Zak FentonWorking notes

· Note

Does AI give HR its time back? Nobody has measured it.

The government's headline figure, 26 minutes a day, is what people said they saved.

Evidence review (in progress)AI is sold to HR on time saved. Does anyone get the time back?

My company has been quietly and subtly rolling out AI features that do admin. I can’t tell you they’ll save you time. On the evidence I’ve just been through, nobody selling an AI tool to an HR team can tell you that yet.

We haven’t made a fuss about it, and that’s deliberate. I don’t believe in big AI announcements in 2026, going into 2027. I’d sooner say what a tool changed once I can count it. I have a commercial stake in the answer all the same, so the review behind this note held anything favourable to these tools to a higher bar.

Where 26 minutes comes from

The UK government’s headline figure for what AI saves people at work is 26 minutes a day. It comes from a cross-government experiment with an AI assistant in the UK civil service.1 Staff were asked how much time they saved and picked from a set of ranges. The 26 minutes is the average of the midpoints of those ranges, with the top range counted as 60 minutes. There was no comparison group and no baseline.

Last week I found that the burnout figure HR is shown about itself came from a survey answered by whoever chose to answer it (Is HR more burnt out than everyone else?). This week’s figure is the same kind of number. It’s what people said.

Better designs, same question

Other evaluations were more careful. The Department for Work and Pensions compared licence holders with matched non-users and estimated 19 minutes a day.2 HMRC handed out licences at random.3 A trial in 68 English schools randomised which teachers were offered an AI assistant.4 In every one of them the time was reported by the person using the tool, in a survey or a diary.

Minutes a day saved, as people reported them

UK government evaluations of the same AI assistant, all self-reported

Cross-government experiment, no comparison group26
Department for Work and Pensions, against matched non-users19
Department for Work and Pensions, unadjusted10
Lines show the 95 per cent interval. The cross-government figure is an average of the midpoints of ranges people picked, with the top range counted as 60 minutes, and has no interval. The Department for Work and Pensions' 19 minutes covers eight routine tasks, from self-reported time bands.

One randomised trial took its time from software logs. Across 66 firms, people offered an AI assistant spent 1.4 hours a week less on email, and the trial detected no shift in the quantity or composition of their tasks.5 Three of its four authors work for the tool’s maker. That is the best measured number there is, and it is about email.

No study of an AI assistant in an office has measured whether the time became more work done. The Department for Business and Trade’s evaluation gets quoted for saying it found no evidence that time savings led to improved productivity.6 The same report says productivity was not a key aim and limited data were collected, so that sentence can’t be read as a finding in either direction. The nearest thing to a measured answer is from Denmark. Researchers linked about 25,000 workers, HR professionals among them, to national registers and found no effect on earnings or recorded hours, ruling out anything larger than 2 per cent.7 Pay and recorded hours stand in for output here, because output itself wasn’t measured.

What people believe about their own time

The one trial that compared what people believed with what was timed found them wrong. Sixteen experienced software developers expected AI to cut their completion time by 24 per cent, believed afterwards it had cut it by 20 per cent, and took 19 per cent longer.8 That was software, with early-2025 tools, and a follow-up hasn’t settled it.9 I wrote about it in Faster, they said. Slower, they were. It’s too far from admin to lean on. It is a reason to ask how any time-saved figure was collected before repeating it.

Expected, believed and measured

Change in task completion time with AI allowed, 16 experienced open-source developers, early 2025

Forecast before the tasks24% less time
Estimated afterwards20% less time
Measured19% more time
Bars run left for less time and right for more, from zero in the middle. The developers logged their own completion times, checked against screen recordings for some tasks. The setting is software development, so this is indirect evidence for admin work.

Where the time goes

When people are asked what they do with time an AI assistant saves, the answer is more work. In the Department for Work and Pensions trial, 89 per cent of users who answered said they moved it to other work tasks.2 In Denmark, fewer than 10 per cent of users expected to take more breaks or leisure.7 The cross-government report, the source of the 26 minutes, says it was not possible to identify how the time saved was spent.1

I recognise the pattern. In September I wrote about spending 98 per cent of a weekly AI allowance the day after it reset (The only thing that stops me is a usage bar). I’ve filled every hour AI has saved me.

What Danish users expected to do with time a chatbot saved

Share choosing each answer, among users reporting time savings. More than one answer allowed.

All users

More time on other tasks84.7%
More time on the same tasks29.4%
More breaks7.2%
More leisure5.8%

HR professionals

More time on other tasks89.9%
More time on the same tasks21.9%
More breaks9.4%
More leisure6.8%
These are answers about what people expected to do with the time. Danish survey data, Table E.3 of the study.

Some of the time goes on checking. In a Home Office pilot, specialists screened every AI summary before anyone used it and removed 9 per cent as inaccurate or incomplete.10 How long checking takes has almost never been timed, and no HR study has timed it.

The promise that less admin means less burnout rests on doctors. In a randomised trial of a note-drafting tool, work exhaustion fell by 0.44 points on a five-point scale over 24 weeks, and a 2026 meta-analysis rated the certainty low.1112 Outside medicine, no controlled study has measured it.

Then I looked for HR

The review found no controlled study that has measured time saved, accuracy in live use or strain from AI in HR, payroll, reward or benefits administration. That’s after 50 logged searches of four research databases and the web.

What exists is one interviewee’s accuracy figure for one firm’s HR bot, and a payroll provider’s count of service desk cases before and after it launched an app.1314 The nearest trial is in US unemployment insurance casework, where examiners rated an AI system highly and it did not improve quality or efficiency against the control group.15 Only its abstract could be read.

