AI is sold to HR on time saved. Does anyone get the time back?
The note this review supportsDoes AI give HR its time back? Nobody has measured it.
AI tools are sold to HR teams on the time they save. This review asks what the evidence shows about that time: whether anyone has measured it, whether it becomes more work done, where it goes, what checking the output costs, and whether taking admin away eases strain. I run a benefits broker, and my company has been rolling out AI features that do admin. I have a commercial stake in the answer, so I held findings favourable to such tools to a higher bar.
How this was checked
The question was split into eleven tiers, from who uses AI at all to HR, payroll, reward and benefits administration itself. Each tier was searched in trials, government evaluations, working papers, official statistics and professional body surveys, then graded with GRADE logic. Randomised trials start high and observational studies start low, and both lose grades for risk of bias, inconsistency, indirectness and imprecision. Surveys and case studies from firms that sell AI, HR or benefits technology, or consultancy, were recorded and labelled and never allowed to set a grade. Most of the evidence concerns tools from 2023 to 2025, older than the ones sold now. Methods, limits and the checking of this page are at the end. Every grade here is provisional.
The headline
| What is said | What the evidence says |
|---|---|
| 49 per cent of HR and reward professionals believe AI will free up time for more strategic work C3 | A belief, from a survey that does not state its sample or recruitment. No study has measured where saved time goes in HR, and the UK cross-government report could not identify how time saved was spent C15. |
| AI saves civil servants 26 minutes a day C13 | An average of the midpoints of self-reported ranges C14. With a matched comparison group, the Department for Work and Pensions’ estimate was 19 minutes, still self-reported C16. The one randomised trial that logged time found 1.4 hours a week less email C22. |
| Time saved becomes productivity | No office evaluation has measured output. The Department for Business and Trade’s sentence on productivity is no evidence of effect, because it did not set out to measure productivity C27. Danish registers show no effect on earnings or recorded hours C34. |
| Taking admin away reduces burnout | Small falls in self-reported exhaustion in trials of doctors’ note-taking scribes, at low certainty C47C50. No controlled study outside medicine has measured it. |
| The tools are accurate enough that checking does not eat the saving | Error rates are measurable and vary by task C42C43C44. Checking time is rarely timed, and no HR study has timed it. |
Eleven tiers at a glance
| Tier | Grade | Direction | In one line |
|---|---|---|---|
| 1. Who uses it | Low | Descriptive | About a third of UK businesses with 10 or more staff use some AI. No official survey measures HR teams’ use. |
| 2. What HR expects | Very low | Descriptive | Mostly sellers’ surveys of belief, with recruitment unstated. |
| 3. Controlled task experiments | Moderate for speed and rated quality; low for accuracy | Improves; mixed where accuracy was checked | Faster on short writing tasks, and less often correct on a task beyond the tool. |
| 4. Time saved at work | Low | Improves, on specific tasks | Mostly self-reported, even with a comparison group. One logged trial found less email. |
| 5. Belief against measurement | Low | Mixed | Developers believed they were faster when they were slower. No adequate office comparison. |
| 6. Does the time become output | Low | Mixed | More resolutions per hour in one customer-support firm. Precise nulls on pay and hours in Denmark. No office study measured output. |
| 7. Where the time goes | Low | Descriptive | Logs show email time falling. Most users say they moved time to other work. |
| 8. Checking, errors and rework | Low | Mixed | Error rates measured mainly in clinical and legal text. Checking time almost never timed. |
| 9. Admin and strain | Low in clinicians; very low elsewhere | Mixed | Small falls in exhaustion with doctors’ scribes. Nothing controlled outside medicine. |
| 10. Who gains | Low; very low for disability and neurodivergence | Mixed | Less experienced workers gain most within a task. |
| 11. HR, payroll, reward and benefits administration | Very low | Unknown | No controlled evaluation found. |
Who uses it
The Office for National Statistics reports that self-reported AI use in UK businesses with 10 or more employees has risen from around 12 per cent to around 35 per cent since late 2023 C1. Its survey counts any use of any AI technology on a list, and it leaves out public administration, so civil service and NHS HR are outside it. The government’s 2026 adoption survey found administration the most common area of use, at 81 per cent of businesses using or planning to use AI C2. Its HR item mixes current use with plans in the same way and is read from chart labels, so it is not quoted here. Eurostat asked European firms about AI for HR management or recruiting once, in 2021, when 0.62 per cent of enterprises reported it. The CIPD’s autumn 2025 survey of 2,019 senior HR professionals and decision-makers found employees in three-quarters of UK organisations using AI tools at work C62. That counts use by anyone in the organisation and says nothing about HR teams’ own use. A survey of 327 European HR professionals, published on the CIPD’s website by a European HR association’s working group, found 19 per cent describing AI as integrated into HR workflows and 66 per cent with no clear guidelines for using it in HR C63.
Grade: low, descriptive. Large official samples with low response rates measure whether a business uses AI for something. No official survey measures whether HR teams use it.
What HR and reward people expect
The figure most often put to HR and reward teams is that 49 per cent of those surveyed believe AI in HR will free up time for more strategic work C3. It comes from a benefits technology firm’s survey, summarised on a reward and benefits association’s website. The summary reports the belief clearly. It gives no sample size, countries, dates or question wording, and the full report sits behind a download form that refused the request. Other expectation surveys come almost entirely from firms that sell HR, benefits or consultancy services, mostly in the US, and most were seen only as search extracts. Some from the same sources point the other way: the association’s own 2024 survey with a benefits consultancy, also seen only as an extract, found 59 per cent expecting AI for benefits management to be not or less important over the next two years. The one academic study of UK HR found Civil Service HR staff personally ready for AI and the function not ready C4.
Grade: very low, descriptive. These surveys describe what respondents expect. Expectations are not evidence of effect in either direction.
Controlled task experiments
In randomised experiments, people given an AI assistant finish short writing tasks faster and graders rate the output higher. In the best known, published in Science, average time fell by 40 per cent and graded quality rose by 18 per cent among 453 college-educated professionals C5. Its working paper drew participants from six occupations, one of them human resource professionals C6, who were a small minority, on 20 to 30 minute tasks such as press releases and delicate emails. Graders scored writing, content and originality, and no check of factual accuracy was found. The authors say their short, self-contained tasks may inflate the estimates C7. A UK experiment with a largely representative online sample found gains of a similar kind on three short text tasks C10.
Where accuracy was checked, the results are mixed. Management consultants using AI completed 12.2 per cent more tasks, 25.1 per cent more quickly, on tasks inside the tool’s capability C8, with quality scores 33.9 and 29.9 per cent above control. The figure of 40 per cent higher quality that circulates for this study was not found in the version read for this review. On a task outside the tool’s capability, consultants with AI were 19 percentage points less likely to be correct C9. In the nearest UK office test, the Department for Business and Trade timed 6 users and 5 non-users on fictional tasks and had the outputs scored blind C11. Users summarised reports faster and more accurately, and did data analysis more slowly and less accurately, the opposite of what their diaries said C12. As the department says, eleven people cannot support a conclusion.
Grade: moderate for speed and rated quality on short tasks, improves; low for accuracy, mixed. No randomised task experiment has been run in HR, payroll, reward or benefits administration.
Time saved at work
Four UK central government evaluations of the same AI assistant are public, alongside a Home Office pilot, a teachers’ trial and evaluations in the NHS and social care. Most of their time figures are what users said.
The cross-government experiment’s 26 minutes a day C13 is an average of the midpoints of the ranges respondents picked, with the top range counted as 60 minutes C14. It comes from 7,115 responses to 20,000 licences, with no comparison group. The Department for Business and Trade’s diary asked staff how long each task took and how long it would have taken without the tool, which is also self-report. A figure of 43 minutes a day for NHS staff comes from a press release with no published method C24. Social care staff in one council self-reported saving about 2.7 hours per assessment, from 29 staff who opted in, with data captured by the tool’s maker C25.
Several evaluations went further than asking. The Department for Work and Pensions compared 1,716 licence holders with 2,535 matched non-users and estimated 19 minutes a day, 95 per cent interval 17 to 22, against 10 unadjusted C16. The time was still self-reported, in bands, and licences went to volunteers and nominees. HMRC allocated licences at random, and its participants self-reported saving 2 to 3 per cent of the working week C17. In a Home Office asylum casework pilot, caseworkers given AI summaries reviewed transcripts 23 minutes quicker than a comparison group, a 32 per cent saving C18, and the Home Office says all its methods relied on self-report C19. The strongest UK design with a time outcome is a cluster randomised trial in 68 English schools. Teachers offered an AI assistant recorded lesson-preparation time at 69 per cent of the control group’s, interval 53 to 90 per cent, in their own diaries C20. In the worst cases, reformatting cancelled out or outweighed the saving C21.
