Zak FentonWorking notes

· Note

Nobody to hand it to

What changes when an agent keeps producing work for someone who has finished for the day?

Evidence reviewDoomworking: what the evidence supports

An agent can keep producing output after the person responsible for checking it has finished for the day. That raises two practical questions: when does the person need to look, and who is responsible while they are away?

Research on on-call work, recovery and automation can help frame those questions. Applying it to agent work still requires care about what each study measured. Care is not the natural state of anyone writing about AI right now, me included.

Where the pattern sits

Blayney describes the appointment of the Health of Munition Workers Committee in September 1915 and the development of fatigue research around productive capacity.1 That history does not, on its own, explain why earlier working-time protections were introduced.

Continuous processes are something the law already describes. The Working Time Directive defines shift work as workers succeeding each other at the same work stations, and says it may be continuous or discontinuous.2 Industry has long run around the clock by rotating people through it, and that has a measured association with harm. A 2012 review included 34 studies overall. Its myocardial-infarction estimate came from 10 studies: a risk ratio of 1.23, with a 95 per cent confidence interval from 1.15 to 1.31.3 That association does not establish the health risk of reviewing agent output.

On-call work has its own literature. A 2004 review said on-call work had received significantly less research attention than shift work or overtime.4 A 2017 systematic review of on-call work and sleep found eight studies, and noted that they did not distinguish nights on call from home with calls from nights without.5 A 2025 review of fatigue management for on-call workers covered 17 studies.6 These reviews concern on-call work and fatigue management. They do not establish the effects on recovery of reviewing agent output.

What the AI-at-work data can and cannot say

In Danish survey data linked to administrative registers, the authors classified the new AI-related tasks workers described, and 35 per cent of those descriptions fell into AI quality review, ethics and compliance.7 That is a classification of what people said their new tasks were, not a share of working time.

Among chatbot users who reported saving time, 7.2 per cent selected more breaks and 5.8 per cent selected more leisure as expected uses of the savings.7 These are responses to a survey question, not percentages of saved minutes.

The same study estimates precise null effects on earnings and recorded hours in its Danish setting, ruling out effects larger than 2 per cent two years after ChatGPT launched.7 Recorded hours are what the registers hold. They do not observe unpaid evening work or recovery.

In a survey of 319 knowledge workers, Lee and colleagues describe a reported shift from gathering information to checking AI outputs.8 This is reported experience, not a timed measure of verification work.

Self-report has a known weakness here. In METR’s early-2025 randomised trial, 16 experienced developers on 246 tasks took 19 per cent longer with AI tools while estimating they had been faster.9 I’ve written about that study and its 2026 follow-up separately. A review of the technostress literature also cautions that its studies may overlap exposure and outcome measures.10

What the automation literature says

Bainbridge discussed the difficulty of monitoring automation that has operated acceptably for a long time.11 Parasuraman and Manzey’s later review describes complacency when manual tasks compete with an automated task for attention.12 Neither finding establishes the effect of reviewing an agent’s overnight output.

Two of Bainbridge’s other observations, made about real-time process control, still frame the question. Automating the easy parts of a task can make the remaining difficult parts harder.11 And an operator can only monitor an automated decision at a meta-level, judging whether it is acceptable.11 She also thought that humans working without time pressure could be impressive problem solvers, and that time pressure was the difficulty.11 Whether morning delivery leaves enough time for review depends on the workload and the deadline. Delivery time alone does not answer that.

What the recovery literature says

In a cross-sectional survey of 157 on-call workers, perceived stress from unpredictability was associated with fatigue and work-home interference, more than the amount of on-call exposure was.13 That is an association in one sample. It does not establish that unpredictability alone causes the harm, or that agent output is predictable enough to escape it.

Across 23 samples and 7,007 workers, psychological detachment from work correlated minus 0.36 with burnout exhaustion.14 The same meta-analysis found no significant average association between detachment and physiological stress indicators, on three samples and 219 people, and negative associations with contextual performance and creativity.14 These are associations. They do not tell us what would happen to performance if an employer changed the conditions for detaching from work.

The law

For Great Britain, the starting point is the Working Time Regulations 1998. Their definition includes time spent working, at the employer’s disposal, carrying out the worker’s duties. It also includes relevant training and periods covered by a relevant agreement.15

The 2021 standby judgment distinguishes the whole period of availability from time actually spent working. Where the whole period does not qualify as working time, work actually done during it still counts.16 It does not justify calling someone who is checking output “probably resting”.

That judgment interprets EU law. UK courts may consider it where relevant but are not bound by it.17 The classification of a particular check still depends on the facts and the applicable rules. I would look at the work done, the response expected and whether the organisation made rest possible.

In the earlier case against Britain, the Court rejected guidance that could leave workers unable to exercise their rest rights. It also said employers are not generally required to force workers to claim their rest.18 That is a judgment about effective rest rights and the limits of the employer’s obligation. It did not decide an AI-notification dispute.

Ontario’s written-policy provisions define disconnecting to include freedom from reviewing work-related messages. The provisions do not themselves create a right to disconnect.19

The EU’s platform work directive prohibits digital labour platforms from using automated monitoring or decision systems in ways that put undue pressure on platform workers or endanger their safety or physical or mental health.20 It applies to digital labour platforms, with a transposition deadline of 2 December 2026. It is not a rule for every employer using AI, and it does not apply in the UK.