What exists for HR, payroll, reward and benefits administration

Nonecontrolled studies of time saved, accuracy in live use or strain
Saidone interviewee's accuracy figure for one firm's HR bot, with accuracy undefined
Countedpayroll service desk cases in two three-month windows, with no control
From 50 logged searches of four research databases and the web. The nearest trial is in public benefits casework, outside HR.

The offers have arrived faster than the studies. At the UK’s main reward and benefits forum, one session in 19 mentioned AI in 2022. On this year’s published agenda it is 20 in 53.16 I counted them from the agendas, by title and description.

What I’d buy it on

The tools may well save time. Nobody has counted the hours in HR, so I wouldn’t buy on them.

I’d buy on things you can count in your own team: mistakes that reach an employee, queries that come back a second time, and jobs that weren’t getting done at all. Before signing, I’d put five questions across the table.

  1. Was that time logged, or did people tell you?
  2. Compared with whom?
  3. What happened to the error rate, and who was checking?
  4. How long does the checking take?
  5. What will our people see that they don’t see today?

The study nobody has run

The study that would settle it is easy to describe. Give HR shared-service, payroll or benefits teams the tool in a random order over six to nine months. Take handling time from the case system’s logs. Audit a sample of the output blind. Measure burnout at the start and every three months. It needs an evaluator with no commercial link to the tool, which rules me out.

If you run one of those teams and would let it be measured, or you’re a researcher who wants to do it, email me. I’d like to see it done.

Taking the admin away is still worth doing, and we’re rolling out a tool that does it. I wouldn’t buy it on the hours, mine included.

Footnotes

  1. Government Digital Service (2025). Microsoft 365 Copilot Experiment: Cross-Government Findings Report. 2 June 2025. gov.uk. Section 1.2, the time savings section and Figure 5, and section 6. ↩ ↩2

  2. Arzilli, F., Lynch, C. S. and Page, L. (2026). An Evaluation of DWP’s Microsoft 365 Copilot Trial. Department for Work and Pensions, 29 January 2026. gov.uk. Chapter 3 and chapter 4, Table 4.3. ↩ ↩2

  3. HM Revenue and Customs (2026). Evaluating the Impact of Microsoft Copilot in HMRC. 9 July 2026. gov.uk. Results, section 2. Published summary; the full methods are not published. ↩

  4. Roy, P. et al. (2024). ChatGPT in Lesson Preparation: A Teacher Choices Trial. National Foundation for Educational Research for the Education Endowment Foundation. Evaluation report, ERIC ED680921. Executive summary and implementation findings. ↩

  5. Dillon, E. W., Jaffe, S., Immorlica, N. and Stanton, C. T. (2025, revised November 2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795. doi:10.3386/w33795. Working paper. ↩

  6. Department for Business and Trade (2025). The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade. August 2025. gov.uk. Executive summary, page 5; page 32. ↩

  7. Humlum, A. and Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. doi:10.3386/w33777. Abstract; sections 2.3 and 3; Table E.3. Working paper. ↩ ↩2

  8. Becker, J., Rush, N., Barnes, E. and Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089, version 2. Not peer reviewed. ↩

  9. Becker, J., Rush, N., Cunningham, T., Rein, D. and Mahamud, K. (2026). We are Changing our Developer Productivity Experiment Design. METR, 24 February 2026. metr.org. ↩

  10. Home Office (2025). Evaluation of AI trials in the asylum decision making process. Research note, 29 April 2025. gov.uk. Sections 4.1, 5.1 and 7. ↩

  11. Afshar, M., Baumann, M. R., Resnik, F. et al. (2025). A Pragmatic Randomized Controlled Trial of Ambient Artificial Intelligence to Improve Health Practitioner Well-Being. NEJM AI. doi:10.1056/aioa2500945. Abstract and methods. ↩

  12. Gong, E. J., Bang, C. S. and Lee, J. J. (2026). Cognitive Workload and Mental Burden in Health Care Professionals Interacting With AI: Systematic Review and Meta-Analysis. Journal of Medical Internet Research, 28, e93618. doi:10.2196/93618. Abstract. ↩

  13. Malik, A., Budhwar, P., Patel, C. and Srikanth, N. R. (2022). May the bots be with you! Delivering HR cost-effectiveness and individualised employee experiences in an MNE. The International Journal of Human Resource Management, 33(6), 1148-1178. doi:10.1080/09585192.2020.1859582. Findings, interviewee quotation. ↩

  14. NHS Shared Business Services (2019). Over 330 NHS Workforce Hours Saved Every Month After Launch of New Payroll App. Press release reproduced by WiredGov, 27 March 2019. wired-gov.net. Paragraph 4. ↩

  15. Martin, O., Magesh, V., Surani, F., Rodolfa, K. T., Ho, D. E. and Perez, A. (2026). Evaluating Generative AI in Benefits Administration: A Demonstration Project. Proceedings of the Symposium on Computer Science and Law. doi:10.1145/3788646.3789527. Abstract. ↩

  16. REBA Future Forum agendas for 2022 and 2026, as published on 17 November 2026. One session is one timed agenda row. A session counts if its title or description contains AI, artificial intelligence, agent, agentic, automation or machine learning. Counted twice, independently, with the same result. ↩

The evidence behind this note (in progress)AI is sold to HR on time saved. Does anyone get the time back?