One randomised trial took time from software logs. Across 66 firms, workers offered an AI assistant cut their email time by 1.4 hours a week, 12 per cent C22, and no shift in the quantity or composition of their tasks was detected C23. Three of the four authors work for the tool’s maker. An NHS evaluation of an ambient voice tool reports emergency department documentation falling from 12 to 6 minutes a task C26. That figure came from record logs at a site that was not observed, the 47 minutes a shift is modelled, and there was no concurrent control.
The Department for Business and Trade’s evaluation is often cited for its sentence that it did not find evidence that time savings led to improved productivity C27. The same report says productivity was not a key aim and limited data were collected, so the sentence is no evidence of effect. It is not evidence of no effect. Nor is it the strongest UK design: the Department for Work and Pensions’ matched comparison, HMRC’s random allocation and the teachers’ cluster randomised trial are all stronger.
Grade: low, improves on specific tasks. Savings are mostly self-reported and mostly from volunteers. Several UK evaluations used a comparison group or randomisation, and all of them recorded time by self-report or participants’ own diaries. The one log-based trial is vendor-linked and found time falling on email. No government evaluation followed users for more than six months.
Belief against measurement
One randomised trial set what people believed against what was measured, in software development. Sixteen experienced open-source developers forecast that AI would cut their completion time by 24 per cent, and afterwards estimated it had cut it by 20 per cent. Measured completion time, logged by the developers and checked against screen recordings for some tasks, rose by 19 per cent C28. A follow-up with later tools produced estimates whose intervals include zero for returning and new developers alike C29, and 30 to 50 per cent of developers withheld tasks they did not want to do without AI C30. It reported no forecasts, so the gap between belief and measurement has not been retested.
In office work, the only check of diaries against observation is the Department for Business and Trade’s eleven observed participants, where observation agreed with the diaries on summaries and emails and contradicted them on data analysis and slides C12. Danish chatbot users self-reported saving about 3 per cent of their hours C31, while registers showed no change in recorded hours C34. Those measure different things, so they are not a test of the gap.
Grade: low, mixed. In one small trial, developers believed they were faster when they were slower. No study has compared logged and self-reported time saved in the same office workers, so for admin there is no evidence either way on the size of the gap.
Does the time become output
In one software firm’s customer support, staggered access to an AI reply-suggestion tool raised issues resolved per hour by 15 per cent on average C32. The 14 per cent often quoted is from the 2023 working paper C33. The tool was in use from late 2020, before the public chatbots, and the outcome is throughput from firm records, the nearest thing to output in this literature.
No study of an AI writing or admin assistant in an office has measured output. The logged trial measured time and detected no shift in the quantity or composition of tasks C23. Danish registers linking about 25,000 workers in exposed occupations, HR professionals among them, show precise nulls on earnings and recorded hours, ruling out effects larger than 2 per cent C34. That is evidence of no effect on pay and recorded hours, which are not output. Aggregate figures rest on self-report. US users reported time savings of 5.4 per cent of their work hours, and the 1.1 per cent quoted alongside it is a modelled productivity gain C35. Employment of 22 to 25 year olds in AI-exposed US occupations is 19 per cent below the path of less-exposed peers, which the authors call descriptive indicators rather than causal estimates C36. The Bank of England’s November 2026 Monetary Policy Report gives UK firms’ own view. Firms in its Decision Maker Panel survey estimate that AI has raised productivity by around 0.5 per cent over the past three years C61, and report close to zero effect on employment. The Bank describes the evidence of an AI-driven rise in UK productivity growth as modest. These are firms’ estimates, so they are self-report too.
Grade: low, mixed. One firm shows more output per hour, and Danish registers show no effect on pay or hours. For office and admin work there is no evidence of effect on output, because no study has measured it.
Where the time goes
The logged trial gives the only measured answer, and only inside the tool maker’s own applications. Email time fell C22 and no other shift in tasks was detected C23. Time that moved elsewhere, or into rest, was not observed.
The rest is what people said. In the Department for Work and Pensions trial, 89 per cent of responding users said they moved saved time to other work tasks C40. In Denmark, fewer than 10 per cent of users said they would expect to take more breaks or leisure if a chatbot saved them time C37. The Department for Business and Trade counted tasks done only because the tool was there, and outputs that went unused, as negative time C46. New work appears too. In workplaces without an employer initiative, 12 per cent of Danish users reported new workloads from chatbots C39, and 35 per cent of the new AI-related tasks they described were quality review or ethics and compliance C38. Researchers who observed one US technology firm for eight months report that AI tools intensified work C41. The cross-government report could not identify where the time went C15.
Grade: low, descriptive. No study has measured total working time, rest or recovery objectively after adoption, and none was in HR.
Checking, errors and rework
Error rates have been measured, mostly in clinical and legal text and often by the tool’s maker. Clinicians annotating notes drafted from mock consultations found hallucinations in 1.47 per cent of sentences, 44 per cent of them major C42, in a study whose authors all worked for a clinical scribe company. Specialist legal research tools hallucinated between 17 and 33 per cent of the time C43. In the Home Office pilot, specialists screened every AI summary and removed 9 per cent as inaccurate or incomplete before anyone used them C44, so the pilot did not measure what happens when staff act on an unscreened summary. In the Department for Business and Trade’s diary, 22 per cent of respondents said they had found hallucinations C45, and scheduling and generating images took longer with the tool than without C46.
Checking time itself is almost never timed. A government press release reports experts taking 22 hours to check the themes an AI tool assigned to consultation responses, and the evaluation reports behind it could not be opened. No study opened for this review measured the cost of acting on a wrong AI output in administration.
Grade: low, mixed. Error rates are measurable and depend on the task. No study has measured checking time or error rates in HR, payroll, reward or benefits administration.
Admin and strain
The best evidence that taking admin away eases strain comes from doctors’ ambient AI scribes. In a stepped-wedge randomised trial with 66 practitioners over 24 weeks, a scribe reduced work exhaustion by 0.44 points on a five-point scale C47. Practitioners had to be willing to adopt the tool C48, the trial was open-label, and its co-primary outcome, professional fulfilment, was not significant at the trial’s threshold. In a trial of 238 physicians over about two months, single-item burnout showed no clear effect with either of two scribes, odds ratios 0.78 and 0.67 with intervals crossing 1 C49. A 2026 meta-analysis rated certainty low for less burnout and moderate for less cognitive workload C50. No randomised trial measured intention to leave. One uncontrolled study found 28.3 per cent before and 26.6 per cent after 84 days C51, a null at very low certainty.
Outside medicine, no controlled study has measured burnout after AI took over admin. The teachers’ trial measured time and no wellbeing outcome C20. One cross-sectional survey points the other way: radiologists who used diagnostic AI regularly had higher odds of burnout C52, though that AI adds a task to the day rather than removing one.
Grade: low for clinician burnout, mixed; very low for any other setting. Clinical notes are checked by trained clinicians under a professional duty, which HR outputs may not be. No trial followed strain beyond 24 weeks.
Who gains
Within a task, less experienced workers gain most. In the customer-support firm, less skilled and less experienced agents resolved about 30 per cent more issues per hour C53, and the writing experiment reports that inequality between workers fell. Beyond the task, employment of younger workers in exposed US jobs has fallen behind C36, and Danish women were 16 percentage points less likely than men in the same occupation to have used a chatbot for work C54. Evidence on disability and neurodivergence is limited to satisfaction ratings. In the Department for Business and Trade’s evaluation, neurodiverse respondents were more likely to be satisfied, at the 90 per cent confidence level only C55, and tests for other conditions were not significant.
Grade: low, mixed; very low for disability and neurodivergence. No study has measured time, output, strain or retention by disability, or any gradient by seniority in HR teams.
HR, payroll, reward and benefits administration
No controlled evaluation of AI time savings, accuracy or staff outcomes in HR shared services, payroll, reward or benefits administration was found in 50 logged searches of four research databases and the web. What exists is company-reported or uncontrolled. In a case study of one multinational, an interviewee said its HR bot had dealt with 11,000 questions with 81 per cent accuracy C57, with accuracy undefined. After an NHS payroll provider launched a payslip app with a chatbot in 2018, service desk cases fell to an average of 9,474 a month from 12,808, across two adjacent three-month windows with no control C58. A synthetic payroll benchmark found models near-exact on simple cases, while on very complex ones three of four models had at most 29 per cent of rows within five cents and one reached 100 per cent. The UK government evaluations mention HR only through interviews C60 and a focus group.