HSE’s Management Standards define Demands as including workload, work patterns and the work environment.21 The step-by-step workbook was published in March 2019. I would use it alongside a specific assessment of what an agent changes: the pace of work, when it arrives and who is expected to check it.

What I would do

I would measure time spent checking output separately from recovery. Established recovery questionnaires may help with the second question, provided the version, population, recall period, scoring and permissions fit the use. A new item about verification needs its own evaluation.

For work that can wait, I would test holding notifications until the next working period. The studies discussed here do not establish what that change would do to recovery from agent work.

One practitioner account is worth reading with its interest declared. SaaStr, a media and events business that runs more than twenty agents with three people and is a Replit customer featured in Replit’s own case study, says its three people work harder than its twenty-person team did.22 That is one organisation describing itself, not a measured effect.

I would start by making the expectation explicit: who checks the output, when they check it, and what can wait. Then I would test whether that arrangement leaves people able to finish for the day. Finishing for the day used to be a thing. I’d like to know whether we still have it before we find out the hard way.

This piece is a reading of primary sources and published research. It is not legal advice. The claims are itemised with sources, passages and limits in the accompanying evidence review.

Footnotes

  1. Blayney, S. (2017). Industrial Fatigue and the Productive Body: the Science of Work in Britain, c. 1900-1918. Social History of Medicine, 32(2), 310-328. doi:10.1093/shm/hkx077. ↩

  2. Directive 2003/88/EC concerning certain aspects of the organisation of working time, Article 2(5). EUR-Lex. ↩

  3. Vyas, M. V. et al. (2012). Shift work and vascular events: systematic review and meta-analysis. BMJ, 345, e4800. doi:10.1136/bmj.e4800. Results: ten studies recorded myocardial infarction. ↩

  4. Nicol, A.-M. and Botterill, J. S. (2004). On-call work and health: a review. Environmental Health, 3, 15. doi:10.1186/1476-069X-3-15. ↩

  5. Hall, S. J. et al. (2017). The effect of working on-call on stress physiology and sleep: a systematic review. Sleep Medicine Reviews, 33, 79-87. doi:10.1016/j.smrv.2016.06.001. ↩

  6. Bumpstead, H. et al. (2025). How should we manage fatigue in on-call workers? A review of guidance materials and a systematic review of the evidence-base. Sleep Medicine Reviews, 79, 102012. doi:10.1016/j.smrv.2024.102012; Massey University repository. Seventeen original studies on fatigue management strategies; it does not establish the effects on recovery of reviewing agent output. ↩

  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. PDF. Task classification at section 2.3, Figure 4 and Appendix B.3; time savings at Table E.3, columns 9 and 10, restricted to users reporting savings. Working paper, not peer reviewed. ↩ ↩2 ↩3

  8. Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI 2025. doi:10.1145/3706598.3713778; author-hosted PDF, sections 5.2 and 5.2.1. ↩

  9. 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. Not peer reviewed. ↩

  10. Borle, P., Reichel, K., Niebuhr, F. and Voelter-Mahlknecht, S. (2021). How Are Techno-Stressors Associated with Mental Health and Work Outcomes? International Journal of Environmental Research and Public Health, 18(16), 8673. doi:10.3390/ijerph18168673. Abstract consulted. ↩

  11. Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779. doi:10.1016/0005-1098(83)90046-8; IFAC-hosted PDF, sections 1.1.3, 3 and 4. ↩ ↩2 ↩3 ↩4

  12. Parasuraman, R. and Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation: An Attentional Integration. Human Factors, 52(3), 381-410. doi:10.1177/0018720810376055. Abstract consulted. ↩

  13. Ziebertz, C. M. et al. (2015). The Relationship of On-Call Work with Fatigue, Work-Home Interference, and Perceived Performance Difficulties. BioMed Research International, 2015, 643413. doi:10.1155/2015/643413. ↩

  14. Wendsche, J. and Lohmann-Haislah, A. (2017). A Meta-Analysis on Antecedents and Outcomes of Detachment from Work. Frontiers in Psychology, 7, 2072. doi:10.3389/fpsyg.2016.02072, Table 1. ↩ ↩2

  15. Working Time Regulations 1998, regulation 2(1), definition of working time, paragraphs (a) to (c). ↩

  16. C-344/19, D.J. v Radiotelevizija Slovenija, 9 March 2021, paragraphs 37 to 40. ↩

  17. European Union (Withdrawal) Act 2018, section 6(1) and (2), treatment of later EU decisions. ↩

  18. C-484/04, Commission v United Kingdom, 7 September 2006, paragraphs 42 to 44. ↩

  19. Ontario Ministry of Labour, Immigration, Training and Skills Development. Employment Standards Act Policy and Interpretation Manual, Part VII.0.1, sections 21.1.1 and 21.1.2. ontario.ca. ↩

  20. Directive (EU) 2024/2831 on improving working conditions in platform work, Articles 1(3), 2, 12(3) and 29(1). ↩

  21. Health and Safety Executive (2019). Tackling work-related stress using the Management Standards approach: a step-by-step workbook, WBK01, March 2019. PDF. ↩

  22. Lemkin, J. (2026). Dear SaaStr: Should I Run 20+ AI Agents the Way SaaStr Does? SaaStr. saastr.com, section “The Part Nobody Tells You”. Grey literature with a declared commercial interest. ↩

The evidence behind this noteDoomworking: what the evidence supports