AI is entering recruitment, employee relations and the boardroom. The prize is faster, better work. The contest is over who gets to judge—and who can challenge the answer.
The interviewer was a machine. The applicant’s chances improved.
In a field experiment recruiting customer-service workers in the Philippines, applicants assigned to an AI voice interviewer received offers at a rate of 9.73%, against 8.70% for those assigned to people: a gain of roughly one percentage point. Human recruiters still evaluated the interviews and made hiring decisions. The September 2026 working paper describes a particular recruitment process, with data collected in 2025. Its result deserves attention without becoming a verdict on every interview. Jabarian and Henkel.
For human resources, the interesting development is the division of labour. A task associated with interpersonal skill was handed to software. Recruiters retained responsibility for interpreting what it produced. It is a glimpse of how a job can change without acquiring a new title.
That arrangement raises a question that follows AI through the entire HR department. When software collects the evidence, selects what matters and proposes an answer, how much judgement is left for the person approving it?
The answer matters to the administrator handling a leave query, the employee-relations specialist investigating a complaint and the chief people officer defending a restructuring. AI offers each of them assistance. It also gives their employers reasons to reconsider what they are paying them to do.
Hiring gets a new interviewer
Recruitment contains several different jobs: attracting applicants, defining a role, screening, interviewing, assessing, negotiating and deciding. A useful tool in one stage can cause trouble in another.
Consider job advertisements. A separate field experiment on an online labour market offered employers AI-written first drafts. Among employers who posted a job, writing time fell by about 44%. More employers posted vacancies, yet the researchers found no discernible increase in matches overall. The September 2026 working paper also found more generic job posts. These were platform experiments using 2023 data, rather than a test of today’s entire recruitment market. Wiles and Horton.
For a talent-acquisition team, the implication is practical: judge the whole funnel. A swiftly written advert is valuable if it helps suitable people understand a real job. A swelling pile of applications can be an expensive distraction. Sourcing and employer branding still require an accurate account of the work, its conditions and the person likely to succeed in it.
Screening adds a harder problem. The UK Information Commissioner’s Office gathered evidence from more than 30 employers between March 2025 and January 2026. It concluded that many were likely making solely automated recruitment decisions without meaningful human involvement, requiring safeguards beyond those it observed. The findings came from voluntary engagement, rather than an enforcement investigation. ICO, Recruitment rewired.
A recruiter who reviews a shortlist may have no opportunity to reconsider everyone excluded before it was drawn up. Oversight needs time, information and authority. Otherwise, a signature can conceal how little choice remains.
The recruiter’s strongest claim to value is therefore an informed assessment: defining relevant evidence, testing whether a selection method works and making decisions that survive scrutiny. Producing more recruitment material is a weaker claim.
The back office changes shape
HR operations offers many plausible starting points: finding the right policy, drafting a routine response, assembling an onboarding checklist or directing a query to its owner. These are candidates for testing, with the organisation’s own information and controls, rather than a promise that every process can run unattended.
There is encouraging experimental evidence on some written tasks. A July 2026 study using ChatGPT 3.5 found faster completion and better-rated outputs on simulated junior HR assignments involving communication, recruitment and selection. Its participants were 91 final-year students in Greece. That supports further testing of routine writing assistance; it gives little basis for calculating the staffing needs of a working HR service centre. Katsavou and colleagues.
Payroll, benefits, pensions and contractual changes demand a more exacting distinction. Established software can apply specified calculation rules. A language model drafting an explanation of those rules is performing a different task. Fluent wording does not establish that an entitlement, deduction or payment is correct.
The useful measures are resolution quality, errors, repeat contacts and total handling time, including checks and corrections. If a chatbot answers quickly but sends a confused employee through three more conversations, the organisation has moved work rather than removed it.
The HR information systems team becomes central to making this work: maintaining reliable records, controlling access and connecting tools to the right sources. Onboarding, contract changes, absence administration, international mobility and offboarding each contain different permissions, local rules and exceptions. A common chat interface should not flatten those differences. Sensitive health or case records need particularly careful handling.
For administrators and coordinators, routine drafting and retrieval may occupy less of the day. Exception handling, records quality and understanding how systems connect may occupy more. That transition can bring better work. It can also bring fewer entry-level opportunities if employers remove routine tasks without building another route into the profession.
Employee relations is the hard case
A grievance can arrive as a mountain of messages. A tool that produces a chronology or identifies relevant passages could help an investigator find their way through it. Such assistance needs to be tested against the underlying record: omitted context can matter as much as an invented fact.
The central ER questions remain questions about evidence and power. Was a warning proportionate? Were comparable cases handled consistently? Did a manager’s account omit something material? Does an apparently neutral rule disadvantage someone who needs an adjustment?