The nearest controlled evidence is from a neighbouring field. In US unemployment insurance adjudication, examiners rated an AI fact-finding system highly, but it did not improve quality or efficiency against the sandbox control group C56. Only the abstract could be read, and these are public welfare benefits. In the Danish registers, no occupation, HR professionals included, showed a significant change in earnings C59, an estimate that is imprecise for HR alone.
Grade: very low, direction unknown. There is no evidence of effect in either direction, because no controlled study has measured time, accuracy, checking or strain in these teams.
The hypothesis, tested
| Part of the hypothesis | Verdict | Why |
|---|---|---|
| In controlled experiments on single writing tasks, AI makes people faster and the output is rated better | Holds | Speed and rated quality improve across randomised experiments C5C8. Accuracy was rarely checked, and it fell on a task beyond the tool C9. |
| In real workplaces, savings are mostly self-reported, come from volunteers and have no comparison group | Too strong on comparison groups | The Department for Work and Pensions C16, the Home Office C18, HMRC C17 and the teachers’ trial C20 used a comparison group or randomisation. All still recorded time by self-report. |
| Where time was measured against a control, the saving is smaller than people report | Not established for office work | Shown once, in developers C28. The only office comparison has eleven people C12. No evidence either way for admin. |
| No workplace evaluation has shown saved time turning into more output | Holds for office and admin work | One customer-support firm showed more output per hour C32. No office evaluation measured output. |
| None has tracked where the time went | Too strong | Logs tracked email time C22, and surveys asked C37C40. No study measured total working time or rest. |
| The best evidence that removing admin reduces strain is from doctors’ scribes, with small effects at low certainty | Holds | C47C49C50 |
| No controlled study has tested AI admin tools in HR, payroll, reward or benefits administration | Holds | Only company-reported and uncontrolled figures exist C57C58. The nearest trial is in public benefits C56. |
Put together, AI makes people faster on short controlled writing tasks, and in real workplaces the time saved rests mostly on self-report, even where there is a comparison group. The one log-based trial found time saved on email. No office study has measured whether saved time became more output, evidence on strain is confined to low-certainty trials of doctors’ scribes, and HR, payroll, reward and benefits administration has no controlled evaluation.
The five claims
| Claim | Verdict | Deciding evidence |
|---|---|---|
| AI will free up HR and reward teams’ time for more strategic work, which 49 per cent of them believe | Unsupported | The 49 per cent is a belief, quoted accurately, from a survey with no stated sample C3. No HR evaluation has measured where time goes, and the cross-government report could not identify it C15. |
| AI saves a stated number of minutes a day, such as 26 for civil servants | Overstated | Self-reported range midpoints C13C14. With a comparison group, 19 minutes, still self-reported C16. Logged: 1.4 hours a week less email, in a vendor-linked trial C22. |
| Time saved on admin becomes productivity | Unsupported | No office study has measured it, and the Department for Business and Trade did not aim to C27. Danish registers rule out effects on earnings and recorded hours above 2 per cent C34. The one output gain is in customer support C32. |
| Taking admin away with AI reduces burnout or improves wellbeing | Overstated | Clinicians only: small short-term falls in self-reported exhaustion at low certainty C47C50, and no clear effect on single-item burnout C49. No controlled study outside medicine. |
| AI admin tools are accurate enough that checking does not eat the saving | Unsupported | No HR study has timed checking, and few studies anywhere have. The Home Office removed 9 per cent of summaries before use C44, and observed users were slower and less accurate at data analysis C12. |
Claims in circulation
The research run checked 48 claims found in circulation: 7 supported, 21 overstated, 3 misattributed, 5 unsupported and 12 unverifiable. The ones that matter most here:
- The customer-support figures of 14 and 34 per cent are from the working paper C33. The published paper reports 15 per cent C32 and about 30 per cent for less experienced workers C53.
- The St Louis Fed’s 1.1 per cent is a modelled gain in productivity. It is not a share of hours saved C35.
- The consultants’ 40 per cent higher quality was not found in the version read for this review, which reports 33.9 and 29.9 per cent C8.
- The written ministerial statement that introduced the cross-government report said many users indicated their productivity gains allowed them to spend more time on strategic and satisfying tasks. The report itself said it was not possible to identify how time saved was spent C15.
- A figure of a 35 per cent fall in clinicians feeling overwhelmed, circulating for the NHS ambient voice evaluation, is not in the report’s text. A claim that disabled or neurodivergent civil servants saved 25 per cent more time was not found in either government report opened.
- A benefits technology blog’s figures of 30 per cent faster processing and 40 per cent fewer errors in benefits administration are attributed only to unnamed studies.
What the evidence does not support
- That a minutes-a-day figure from a self-report survey is a measured saving, or that it carries over from civil servants to an HR team.
- That time saved by an AI assistant has become more output in any office or admin setting. No study has measured it, and the pay and recorded-hours data that exist show no effect.
- That saved time goes to strategic work. Most users say they moved it to other work. No study measured whether that work was strategic.
- That AI admin reduces burnout outside medicine. No controlled study has measured it.
- That checking is negligible. Error rates are measurable and depend on the task, and checking time is almost never measured.
- Any time, accuracy or staff-outcome figure for HR, payroll, reward or benefits administration from a controlled design.
- Conversions of minutes a day into days a year, money or headcount. None of the evaluations measured these, and this review does not compute them.
- That these results transfer to the tools sold now, in either direction.
What would change the verdict
- A randomised or stepped-wedge evaluation in an HR shared service, payroll or benefits team, with time taken from case-management or software logs rather than self-report.
- Any office evaluation that measured output, such as cases closed or errors per case, alongside time and against a control.
- A same-person comparison of self-reported and logged time saved in office work, large enough to estimate the gap with useful precision.
- A controlled study of an AI admin tool outside medicine, with a validated burnout or workload measure and follow-up beyond six months.
- Measured checking time and field error rates for AI-drafted administrative text, with the cost of the errors that reached employees.
- The full text of the unemployment insurance adjudication trial C56, which could move the HR tier if its effects are precise.
- Evaluations of current tools, which could supersede much of what is here.
Five things an HR lead should know before buying
- Most time-saved figures, including the 26 minutes a day quoted for civil servants, are what users said they saved C13C14. Ask whether time was logged or observed, and against what comparison group.
- Where time was logged in a randomised trial, the saving it found was on email C22, and no other shift in work was detected C23. The tool’s maker co-ran that trial, so treat it as indicative.
- No office evaluation has measured whether saved time became more output. Danish registers, which include HR professionals, show no effect on pay or recorded hours C34C59.
- AI-drafted text contains errors at measurable rates C42C44, and some tasks took longer with the tool C46. Checking time is rarely measured, so test accuracy on your own cases and budget time for review.
- The evidence that removing admin reduces burnout comes from doctors using note-taking scribes, with small effects at low certainty, followed for weeks C47C50. No controlled study has tested it in HR, payroll or benefits teams.
The study worth running next
A stepped-wedge cluster randomised trial across HR shared-service or benefits administration teams. Teams would get the AI assistant in a random order over six to nine months, so each acts as its own control before it switches. The design would be pre-registered and run by an evaluator with no commercial link to the tool.
It would measure handling time per case from case-management and software logs, with self-reported time collected alongside so the gap can be estimated. Output would be cases closed per staffed hour and reopened queries, with a blind audit of a random sample of assisted and unassisted outputs against source records for factual and policy errors. Checking would be the logged time spent reviewing and correcting drafts, and the errors that reached employees. Short time-use diaries on sample days would show where the time went. Strain would be a validated burnout and workload measure at baseline, three, six and nine months, with intention to leave. Experience, role, and disability or neurodivergence would be pre-specified subgroups. The customer-support study measured output from routine firm records C32 and the adjudication trial tested an assistant on casework C56, so these outcomes can be measured in practice.
Methods and limits
The search and first grading were done on 17 November 2026 by AI agents working in five parallel tracks: government evaluations, output and time use, strain, HR administration and claims, and task experiments and errors. They logged 145 searches across OpenAlex, Crossref, Europe PMC, PubMed, arXiv, NBER, SSRN, gov.uk, the Parliament written statements API, other government sites, professional bodies and the web, and opened 159 source records. Of these, 102 were read at full text, 14 at abstract only and 4 as press releases only, 30 were seen as summary pages or search-engine extracts, and 9 were blocked. Material from anyone selling AI, HR or benefits technology, or consultancy, sat in a separate industry column and set no grade. Department-run evaluations of tools the department bought are labelled as such, and studies with authors or funding from a tool’s maker are labelled vendor-linked.
