy, and where personalization raises its own questions of trust.