In Britain, the Acas disciplinary and grievance Code calls for investigation, an opportunity for the employee to respond and an appeal against formal decisions. Its emphasis is on a fair process. A convincing case summary cannot substitute for those steps. Acas Code, particularly paragraph 4.
For ER specialists, the opportunity is to spend less effort assembling documents and more on testing accounts, resolving disputes and advising managers. The danger is accepting the first coherent explanation because it arrives with the polish of a finished report.
Industrial relations adds another dimension. Software can help compare proposals or model the cost of a settlement. A negotiation still requires parties to decide what they will accept and whether commitments are credible. Consultation, collective bargaining and relationships with employee representatives should be designed into technological change, rather than invited in once the decision is effectively settled.
AI can also create ER work. An employee may challenge a recruitment score, an automated performance flag or a decision based on inaccurate data. Someone needs the ability to investigate the system as well as the employee.
A performance score is a policy
Performance management, reward and inclusion meet at an awkward point: deciding what counts.
An appraisal assistant can organise a manager’s notes. A reward model can explore the cost of different pay proposals. Neither activity settles whether the underlying measures recognise good work fairly. Fast case closure might reflect skill, simpler cases or an incentive to pass difficult work elsewhere.
Historical promotion and pay records also require interpretation. If past opportunities were distributed unequally, treating those outcomes as an uncomplicated definition of merit risks preserving the pattern. Diversity, equity and inclusion specialists have a role in examining access to opportunities, selection outcomes and the possibility of challenge, alongside the quality of generated language.
Monitoring deserves particular care. Eurofound’s October 2026 report, using the European Working Conditions Survey 2024, associates algorithmic management with greater work demands and job strain. Performance monitoring is especially associated with poorer health and wellbeing outcomes. These are survey associations, and the systems examined extend beyond generative AI. They nevertheless give HR a reason to test effects on working life alongside efficiency. Eurofound.
For wellbeing and employee experience, that means asking whether a system reduces frustration or simply makes work more intense. Summarising an engagement survey may be useful. Inferring an individual’s mental health from their digital behaviour raises a much more intrusive proposition. Aggregated findings can also expose individuals when groups are small.
A model’s tidy score should begin a discussion about its assumptions. It should not end one.
Training without learning?
Learning and development teams have an enticing prospect: material adapted to a learner’s needs, practice available on demand and help delivered during the task itself. Talent teams could use similar tools to explore internal moves, development opportunities and potential succession gaps.
The test is whether people become more capable.
A study published in Nature Human Behaviour in October 2026 examined consultants tackling unfamiliar data-science tasks. Its analysis included 487 participants. ChatGPT assistance improved task performance, but the researchers found no evidence of better technical knowledge on a subsequent test without the tool. The study involved BCG and OpenAI researchers and does not establish the effect of every form of AI-supported learning. Wiles and colleagues.
For L&D, the distinction should change evaluation. Successful work with assistance, independent understanding and the ability to detect a bad answer are separate outcomes. A polished training product or a completed module tells us little about all three.
It also changes HR’s own apprenticeship. Junior practitioners traditionally encounter recurring questions, work through drafts and learn why a senior colleague changes them. If software supplies the first answer, teams need deliberate opportunities to explain reasoning, practise difficult cases and learn from corrections. Removing the practice while retaining the expectation of expertise would be a poor bargain.
The spreadsheet acquires a voice
People analytics may become easier to use. A well-designed assistant could help an HR colleague explore absence patterns, query turnover data or prepare a workforce scenario without constructing every spreadsheet or line of code themselves.
The accessibility of the answer makes the quality of the question more important.
Suppose an analysis finds that a department’s leavers are concentrated under one manager. That is a lead for investigation. The department might also have lower pay, scarce skills, difficult shifts or a recent restructuring. A prediction about who may leave does not identify which intervention will persuade them to stay.
Workforce data analysis therefore requires reliable definitions, sensible comparisons and checks against underlying records. Headcount, full-time equivalents, vacancies and productive capacity answer different questions. A model that mixes them can produce a persuasive account of a workforce that does not exist.
Strategic workforce planning adds uncertainty about demand. If a task becomes faster, an organisation might reduce staffing, expand its service, improve quality or absorb an existing backlog. Each choice implies a different hiring plan. The time saved on one activity must be translated into the capacity of the whole service before it becomes a headcount assumption.
The same discipline matters in succession planning and internal mobility. A skills profile can suggest possibilities; it cannot establish that someone is ready for a role merely because the vocabulary matches. People need ways to correct their profiles, demonstrate new abilities and contest a recommendation that narrows their prospects.
For analysts, statistical reasoning, data stewardship and the ability to explain uncertainty may become more valuable as routine reporting becomes easier. For everyone consuming their work, the crucial habit is asking what evidence could make the apparent explanation wrong.