I run a benefits broker, and my company has been rolling out AI features that do admin, so I have a commercial stake in the answer. The brief for this review asked for findings favourable to such tools to meet a higher bar.
Some things were not done, and the grades they touch are provisional.
- The Bank of England’s November 2026 Monetary Policy Report box and the CIPD material were read in a browser on the evening of 17 November, after the research run had finished. They are in the ledger as C61 to C63, and they were not part of the run’s graded search or its verification sample.
- The full report behind the 49 per cent figure is gated, so its sample, recruitment and question wording are unknown.
- Several sources that matter were read at abstract only: the published Science version of the writing experiment (its working paper was read in full), the unemployment insurance adjudication trial, three smaller scribe trials and three HR papers. The ledger marks each one.
- About 25 trade-press and vendor pages in the HR track were seen only as search-engine extracts, because the sites blocked direct fetching. They are labelled, sit in the industry column, set no grade and are not in the ledger.
- The Australian whole-of-government evaluation could not be opened and feeds no grade. The Scottish, Welsh and Northern Ireland governments, and New Zealand, were searched only lightly, so finding nothing there is provisional.
Verification
A seeded random 30 per cent of cited source ids was drawn with seed 20261118, in two waves because one track finished later: 36 of 118, then 8 of 24, so 44 of 142 in all. Separate verifier agents reopened each source from its citation and link alone. 26 passed, 18 partly, and none failed. Seven of the partial results were partial only because the verifier could not open the page or the full text. The other eleven led to corrections, all applied. The most material was the NHS emergency department figure, which had been described as measured by observers when it came from record logs at a site that was not observed, with the per-shift figure modelled C26.
On 17 November 2026 every passage in the ledger below was checked by machine against the passage stored for its source, and all 60 matched word for word. Each cited page was then fetched again. 43 passages were found on the live page or the document it hosts, none was missing, and 17 could not be checked because the publisher refused the request or did not respond. Where only an abstract or a press release could be read, the ledger says so. Claims C61 to C63 were added later the same evening. Their passages were read on the Bank of England and CIPD websites, and the CIPD report’s sentence was checked against its PDF.
Claims ledger
Each claim is worded exactly as it was checked. Assessed claims show the source, passage, limits and date checked. If a source could not be checked, the entry records what was attempted and why. Claims still being checked are marked in review.
- C1Supported
The Office for National Statistics reports that self-reported AI use in UK businesses with 10 or more employees has risen from around 12 per cent to around 35 per cent since late 2023.
- Source
- Office for National Statistics (2026). Artificial intelligence in UK businesses: 2023 to 2026. 20 July 2026. Open source (full text read)
- Passage
The self-reported use of artificial intelligence (AI) in UK businesses with 10 or more employees has increased from around 12% to around 35% since late 2023
- Where
- Main points. The latest wave, 5 to 28 June 2026, had 38,637 responses, a response rate of 26.7 per cent
- Limits
- A voluntary business survey that counts any use of any AI technology on a list. It excludes public administration and defence, public health and education, and finance and insurance, so civil service and NHS HR are outside it. It has no item on HR teams.
- Checked
Back to text↩ 1
- C2Supported
In the government's 2026 AI adoption survey, administration was the most common business area for AI, at 81 per cent of businesses using or planning to use it.
- Source
- Department for Science, Innovation and Technology (2026). Artificial intelligence adoption research 2026 (AI adoption survey report 2026), research by BMG Research. October 2026. GOV.UK. Open source (full text read)
- Passage
most commonly in administration (81% of those using or planning to use AI)
- Where
- Executive summary
- Limits
- Telephone survey of 3,503 private businesses with 5 or more staff, response rate 5 per cent. The question combines current use with plans, and businesses with neither are left out of the base. The survey's HR figure is read from chart labels and is not quoted here.
- Checked
Back to text↩ 1
- C3Supported
A survey summary from a benefits technology firm, hosted on a reward and benefits association's website, says 49 per cent of the HR and reward professionals surveyed believe AI in HR will free up time for more strategic work.
- Source
- REBA (2025). Benifex report: AI in benefits and reward. REBA resource page, 28 May 2025. Open source (full text read)
- Passage
Nearly half (49%) of the HR and reward professionals surveyed in the report believe AI in HR will free up time for more strategic work
- Where
- Resource page dated 28 May 2025
- Limits
- A belief, quoted accurately. The page gives no sample size, countries, dates, recruitment method or question wording. The full report sits behind a download form, and the request for it was refused (HTTP 403), so none of these could be checked. The firm sells benefits, reward and recognition technology.
- Checked
- C4Supported
A survey of UK Civil Service HR staff suggested the function was not ready to use AI, while staff were personally ready.
- Source
- Smith, M., Prabhakar, G., Nisar, T., & Tseng, H. (2024). Adoption of AI by the HR function in the civil service. Employee Relations, 47(1):104-126. Open source (abstract only)
- Passage
While the results suggest that the CSHR is not ready to harness AI opportunities, employees were personally ready
- Where
- Abstract, findings
- Limits
- Abstract only; the full text was blocked. Readiness and intentions, not use or outcomes.
- Checked
Back to text↩ 1
- C5Supported
In a randomised experiment published in Science, the average time taken on short writing tasks fell by 40 per cent and graded output quality rose by 18 per cent.
- Source
- Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192. Open source (abstract only)
- Passage
The average time taken decreased by 40% and output quality rose by 18%.
- Where
- Abstract
- Limits
- Abstract only; the publisher's full text and supplement could not be opened, so whether the published version checked factual accuracy is unverified. Quality was graded by evaluators.
- Checked
- C6Supported
The working paper behind that experiment drew its participants from six occupations, one of them human resource professionals.
- Source
- Noy, S., & Zhang, W. (2023, 2 March). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Working paper (not peer reviewed), MIT. Open source (full text read)
- Passage
The occupations we draw on are marketers, grant writers, consultants, data analysts, human resource professionals, and managers.
- Where
- Introduction; Table 1 gives HR professionals as 6 per cent of the control group and 11 per cent of the treatment group
- Limits
- Working paper with 444 participants, against 453 in the published version. Evaluators graded writing quality, content quality and originality; no check of factual accuracy was found. No task was an HR task.
- Checked
Back to text↩ 1
- C7Supported
The authors say their tasks were short, self-contained and lacked context-specific knowledge, which may inflate their estimates.
- Source
- Noy, S., & Zhang, W. (2023, 2 March). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Working paper (not peer reviewed), MIT. Open source (full text read)
- Passage
the tasks are relatively short, self-contained, and lack a dimension of context-specific knowledge, which may inflate our estimates
- Where
- Conclusion
- Limits
- The authors' own caution. The tasks took 20 to 30 minutes.
- Checked
Back to text↩ 1
- C8Supported
In a randomised experiment with management consultants, those using AI completed 12.2 per cent more tasks, 25.1 per cent more quickly, on tasks inside the tool's capability.
- Source
- Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403-423. Open source (full text read)
- Passage
completing 12.2% more tasks and completing them 25.1% more quickly on average while also delivering solutions of significantly improved quality.
- Where
- Abstract; the results give quality scores 33.9 and 29.9 per cent above control in the two AI arms
- Limits
- Inside the tool's capability, quality was graded on subjective rubrics because, the authors say, no single correct answer exists. Three authors are at the consultancy's research institute and the consultancy helped design the tasks. A figure of 40 per cent higher quality that circulates for this study is not in the published paper.
- Checked
- C9Supported
On a task outside the tool's capability, consultants using AI were 19 percentage points less likely to produce correct solutions.
- Source
- Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403-423. Open source (full text read)
- Passage
However, for a task outside the frontier, subjects using AI were 19 percentage points less likely to produce correct solutions.
- Where
- Results; the control group was correct about 84.5 per cent of the time and the AI groups 60 and 70.6 per cent
- Limits
- Accuracy was checked on this one task only. The published abstract words the same result as 19 per cent less likely.
- Checked
- C10Supported
In a UK experiment with a largely representative online sample of 1,041 working-age adults, predicted time savings from AI access ranged from 0.32 to 0.69 standard deviations and performance gains from 0.41 to 0.92 on three short text tasks.
- Source
- Haslberger, M., Gingrich, J., & Bhatia, J. (2023, 6 October). No Great Equalizer: Experimental Evidence on AI in the UK Labor Market. Working paper (SSRN 4594466). Open source (full text read)
- Passage
Predicted time savings range from 0.32 to 0.69 standard deviations (SD), while performance increases amount to between 0.41 and 0.92 SD.
