Where AI Actually Helps Inside a People Function

Two years ago I ran an experiment I did not tell anyone about. I had a model score 200 CVs for a role we had already hired for, then compared its ranking against how those candidates actually performed in interviews.

It ranked our eventual hire 74th.

It was not stupid. It was doing exactly what it was asked: matching surface features against a job description. Our hire had switched industries, had a two-year gap and had none of the listed keywords. She was also, obviously, the best person in the pipeline within about eight minutes of talking to her.

That result shaped how I have used AI in people work ever since. The line is not between simple and complex tasks. It is between tasks where the text is the work and tasks where the text is evidence about a human being.

Where it genuinely earns its place

Drafting anything that starts from a blank page

Job descriptions, interview guides, rejection notes, onboarding checklists, policy first drafts, offer explainers. A model gets you to a solid 70 percent draft in two minutes instead of forty.

The gain is not just speed. It is that things which used to get skipped now get done. Before, a structured interview guide existed for our top three roles because that is all anyone had time to write. Now every role has one, because the first version takes four minutes. Consistency across 30 roles beats brilliance across three.

Retrieval over your own documents

This is the most underrated use and the least discussed. Our policy handbook, benefits documentation and process guides run to a few hundred pages. Nobody reads them. Everybody asks a person instead, and that person is usually in HR.

Putting a retrieval assistant over those documents cut routine policy questions to my team by roughly 60 percent. Two conditions made it work: every answer cites the source document and section, and anything about pay, termination or immigration is deliberately routed to a human. It answers "how many days of leave carry over" and refuses to answer "am I being paid fairly".

The citation rule is not decoration. Grounding a model in retrieved documents does reduce invented answers; that has been measured since Shuster and colleagues' 2021 work on retrieval-augmented chatbots. It does not remove them. When a Stanford team tested commercial legal research tools built exactly this way, they found each one still hallucinated between 17 and 33 percent of the time. Fewer errors than a general chatbot, nowhere near zero. A citation is what lets an employee check the answer in ten seconds instead of trusting it.

Summarising and structuring what humans produced

Turning your own interview notes into a structured scorecard. Clustering 300 free-text engagement survey responses into themes. Summarising a long performance conversation into agreed actions.

Note the direction of travel: a human generated the judgement, the model organised it. That is the safe orientation. When it runs the other way, you are in trouble.

Removing scheduling and coordination toil

Unglamorous and worth real money. Availability matching, reminder sequences, document chasing, interview logistics. Our coordinator went from spending most of her week on calendar work to spending most of it on candidate experience, the part that actually differentiates us.

Where it does not belong

Ranking or scoring candidates

I have tested this repeatedly and I will not do it in production. Three reasons.

It optimises for legible signals (brand-name employers, keyword density, tidy career paths), and those correlate with access and privilege far more than with capability. The candidate who took two years out, or studied somewhere you have not heard of, or changed fields at 34, gets filtered by a system that cannot see the thing that makes them good.

This is not a hunch, and it is not new. In 2018 Reuters reported that Amazon had scrapped an experimental CV-scoring model after finding it penalised CVs containing the word "women's", as in "women's chess club captain". (Amazon said recruiters never used it to evaluate candidates. The point stands: the model learned the bias from ten years of the company's own hiring history.) The newer models have not fixed it. In 2024 Bloomberg ran GPT-3.5 through a resume-ranking test with identical CVs under different names and concluded that if a company hired solely on its rankings, protected groups would be adversely affected. The same year, researchers at the University of Washington tested the embedding models used for retrieval-based screening and found White-associated names preferred in 85.1 percent of their race comparisons.

It is confidently wrong in a way humans are not. A recruiter who is unsure says so. A model returns 8.4 out of 10 with the same fluency whether it is right or badly out of its depth. False precision is worse than acknowledged uncertainty because it stops the conversation. There is a mechanical reason for this: a 2025 paper from OpenAI researchers argues that models hallucinate partly because training and evaluation reward guessing over admitting uncertainty. And the people reading the output make it worse. A 2025 study in Nature Machine Intelligence found that users overestimate how accurate a model's answers are, and that longer explanations raised their confidence even when the extra length did not make the answer any more correct. A fluent paragraph justifying a score is exactly the wrong thing to hand a tired hiring manager.

And if you hire in Europe, you must be able to explain an adverse decision. GDPR Article 22 already gives candidates the right not to be subject to a decision based solely on automated processing, and Recital 71 names "e-recruiting practices without any human intervention" as the example. On top of that, the EU AI Act lists systems used "to analyse and filter job applications, and to evaluate candidates" as high-risk. High-risk does not mean banned. It means obligations: human oversight, telling the people affected, and giving them a meaningful explanation of the system's role in the decision. After the 2026 Digital Omnibus amendment, those obligations apply to recruitment systems from 2 December 2027. That is a delay, not a reprieve. Anyone building a screening pipeline now is building it for a world where "the model ranked them lower" is not an acceptable answer.