OD moves beyond the organisation chart
Organisation development concerns how a business works: its roles, relationships, behaviour and capacity to change. Organisation design concerns the distribution of work and authority. AI gives both disciplines something substantial to investigate.
A 2026 Organization Science study involved 791 P&G professionals working on product-innovation challenges. Individuals using AI matched the assessed performance of teams working without it, and AI-assisted proposals drew more evenly on technical and commercial perspectives. This was a bounded experiment on innovation work, rather than evidence that an organisation can dispense with teams or managers. Dell’Acqua and colleagues.
For OD, the interesting possibility is that some expertise becomes easier to access across departmental boundaries. That could alter handovers, decision rights and the mix of skills a team needs. It requires testing in the organisation’s real work.
A role should be redesigned around complete responsibilities: doing the task, checking its output, handling exceptions and owning the consequences. Moving the first step to AI can increase the demands of the others. Wider managerial spans, for example, could leave too little time for the review on which the new arrangement depends.
Change management must also reach beyond explaining the software. Employees will reasonably want to know how targets, development, workload and job security may change. A demonstration of a better draft does not answer those questions.
The career ladder gets a different test
These findings do not produce a reliable league table of safe and doomed HR jobs. Titles and responsibilities vary between organisations. The role implications below are an assessment of how work could shift, rather than measured forecasts of redundancies.
HR assistants, administrators and coordinators face direct pressure on routine writing, retrieval and coordination. Their development needs to include systems, data quality and exceptions, with supervised exposure to real decisions. Leaving them to approve machine output without understanding it would weaken the service and their prospects.
HR advisers, generalists and specialist teams need enough knowledge to recognise when a plausible answer fails the circumstances. Recruiters, ER practitioners, reward specialists and L&D teams should test tools against the outcomes their function exists to deliver. Producing more documents is an inadequate substitute.
HR business partners may find routine reporting and briefing preparation easier. Their harder work is interpreting what the business needs, challenging a manager’s diagnosis and connecting a people decision to performance, cost and risk. An HRBP who mainly relays information is more exposed than one who can change a consequential decision. Even the latter needs time and authority to do so.
Associate HR directors and heads of function are positioned to turn experiments into a workable service. They need to decide which work stays local, which moves into shared services, how specialists support difficult cases and where decisions can be escalated. Their performance should include the quality and resilience of the resulting system, alongside its cost.
HR directors face choices about investment, staffing, capability and the balance between central standards and local needs. They also need to understand what happens when a supplier changes its model or a widely used service becomes unavailable. A cheaper operating model with unrecognised dependencies may prove expensive when it fails.
Chief people officers face the widest question: what kind of employment relationship will the organisation build around AI? They need to negotiate decisions about productivity gains, redeployment, development and workforce trust with the chief executive, finance and technology leaders. Their own reporting and presentation work can be assisted. The value of the role depends on influencing the choices behind the presentation.
Readiness should not be assumed at the top. CIPD’s 2026 Ireland report, based on 330 responses collected in November 2025, found that only 19% of respondents believed their leadership was fully equipped to lead in an AI-enabled environment. This is a respondent survey in one country, rather than a census of executive capability. CIPD report, methodology and key findings.
Seniority is not a technical qualification. Nor is familiarity with a chatbot evidence that someone can redesign a people function.
What would count as progress?
An HR team needs a starting measure and a real outcome. Recruitment can examine suitable hires and subsequent performance. Operations can measure accurate resolution. ER can examine the quality of investigation and the handling of challenges. L&D can test retained capability. Workforce planning can compare its assumptions with what subsequently happens.
Those measures should include the full cost of review, correction, implementation and the work passed to managers or employees. A saving in one team can be another team’s new burden.
The evidence also argues against confident timetables for HR’s disappearance. Research linking surveys of 25,000 Danish workers across 11 occupations, including HR, to administrative records found no detectable average earnings or hours effects in the period studied, through December 2024. That early national result cannot settle the effects of later systems or changes elsewhere. It does show why capability claims and labour-market outcomes must be kept separate. Humlum and Vestergaard’s account of their research.
AI gives HR a chance to make some work quicker, some expertise more accessible and some decisions better informed. Employers will decide how much of that capacity becomes a better service, a smaller team or a higher target. HR has an interest in shaping that choice before it is made elsewhere.
Eventually, an employee will ask why they were rejected, overlooked, underpaid or selected for a change. The organisation needs someone who can inspect the evidence, explain the reasoning and correct a mistake. That capability takes knowledge, time and authority. HR’s new power struggle is over whether it retains all three.
Research note: This is a desk-researched analysis with a UK focus and international evidence. It distinguishes experiments, surveys, regulatory findings and the publication’s assessment of future role changes. No practitioner interviews were conducted. Research and drafting were assisted by AI. The cited studies do not establish a single automation percentage or a replacement timetable for the HR profession.