- Where
- Section 4.1; fieldwork 19 to 28 July 2023
- Limits
- Working paper. Three very short tasks, one session.
- Checked
Back to text↩ 1
- C11Supported
In the Department for Business and Trade's observed tasks, a second analyst scored outputs blind for quality and accuracy, and the facilitator recorded the time each task took.
- Source
- Department for Business and Trade (2025, August). The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade. Open source (full text read)
- Passage
Outputs of the observed task sessions were blind assessed by a secondary analyst to score each output on quality and accuracy, and the time taken for each task was recorded by the interview facilitator.
- Where
- Methodology, observed tasks; Table 6 gives summarising reports 12:37 against 41:34, emails 07:30 against 07:43, data analysis 25:01 against 20:33 with accuracy 1.5 against 2.7 out of 5, and slides 10:47 against 18:30 with accuracy 1.5 against 5
- Limits
- 6 users and 5 colleague-nominated non-users, not randomised, on fictional one-hour tasks after the pilot. The department says the small samples limit the findings.
- Checked
Back to text↩ 1
- C12Supported
Observed users did data analysis more slowly and less accurately than non-users, which conflicted with the time savings reported for data analysis in the diary.
- Source
- Department for Business and Trade (2025, August). The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade. Open source (full text read)
- Passage
data analysis more slowly and to a worse quality and accuracy than non-users, conflicting time savings reported in the diary study for data analysis.
- Where
- Findings, observed tasks; the sentence opens by naming the tool's users and the spreadsheet software
- Limits
- 11 people in all. The report's conclusion describes the observed tasks as showing similar trends to the diary.
- Checked
- C13Supported
The cross-government experiment reports that participants saved an average of 26 minutes a day.
- Source
- Government Digital Service (2025). Microsoft 365 Copilot Experiment: Cross-Government Findings Report. 2 June 2025. gov.uk (HTML and PDF). Open source (full text read)
- Passage
Trial participants saved an average of 26 minutes a day
- Where
- Section 1.2, key findings; the sentence ends by naming the tool
- Limits
- Self-reported. 7,115 survey responses from 20,000 licences, no comparison group and no baseline. Fieldwork 30 September to 31 December 2024.
- Checked
- C14Supported
The 26 minutes is an average of the midpoints of the ranges respondents chose, with the top range counted as 60 minutes.
- Source
- Government Digital Service (2025). Microsoft 365 Copilot Experiment: Cross-Government Findings Report. GOV.UK. Open source (full text read)
- Passage
This estimate was calculated by using the middle point of each range, with the largest savings being estimated at 60 minutes.
- Where
- Section on time savings, Figure 5 commentary; the preceding sentence converts the figure to 13 days a year
- Limits
- Self-reported ranges converted at midpoints. The annual conversion is the report's own.
- Checked
- C15Supported
The cross-government report says it was not possible to identify how the time saved was spent.
- Source
- Government Digital Service (2025). Microsoft 365 Copilot Experiment: Cross-Government Findings Report. Open source (full text read)
- Passage
due to experimental constraints it was not possible to identify how time saved was spent.
- Where
- Section 6, Conclusions
- Limits
- The report also records an HR focus group, without an HR time figure.
- Checked
- C16Supported
The Department for Work and Pensions estimated that licence holders saved an average of 19 minutes a day across eight routine tasks, against a matched comparison group of non-users.
- Source
- Arzilli, F., Lynch, C. S., & Page, L. (2026). An Evaluation of DWP’s Microsoft 365 Copilot Trial. Department for Work and Pensions. 29 January 2026. gov.uk (HTML). Open source (full text read)
- Passage
The analysis estimated that users in the treatment group saved an average of 19 minutes per day across eight routine tasks
- Where
- Chapter 4, task efficiency; Table 4.3 gives 19 minutes, 95 per cent interval 17 to 22, and the unadjusted estimate as 10 minutes, 7 to 12
- Limits
- Self-reported time bands from 1,716 licence holders and 2,535 non-users. Licences went to volunteers and nominees, there was no baseline, and the authors say they cannot assume all bias from voluntary uptake has been addressed.
- Checked
- C17Supported
HMRC participants self-reported saving 2 to 3 per cent of their working week, around 60 minutes a week for a full-time worker.
- Source
- HM Revenue and Customs (2026). Evaluating the Impact of Microsoft Copilot in HMRC (Evaluation report, phase 3 trial of Microsoft Copilot). 9 July 2026. gov.uk (HTML summary). Open source (full text read)
- Passage
on average, HMRC participants self-reported saving 2-3% of their working week, equating to around 60 minutes per week for a full-time equivalent working pattern
- Where
- Results, section 2, time savings
- Limits
- Licences were allocated at random, the only UK government licence evaluation found to do so, but the time figure is self-reported and the published summary reports no non-user comparison. The full methods are not published.
- Checked
- C18Supported
In a Home Office asylum casework pilot, the test group reviewed interview transcripts 23 minutes quicker than a comparison group, a 32 per cent time saving.
- Source
- Home Office (2025, 29 April). Evaluation of AI trials in the asylum decision making process. Research note, GOV.UK. Open source (full text read)
- Passage
On average, the test group reviewed transcripts 23 minutes quicker than the comparison group (a 32% time saving).
- Where
- Section 4.1
- Limits
- Test and comparison groups were chosen for availability, not at random. Time was recorded by participants (next claim).
- Checked
- C19Supported
The Home Office says all its evaluation methods used self-reporting.
- Source
- Home Office (2025, 29 April). Evaluation of AI trials in the asylum decision making process. Research note, GOV.UK. Open source (full text read)
- Passage
all the evaluation methodologies use self-reporting, so may reflect inaccuracies in recall and biases associated with those that chose to respond
- Where
- Section 7, limitations
- Limits
- The authors' own caution.
- Checked
Back to text↩ 1
- C20Supported
In a cluster randomised trial in English schools, teachers offered an AI assistant recorded lesson-preparation time at 69 per cent of the control group's, 95 per cent interval 53 to 90 per cent.
- Source
- Roy P, et al.; National Foundation for Educational Research for the Education Endowment Foundation (2024). ChatGPT in Lesson Preparation: A Teacher Choices Trial. Evaluation Report (ERIC ED680921). Open source (full text read)
- Passage
69% (53%, 90%) 4.68 (3.88, 5.65) 6.79 (5.66, 8.16) 211 (43) 0.006
- Where
- Executive summary table (columns: time as a share of control with 95 per cent interval; adjusted mean hours over five weeks for each group; teachers analysed, with missing; P). 68 schools were randomised
- Limits
- Self-completed workload diaries over five weeks. Teaching, not office administration. No wellbeing outcome was measured. Read from the copy held by ERIC because the funder's site blocked the request.
- Checked
- C21Supported
The trial report says that in the worst cases, reformatting or finding images cancelled out or outweighed the time saved.
- Source
- Roy P, et al.; National Foundation for Educational Research for the Education Endowment Foundation (2024). ChatGPT in Lesson Preparation: A Teacher Choices Trial. Evaluation Report (ERIC ED680921). Open source (full text read)
- Passage
In the worst cases, reformatting or finding images negated or even outweighed the time saved
- Where
- Implementation findings; the sentence ends by naming the tool
- Limits
- Interview and survey findings, not timed.
- Checked
Back to text↩ 1
- C22Supported
In a randomised field experiment with time taken from software logs, workers offered an AI assistant cut their email time by 1.4 hours a week, a 12 per cent reduction (intention to treat).
- Source
- Dillon, E. W., Jaffe, S., Immorlica, N., & Stanton, C. T. (2025, revised November 2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795. doi:10.3386/w33795. Open source (full text read)
- Passage
email by 1.4 hours per week, a 12% reduction relative to the pre-period mean.
- Where
- Introduction, printed page 2; the sentence opens by naming the tool and the email application. 7,137 workers in 66 firms
- Limits
- Vendor-linked: three of the four authors work at the tool maker's research arm, and the maker's programme recruited the firms. Logs cover only the maker's own applications. Working paper.
- Checked
- C23Supported
The same experiment did not detect shifts in the quantity or composition of workers' tasks.
- Source
- Dillon, E. W., Jaffe, S., Immorlica, N., & Stanton, C. T. (2025, revised November 2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795. Open source (full text read)
- Passage
we do not detect shifts in the quantity or composition of workers’ tasks resulting from individual-level AI provision.
- Where
- Abstract, November 2025 revision
- Limits
- The authors observe no content of work and no measure of productivity or performance, so output was not measured. Vendor-linked.