Outside Europe the rules are patchier but heading the same way. New York City's Local Law 144 has, since July 2023, barred employers from using an automated hiring tool unless it has had an independent bias audit within the past year and candidates have been given notice.

Anything close to a termination, promotion or pay decision

These are the decisions where being wrong changes someone's life. They require context that is nowhere in your data: the six months someone carried a struggling team, the manager change that derailed a quarter, the medical situation nobody wrote down. A model sees the record. The record is not the person.

Writing the message that should have cost you something

An AI-drafted note declining a candidate after five rounds is technically fine and emotionally hollow, and people can tell. The same is true of a condolence note, a difficult performance message, or a company-wide announcement about layoffs.

If the point of a message is to show that a person spent time on you, outsourcing it defeats the message. Write those yourself. They are a small fraction of your volume and nearly all of your reputation.

Culture and values work

Ask a model to write your values and you get a competent, generic list that could belong to any company. Values are the record of trade-offs a specific group of people actually made under pressure. They have to be extracted from your own history, not generated.

Reading people's emotions

Some vendors sell tools that claim to read engagement, stress or honesty from faces and voices in interviews and meetings. Leave them alone. The science behind them is weak, and in the EU the question is already settled: since 2 February 2025, Article 5 of the AI Act has prohibited AI systems that infer the emotions of people in the workplace, except for medical or safety reasons.

The rules I actually operate by

After enough iterations, this is what we hold to:

1. AI drafts, humans decide. Any output that affects a person's status is reviewed and owned by a named human before it goes anywhere.
2. Never the sole basis for an adverse decision. No rejection, no performance rating, no termination traceable to a model output.
3. Cite or refuse. Anything answering a policy question links to the source. No source, no answer.
4. Candidates are told. If AI touches a hiring process, we say where and how. Nobody has ever objected to a clear explanation. This is also becoming the law rather than a courtesy: New York City already requires notice, and the EU AI Act's Article 26 requires employers using high-risk systems to inform the people subject to them.
5. Sensitive data stays out. Health information, immigration status, compensation records and grievance details do not go into general-purpose tools.
6. Audit quarterly. We sample outputs and look for drift. Twice this has caught tone problems in candidate communications that nobody would have flagged individually.

What this is really worth

Honest accounting. AI has not reduced my headcount. It has moved where the hours go.

Roughly 30 percent of our people-operations time was administrative production: drafting, formatting, answering the same question, coordinating calendars. Most of that is now assisted, and the recovered time went into candidate experience, manager coaching and actually talking to employees.

That is the real return, and it is bigger than it sounds. The work that was crowded out is the work that only humans can do, and it was always the work that determined whether people stayed.

The best large-scale evidence points the same way. Brynjolfsson, Li and Raymond studied 5,179 customer-support agents given an AI assistant, in a paper published in the Quarterly Journal of Economics in 2025. Productivity rose 15 percent on average. The detail that matters for HR teams is who gained: less experienced and lower-skilled agents improved in both speed and quality, while the most experienced saw small gains in speed and small declines in quality. AI lifts the floor of routine work. It does not raise the ceiling of judgement, and it can quietly lower it if your best people lean on it for the work they are best at.

If you are starting, start with retrieval over your own documents. It is the lowest risk, the fastest to show value, and it teaches your team where the technology is trustworthy before you point it anywhere near a decision about a person.

Sources

1. Shuster, K., Poff, S., Chen, M., Kiela, D. and Weston, J. (2021). Retrieval Augmentation Reduces Hallucination in Conversation. arXiv:2104.07567.
2. Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D. and Ho, D. E. (2024). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. arXiv:2405.20362.
3. Dastin, J. (10 October 2018). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters (archived copy).
4. Yin, L., Alba, D. and Nicoletti, L. (7 March 2024). OpenAI's GPT Is a Recruiter's Dream Tool. Tests Show There's Racial Bias. Bloomberg. Methodology and code.
5. Wilson, K. and Caliskan, A. (2024). Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. The authors issued an erratum in August 2026 inverting the gender-only results; the race and intersectional results cited here were unaffected.
6. Kalai, A. T., Nachum, O., Vempala, S. S. and Zhang, E. (2025). Why Language Models Hallucinate. arXiv:2509.04664.
7. Steyvers, M. et al. (2025). What large language models know and what people think they know. Nature Machine Intelligence 7, 221–231.
8. Regulation (EU) 2016/679 (General Data Protection Regulation), Article 22 and Recital 71.
9. Regulation (EU) 2024/1689 (the AI Act), Annex III point 4, Article 5 and Article 26, via the European Commission's AI Act Service Desk.
10. Regulation (EU) 2026/1744 (Digital Omnibus on AI), 8 July 2026.
11. NYC Department of Consumer and Worker Protection. Automated Employment Decision Tools (Local Law 144 of 2021).
12. Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics 140(2), 889–942. The 2023 NBER working paper version reports 14 percent; the published figure is 15.

Legal summaries are accurate as of October 2026 and are not legal advice. AI regulation is moving quickly; check the current text before relying on any date.