- Checked
- C24Supported
A government press release says an AI assistant could save NHS staff on average 43 minutes per staff member per day.
- Source
- Department of Health and Social Care and NHS England (2025). Major NHS AI trial delivers unprecedented time and cost savings. Press release, 21 October 2025. gov.uk. Open source (full text read)
- Passage
could save NHS staff on average 43 minutes per staff member per day or more
- Where
- Paragraph 2
- Limits
- Press release only. No method, comparison group or evaluation report has been published, so how the figure was measured is unknown. A joint announcement that quotes the tool's maker.
- Checked
Back to text↩ 1
- C25Supported
In one council's adult social care team, staff reported an average saving of about 2.7 hours per assessment with a note-drafting tool.
- Source
- Health Innovation Kent Surrey Sussex, with Kent County Council Adult Social Care, Beam and Unity Insights (2025). Evaluation of test and learn of Magic Notes in Kent County Council Adult Social Care Team. November 2025. Open source (full text read)
- Passage
Staff reported a time saving for writing care needs assessments and occupational therapy assessments with an average approximate saving of 2.7 hours per assessment (48% reduction).
- Where
- Printed page 5, results
- Limits
- Self-reported, 29 staff who opted in, eight weeks. The tool's maker captured and analysed the data and contributed to the report. The evaluators recommend a time-and-motion study.
- Checked
Back to text↩ 1
- C26Supported
An NHS evaluation of an ambient voice tool reports emergency department documentation falling from 12 to 6 minutes per task, which it equates to 47 minutes a shift.
- Source
- Great Ormond Street Hospital DRIVE Unit (2026). The use of Ambient Voice Technology with Generative Artificial Intelligence in Multiple Clinical Settings Across the NHS (AVT Phase 4 final report). Version 1.3, updated 11 February 2026. Open source (full text read)
- Passage
Use of the AVT tool saved an average of 6 minutes per documentation task (12mins pre), equating to 47 minutes saved per shift.
- Where
- Executive summary, key findings; the emergency department chapter reports a 7.02-minute average reduction
- Limits
- Before-and-after design with no concurrent control. This figure came from health record logs at a site that was not directly observed, and the per-shift figure is a modelled estimate. Clinical documentation, not office admin. The tool's maker developed and provided the app.
- Checked
- C27Supported
The Department for Business and Trade's evaluation did not find evidence that time savings had led to improved productivity.
- Source
- Department for Business and Trade, Digital Data and Technology Monitoring and Evaluation team (2025). The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade. August 2025. gov.uk. Open source (full text read)
- Passage
The evaluation did not find evidence that time savings have led to improved productivity
- Where
- Executive summary, printed page 5. Page 32 says productivity was not a key aim of the evaluation and limited data were collected
- Limits
- No evidence of effect, not evidence of no effect: the evaluation did not set out to measure productivity. 1,000 licences, about 70 per cent to volunteers; 300 diaries, a 32 per cent response rate.
- Checked
- C28Supported
In a randomised trial with 16 experienced open-source developers, they forecast AI would cut completion time by 24 per cent and afterwards estimated it had cut it by 20 per cent; measured completion time rose by 19 per cent.
- Source
- Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089 (v2, 25 July 2025). Open source (full text read)
- Passage
Before starting tasks, developers forecast that allowing AI will reduce completion time by 24%. After completing the study, developers estimate that allowing AI reduced completion time by 20%. Surprisingly, we find that allowing AI actually increases completion time by 19%
- Where
- Abstract, version 2; tasks were done between February and June 2025
- Limits
- Completion times were logged by the developers for each task and checked against screen recordings for a subset. Software development, not admin. Preprint.
- Checked
- C29Supported
In a follow-up with later tools, the estimates for returning and newly recruited developers both had confidence intervals that include zero.
- Source
- Becker, J., Rush, N., Cunningham, T., Rein, D., & Mahamud, K. (2026). We are Changing our Developer Productivity Experiment Design. METR, 24 February 2026. Open source (full text read)
- Passage
For the subset of the original developers who participated in the later study, we now estimate a speedup of -18% with a confidence interval between -38% and +9%. Among newly-recruited developers the estimated speedup is -4%, with a confidence interval between -15% and +9%.
- Where
- Opening section; 57 developers, 10 of them returning
- Limits
- The follow-up reports no forecasts or after-the-fact estimates, so the gap between belief and measurement cannot be updated.
- Checked
Back to text↩ 1
- C30Supported
In the follow-up, 30 to 50 per cent of developers said they were not submitting some tasks because they did not want to do them without AI.
- Source
- Becker, J., Rush, N., Cunningham, T., Rein, D., & Mahamud, K. (2026, 24 February). We are Changing our Developer Productivity Experiment Design. METR blog. Open source (full text read)
- Passage
When surveyed, 30% to 50% of developers told us that they were choosing not to submit some tasks because they did not want to do them without AI.
- Where
- Blog post body
- Limits
- The authors call the new data only very weak evidence for this reason.
- Checked
Back to text↩ 1
- C31Supported
Danish workers who used AI chatbots self-reported saving about 3 per cent of their work hours.
- Source
- Humlum, A., & Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Open source (full text read)
- Passage
On average, adopters in our sample report savings of about 3% of their work hours.
- Where
- Appendix discussion of reported benefits
- Limits
- Self-reported. Set against register hours and earnings in the same study (claim C34), which measure something different.
- Checked
Back to text↩ 1
- C32Supported
In one software firm's customer support, access to an AI reply-suggestion tool increased issues resolved per hour by 15 per cent on average.
- Source
- Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889-942. Open source (full text read)
- Passage
Access to AI assistance increases worker productivity, as measured by issues resolved per hour, by 15% on average
- Where
- Abstract
- Limits
- Read in the arXiv version that matches the published abstract; the journal's text was blocked. One firm, live chat, a tool in use from late 2020, before the public chatbots. Output per hour from firm records. The resolution rate rose by a non-significant 1.3 points and customer satisfaction did not change.
- Checked
- C33Supported
The November 2023 working paper reported 14 per cent on average and 34 per cent for novice and low-skilled workers.
- Source
- Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at Work. NBER Working Paper 31161, revised November 2023. Open source (full text read)
- Passage
by 14% on average, including a 34% improvement for novice and low-skilled workers
- Where
- Abstract, page 2 (spacing as extracted)
- Limits
- Superseded by the published figures: 15 per cent on average (claim C32) and about 30 per cent for less experienced workers (claim C53).
- Checked
- C34Supported
Linking surveys of Danish workers in exposed occupations to national registers, the authors estimate precise null effects on earnings and recorded hours, ruling out effects larger than 2 per cent.
- Source
- Humlum, A., & Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Previously circulated as "Large Language Models, Small Labor Market Effects". Open source (full text read)
- Passage
we estimate precise null effects on earnings and recorded hours at both the worker and workplace levels, ruling out effects larger than 2%
- Where
- Abstract, March 2026 revision; section 3 adds that dynamic estimates rule out changes larger than 3 per cent
- Limits
- Working paper. Earnings and recorded hours are not output, and unpaid or unrecorded work is not captured. Registers run to December 2024.
- Checked
- C35Supported
In a US survey, generative AI users reported average time savings of 5.4 per cent of their work hours in November 2024.
- Source
- Bick, A., Blandin, A., & Deming, D. (2025, 27 February). The Impact of Generative AI on Work Productivity. Federal Reserve Bank of St. Louis, On the Economy blog. Open source (full text read)
- Passage
We found an average time savings of 5.4% of work hours in the November 2024 survey.
- Where
- Section 'Time Savings and Productivity Gains'; the same section gives 1.4 per cent of all workers' hours, and 1.1 per cent as the model-implied increase in aggregate productivity
- Limits
- Self-reported. The 1.1 per cent often quoted with it is modelled productivity, not hours. A later post revised the series.
- Checked
- C36Supported
Employment of 22 to 25 year olds in AI-exposed US occupations is 19 per cent below where it would be had it kept pace with less-exposed peers, with no comparable gap for experienced workers.
- Source
- Brynjolfsson, E., Chandar, B., & Chen, R. (2025, revised August 2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab working paper. Open source (full text read)
- Passage
now stands 19% below where it would be had it kept pace with that of their less-exposed peers; experienced workers show no comparable gap.
- Where
- Abstract, August 2026 revision; the sentence opens with the age group
- Limits
- The authors call these descriptive indicators, not causal estimates, and report some trends that predate generative AI. The payroll data come from a firm that sells payroll and HR software.
- Checked
- C37Supported
Among Danish chatbot users reporting time savings, fewer than 10 per cent said they took more breaks or leisure, and about 30 per cent spent more time on the same tasks.
- Source
- Humlum, A., & Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Previously circulated as "Large Language Models, Small Labor Market Effects". Open source (full text read)
- Passage
fewer than 10% report taking additional breaks or leisure, and about 30% spend more time on the same tasks they initially saved time on.
- Where
- Section 2.3, footnote 9; Table E.3 gives other tasks .847, same tasks .294, breaks .072 and leisure .058, and for HR professionals .899, .219, .094 and .068
- Limits
- The survey question asks what people expect to do if a chatbot saves time, so these are expectations, not observed time. More than one answer was allowed.
- Checked
- C38Supported
In the same study, 35 per cent of the new AI-related work people described was quality review or ethics and compliance.
- Source
- Humlum, A., & Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Previously circulated as "Large Language Models, Small Labor Market Effects". Open source (full text read)
- Passage
Notably, 35% of new AI-related work is dedicated to AI Quality Review and AI Ethics & Compliance
- Where
- Section 2.3, printed page 15
- Limits
- A classification of tasks people described, not a share of working time.
- Checked
Back to text↩ 1
- C39Supported
The same study reports that AI chatbots generated new workloads for 12 per cent of users, a figure for workplaces without an employer initiative.
- Source
- Humlum A, Vestergaard E (2025, revised 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Open source (full text read)
- Passage
AI chatbots have generated new workloads for 12% of users
- Where
- Section 2.3; the figure is for workplaces without an employer initiative
- Limits
- Self-reported.
- Checked
Back to text↩ 1
- C40Supported
In the Department for Work and Pensions trial, 89 per cent of responding users said they reallocated time saved to other work tasks.
- Source
- Department for Work and Pensions (2026, 29 January). An Evaluation of DWP's Microsoft 365 Copilot Trial. GOV.UK. Open source (full text read)
- Passage
The majority of users (89%) reported reallocating the time saved to completing other work tasks.
- Where
- Chapter 3, use of time saved; select all that apply
- Limits
- Self-reported. 48 per cent of licence holders responded and users were self-selected.
- Checked
- C41Supported
Researchers who spent eight months observing one US technology firm report that AI tools consistently intensified work.
- Source
- Ranganathan, A., & Ye, X. M. (2026, 9 February). AI Doesn't Reduce Work, It Intensifies It. Harvard Business Review. Open source (full text read)
- Passage
In our in-progress research, we discovered that AI tools didn’t reduce work, they consistently intensified it.
- Where
- Opening section
- Limits
- One firm of about 200 staff, no comparison group, no hours measured. A practitioner magazine article ahead of peer review.
- Checked
Back to text↩ 1
- C42Supported
Clinicians annotating notes drafted by a language model from mock consultations found hallucinations in 1.47 per cent of 12,999 sentences, 44 per cent of them major.
- Source
- Asgari, E., Montaña-Brown, N., Dubois, M., Khalil, S., Balloch, J., Au Yeung, J., & Pimenta, D. (2025). A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine, 8, 274. Open source (full text read)
- Passage
Of 12,999 sentences in 450 clinical notes, 191 sentences had hallucinations (1.47%), of which 84 sentences (44%) were major (could impact patient diagnosis and management if left uncorrected).
- Where
- Results, hallucinations
- Limits
- All authors were employees of a UK clinical scribe company. Mock consultations, not live clinics; one model.
- Checked
- C43Supported
Specialist AI legal research tools tested in 2024 hallucinated between 17 and 33 per cent of the time.
- Source
- Magesh, V., Surani, F., et al. (2025). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Journal of Empirical Legal Studies. (arXiv 2405.20362 version of 3 June 2024 opened) Open source (full text read)
- Passage
each hallucinate between 17% and 33% of the time.
- Where
- Abstract; the sentence opens by naming the tools and their publishers
- Limits
- Preprint version opened. Benchmark queries, not live use. Independent of the tools' makers.
- Checked
- C44Supported
In the Home Office pilot, specialists reviewed every AI summary before use and removed 9 per cent as inaccurate or missing information.
- Source
- Home Office (2025, 29 April). Evaluation of AI trials in the asylum decision making process. Research note, GOV.UK. Open source (full text read)
- Passage
Technical specialists reviewed all summaries for accuracy prior to use in the pilot. A small proportion of summaries produced (9%) were deemed to be inaccurate or had missing information and were therefore removed from the pilot
- Where
- Section 5.1
- Limits
- Because faulty summaries were removed first, the pilot did not measure what happens when staff act on them. Decision quality on a dip-sample was similar in the two groups.
- Checked
- C45Supported
In the Department for Business and Trade's diary, 22 per cent of respondents said they had identified hallucinations.
- Source
- Department for Business and Trade (2025, August). The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade. Open source (full text read)
- Passage
22% of respondents did identify hallucinations
- Where
- Findings
- Limits
- Self-reported. How often hallucinations occurred, and how long they took to fix, was not measured.
- Checked
Back to text↩ 1
- C46Supported
In the same diary, scheduling and generating images took longer with the tool, with mean time savings of minus 0.6 and minus 0.5 hours per task.
- Source
- Department for Business and Trade (2025, August). The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade. Open source (full text read)
- Passage
“scheduling” resulted in mean time savings of -0.6 hours per task, and “generating images” resulted in mean time savings of -0.5 hours per task, meaning they both took longer
- Where
- Findings, Table 5; these are after tasks done only because of the tool, and unused outputs, were counted as negative time
- Limits
- Self-reported diary times.
- Checked
- C47Supported
In a stepped-wedge randomised trial with 66 practitioners, an ambient AI scribe reduced work exhaustion and interpersonal disengagement by 0.44 points on a five-point scale.
- Source
- Afshar M, Baumann MR, Resnik F, Hintzke J, Sullivan AG, Wills G, et al. (2025). A Pragmatic Randomized Controlled Trial of Ambient Artificial Intelligence to Improve Health Practitioner Well-Being. NEJM AI. Open source (full text read)
- Passage
Ambient AI use had a significant reduction in work exhaustion/interpersonal disengagement (−0.44 points; 95% confidence interval [CI], −0.62 to −0.25; P<0.001)
- Where
- Abstract, results; 24 weeks
- Limits
- Open-label, with self-reported outcomes. The co-primary outcome, professional fulfilment, was not significant at the trial's threshold. Licences were bought from the tool's maker, which gave technical support and is thanked in connection with review of the manuscript.
- Checked
- C48Supported
Practitioners had to be willing to adopt the tool to take part.
- Source
- Afshar M, Baumann MR, Resnik F, Hintzke J, Sullivan AG, Wills G, et al. (2025). A Pragmatic Randomized Controlled Trial of Ambient Artificial Intelligence to Improve Health Practitioner Well-Being. NEJM AI. Open source (full text read)
- Passage
Additional eligibility required a willingness to adopt the ambient AI tool.
- Where
- Methods, recruitment and eligibility
- Limits
- Volunteers, so the result may not hold for staff who would not choose the tool.
- Checked
Back to text↩ 1
- C49Supported
In a randomised trial of 238 physicians, the odds ratios for single-item burnout with two scribes against usual care were 0.78 (0.40 to 1.54) and 0.67 (0.34 to 1.34).
- Source
- Lukac PJ, Turner W, Vangala S, Chin AT, Khalili J, Shih YT, Sarkisian C, Cheng EM, Mafi JN (2025). Ambient AI Scribes in Clinical Practice: A Randomized Trial. NEJM AI 2(12). Open source (full text read)
- Passage
Single-item burnout 0.78 0.40 to 1.54 0.67 0.34 to 1.34
- Where
- Table 2 (odds ratio and 95 per cent interval for each scribe against control)
- Limits
- A secondary outcome, not hypothesis-tested. About two months of follow-up. Work exhaustion changed by minus 0.32 (minus 0.55 to minus 0.08) and minus 0.23 (minus 0.46 to 0.01).
- Checked
- C50Supported
A 2026 meta-analysis rated certainty moderate for less cognitive workload with ambient AI scribes and low for less burnout.
- Source
- Gong EJ, Bang CS, Lee JJ (2026). Cognitive Workload and Mental Burden in Health Care Professionals Interacting With AI: Systematic Review and Meta-Analysis. Journal of Medical Internet Research 2026;28:e93618. Open source (full text read)
- Passage
GRADE certainty was moderate for cognitive workload reduction with ambient AI, low for burnout reduction with ambient AI
- Where
- Abstract; 21 studies, 2,885 health care professionals
- Limits
- Prediction intervals crossed the null. The work exhaustion pool includes a study of AI-drafted replies to patient messages, and the burnout pool is three uncontrolled before-and-after studies.
- Checked
- C51Supported
In an uncontrolled study, intention to leave was 28.3 per cent before and 26.6 per cent 84 days after clinicians began using an ambient scribe, P = .61.
- Source
- You JG, Dbouk RH, Landman A, Ting DY, Dutta S, Wang JC, et al. (2025). Ambient Documentation Technology in Clinician Experience of Documentation Burden and Burnout. JAMA Network Open 8(8):e2528056. Open source (full text read)
- Passage
Intention to leave 54 (28.3) NA 51 (26.6) .61
- Where
- Table 3 (number and per cent at baseline, then at 84 days, then P), one health system, 192 clinicians
- Limits
- No control group and a low response rate. The only study found that measured intention to leave.
- Checked
Back to text↩ 1
- C52Supported
In a cross-sectional survey of radiologists, those who reported regular AI use had higher odds of burnout, odds ratio 1.20 (1.10 to 1.30).
- Source
- Liu H, Ding N, Li X, Chen Y, Sun H, Huang Y, et al. (2024). Artificial Intelligence and Radiologist Burnout. JAMA Network Open 7(11):e2448714. Open source (full text read)
- Passage
In the primary analysis with IPW, AI use was significantly associated with higher odds of burnout (OR, 1.20; 95% CI, 1.10-1.30).
- Where
- Results, primary weighted analysis
- Limits
- Cross-sectional. Diagnostic AI, which adds a task rather than removing admin.
- Checked
Back to text↩ 1
- C53Supported
In the customer-support study, less skilled and less experienced agents gained about 30 per cent in issues resolved per hour.
- Source
- Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889-942 (passage quoted from arXiv:2304.11771v2, section 4.1; published QJE text blocked, HTTP 403). Open source (full text read)
- Passage
less-skilled and less-experienced workers improve significantly across all productivity measures we consider, including approximate increase of 30% in the number of issues they are able to resolve per hour.
- Where
- Section 4.1, arXiv version 2
- Limits
- One firm, live chat. Read in the arXiv version matching the published abstract; the sentence was not checked against the journal text.
- Checked
- C54Supported
In Denmark, women were 16 percentage points less likely than men in the same occupation to have used an AI chatbot for work.
- Source
- Humlum, A., & Vestergaard, E. (2025). The unequal adoption of ChatGPT exacerbates existing inequalities among workers. Proceedings of the National Academy of Sciences, 122(1), e2414972121. Open source (full text read)
- Passage
women are 16 percentage points less likely to have used ChatGPT for work than men in the same occupation.
- Where
- Introduction
- Limits
- Adoption in late 2023, not outcomes. Cross-sectional.
- Checked
Back to text↩ 1
- C55Supported
In the Department for Business and Trade's evaluation, respondents identifying as neurodiverse were more likely to be satisfied, at the 90 per cent confidence level only.
- Source
- Department for Business and Trade (2025, August). The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade. GOV.UK. Open source (full text read)
- Passage
Mann-Whitney U tests identified that respondents identifying as neurodiverse were statistically more likely to be satisfied (to a 90% confidence level)
- Where
- Findings, health conditions and disability; the test table gives p = 0.09978 for satisfaction and 0.03484 for recommending the tool
- Limits
- Satisfaction only, no comparison group, no adjustment for multiple tests. Tests for hearing, vision, learning difficulty and mental health conditions were not significant.
- Checked
Back to text↩ 1
- C56Supported
In a US unemployment insurance adjudication project, examiners rated an AI fact-finding system highly, but it did not improve quality or efficiency against the sandbox control group.
- Source
- Martin, O., Magesh, V., Surani, F., Rodolfa, K. T., Ho, D. E., & Perez, A. (2026). Evaluating Generative AI in Benefits Administration: A Demonstration Project. Proceedings of the Symposium on Computer Science and Law (CSLAW). Open source (abstract only)
- Passage
AI fact-finding was a substantial improvement to historical (observational) baselines and examiners subjectively rated the system highly; but the system did not improve quality or efficiency in the sandbox control group, though it may reduce inter-adjudicator variance.
- Where
- Abstract
- Limits
- Abstract only, so sample size, effect sizes and precision were not seen. Public welfare benefits, not employee benefits, tested on historical cases.
- Checked
- C57Supported
In a case study of one multinational, an interviewee said its HR bot had dealt with 11,000 questions with 81 per cent accuracy.
- Source
- Malik, A., Budhwar, P., Patel, C., & Srikanth, N. R. (2022; online 2020). 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 to 1178. Open source (full text read)
- Passage
has dealt with 11,000 questions with 81 per cent accuracy
- Where
- Findings, interviewee quotation; the sentence opens with the bot's name
- Limits
- An interviewee's statement from 2019, not an audited measure. Accuracy is not defined and there is no comparison. The case firm builds AI applications.
- Checked
- C58Supported
After an NHS payroll provider launched a payslip app with a chatbot, payroll service desk cases fell to an average of 9,474 a month.
- Source
- 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. Open source (full text read)
- Passage
Between November 2018 and January 2019, however, the total number of cases was down significantly to an average of 9,474 per month
- Where
- Paragraph 4; the sentence continues 'a 26 per cent decrease', from 12,808 a month
- Limits
- Press release from the provider, reproduced by a news service. Two adjacent three-month windows, no control. The chatbot is one feature of the app, and it predates generative AI.
- Checked
- C59Supported
In the Danish registers, no occupation, HR professionals included, showed a statistically significant effect on earnings.
- Source
- Humlum, A., & Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Open source (full text read)
- Passage
Nonetheless, we find no statistically significant effects in any occupation, and point estimates are generally near zero.
- Where
- Section 3, occupational heterogeneity, Figure 7; the preceding sentence says bands are wider for single occupations
- Limits
- Imprecise for HR alone. Earnings, not admin time or accuracy.
- Checked
- C60Supported
The Department for Business and Trade's interviews identified benefits for colleagues in HR, among other roles.
- Source
- Department for Business and Trade (2025). The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade. August 2025. GOV.UK. Open source (full text read)
- Passage
Benefits were identified for colleagues in Human Resources, Communications, and Commercial roles.
- Where
- Findings, satisfaction and impact to specific roles: interviews, page 27
- Limits
- Interviews, not measurement. The report did not run statistical tests by role because some samples were small. No HR time figure.
- Checked
Back to text↩ 1
- C61Supported
UK firms in the Bank of England's Decision Maker Panel survey estimate that AI has raised their productivity by around 0.5 per cent over the past three years.
- Source
- Bank of England (2026). Monetary Policy Report, November 2026. Box D: Developments in UK productivity growth. Open source (full text read)
- Passage
Firms responding to the DMP survey estimate that AI has increased productivity by around 0.5% over the past three years, in line with similar impacts reported in other countries.
- Where
- Box D. The same box says there is so far only evidence of a modest AI-driven rise in UK productivity growth, and that firms report close to zero impact of AI on employment
- Limits
- Firms' own estimates in a survey, not measured output. The whole economy, not HR. Read in a browser after the research run.
- Checked
Back to text↩ 1
- C62Supported
In the CIPD's autumn 2025 Labour Market Outlook, employees in 76 per cent of UK organisations were using AI tools at work.
- Source
- CIPD (2025). Labour Market Outlook, autumn 2025. Fieldwork by YouGov, 19 September to 14 October 2025. Open source (full text read)
- Passage
Employees in three-quarters of UK organisations are using AI tools at work (76%).
- Where
- Section on employees' use of AI; a survey of 2,019 senior HR professionals and decision-makers in the UK, weighted
- Limits
- Use of any AI tool by any employee, as reported by HR respondents. It does not measure HR teams' own use, or buying. Added after the research run.
- Checked
Back to text↩ 1
- C63Supported
In a survey of European HR professionals published on the CIPD's website, 19 per cent described AI as integrated into HR workflows and 66 per cent said their organisation had no clear guidelines for AI use in HR.
- Source
- European Association for People Management AI Working Group (2026). HR and AI: what Europe is actually doing (and what it isn't). CIPD blog. Open source (full text read)
- Passage
Only 19% describe AI as genuinely integrated into workflows. And 66% say their organisation has no clear guidelines for AI use in HR.
- Where
- Body of the blog; the survey covered 327 European HR professionals in 30 countries, most of them heads of HR
- Limits
- A blog reporting an association's own survey. Recruitment and response rate are not stated. Read in a browser after the research run.
- Checked
Back to text↩ 1
Back to the noteDoes AI give HR its time back? Nobody has measured it.