A machine that doesn't get sick

This morning I wrote about the models giving us more work than we can check, and that post was about the jobs that stay. This post is the other half of the question, the jobs that go, and I want to go at it through two things everyone who has run an office knows and nobody puts in the plan. People are ill some of the time, and two people given the same case don't give the same answer. The scenario I have in mind is admin and office work running on something almost as good as average human judgement, maybe a bit better, that works every hour of the year, and I don't think we're far from it.

The sick day

The Office for National Statistics counted 148.8 million working days lost to sickness or injury in the UK in 2025, 4.4 days a worker, 2.0% of all working hours, and that's a rate that hasn't moved since 2024 and is still above 2019 ^[1]. The American numbers say the same thing by occupation. The Bureau of Labor Statistics puts the illness absence rate for office and administrative support workers at 2.8% in 2025, with 1.4% of work time lost, against 1.9% and 0.9% for managers and professionals ^[4]. So the people doing the admin are off ill more often than the people deciding things, which is the opposite of what you'd want if you were designing it.

The days off are the smaller cost. Deloitte's 2024 estimate of what poor mental health costs UK employers is £51 billion a year, and the largest part of it, about £24 billion, is presenteeism, people at their desks and not working well ^[5]. You can't see presenteeism on a rota, and it's the thing that turns a 2% absence rate into something much larger, because the person who came in with a cold is answering the emails a bit worse than they did last week, and nobody measures that.

The noise

The second cost is that human judgement doesn't repeat. Kahneman, Sibony and Sunstein built a book on it in 2021, and the example I keep coming back to is an insurance company that asked its underwriters to price the same risks. The executives expected two underwriters to differ by about 10% and thought that was fine, and the median difference was 55% ^[2]. The same book found a mean gap of three and a half years between the sentences different judges gave for the same simplified case ^[2]. Nobody in either case was being careless.

It gets worse with the hour. Landrigan and colleagues put interns on the traditional hospital schedule, with shifts over 24 hours, against a schedule with shorter shifts, and the long-shift interns made nearly 36% more serious medical errors and 5.6 times as many serious diagnostic errors ^[6]. Torres and Williams looked at 284 bail hearings across two judges in New Jersey and found that how far into the session a case came affected the bail set by both judges ^[7]. The one I didn't expect is in a benchmark. When OpenAI built GDPval, its benchmark of real work tasks graded by people in those jobs, it had to measure how often two expert graders agreed with each other about which of two pieces of work was better, and the answer was 70.8% of the time ^[8]. The people hired to judge the models couldn't agree with each other three times in ten.

We don't count this as a cost because there's no line for it. There's no "variance" column in a payroll. But a 55% spread in how a risk gets priced is money, and a 30% disagreement between experts is a lot of work being done twice, or wrongly once.

The machine seat

Now put a machine in the seat. I'm not going to argue it's better than the best person, only that in a lot of this work it's already at or past the average one, which is a different and lower bar. Agarwal, Moehring, Rajpurkar and Salz ran 227 radiologists through 324 chest X-rays and found a model that was more accurate than 78% of them, and the odd part is that giving the radiologists the model's answer didn't improve them on average, because it helped when the model was sure and they weren't, and hurt when the model was unsure and they were sure, and over the whole set the two cancelled out ^[9]. On GDPval, the best model in September 2025 won or tied against the industry expert in 47.6% of tasks, in work like legal memos, spreadsheets and slide decks ^[10]. That number is a year old.

The hours are the easy part of the arithmetic. A UK worker works 1,533 hours a year on average and an American one 1,800 ^[3], and a year has 8,760. Take the 2% off for sickness and a human seat gives you about 1,500 hours of work, and a machine seat gives you all 8,760 of them, or nearly six seats' worth before anyone has said a word about quality. Then there's the variance. A model asked the same question twice doesn't give the same answer either, I go into that below, but you can ask it three times and take the majority, which is what Haldar and Hockenmaier found improved accuracy ^[11], and a second opinion from a machine costs seconds. A second opinion from an underwriter is a second underwriter.

So on the office and admin work the Bureau counts, the 17.8 million jobs I used this morning ^[12], the scenario is a seat that never takes the 4.4 days, never comes in with a cold, never prices the risk at 55% off its colleague, and does it at four in the morning. I don't think any of that is ten years away, and most of what's left of it is a buying decision.

The machine's sick day

What I don't want to do is pretend the machine doesn't have bad days, because it has two kinds and they're both worse than ours in one way.

The first is that it isn't deterministic either. Thinking Machines showed in September 2025 that the usual explanation for a model giving different answers to the same prompt, floating point arithmetic on a busy GPU, isn't the main cause. The main cause is that the batch size changes with how busy the server is, and the same request in a different batch gives a different answer, and they had to write batch-invariant kernels to make a model answer the same way twice ^[13]. Haldar and Hockenmaier measured it on models used as judges, and one of the better ones gave the same score on all three runs only 61.3% of the time ^[11]. So the noise is there. The difference is that it's cheap to sample and vote it away, and that the noise in a human can't be sampled, because the human has gone home.

The second is the outage, and this is the one that matters. IncidentHub counted 30,246 outages across 1,082 providers in the first half of 2026, with Azure's East US control plane down for over twelve hours in April and an AWS region in the Middle East out for months after physical damage in March ^[14]. And the one that makes the point best was Anthropic suspending access to Claude Fable 5 and Mythos 5 from 13 June to 1 July, eighteen days and nineteen hours, to comply with export controls, which IncidentHub keeps out of its count, as a suspension rather than a failure ^[14]. A policy rather than a bug, and the seat was empty either way.

The comparison turns on one word for me. Human sick days are uncorrelated. When one of your hundred clerks is ill, ninety-nine are at work, and the 2% is a 2% you can plan around. Machine sick days are correlated. When the model is down, all hundred seats are down at once, and when the model is updated, all hundred seats change their minds at once, in the same direction, and nobody tells you. We track the second kind on the Sunsets page because it's the one killing agent workflows, a model retired or changed under something that depended on it. Eighteen days is 5% of a year, more than the 2% a person loses, and it lands on every seat on the same morning. A 2% loss spread across a hundred people and a 5% loss taken all at once are not the same risk, and most of the people buying the machine seat are pricing them as if they were.

What I think happens

So the office I'm describing has fewer human seats and the ones left are doing what I wrote about this morning, checking, and the checking is the thing that doesn't scale to 8,760 hours, because the people doing it still sleep. The machine's variance gets handled by sampling, which is cheap. The machine's outages get handled the way we handle any correlated risk, with a second provider, which most companies won't buy until the first eighteen-day gap, and then all of them will. And the human variance, the 55%, goes away in the parts of the work that get automated and shows up again, concentrated, in the parts that don't, which are the judgements nobody trusts a model with yet.

I'm not sure about the timing. The radiology result says the model is already past most of the radiologists and the radiologists still have jobs, because nobody has yet decided to let it decide. The GDPval number says the model wins or ties in about half the office tasks, which means it loses the other half, and I don't know how fast that moves. And I'm not sure the presenteeism cost comes back as savings, because a model with a bad batch size is a kind of presenteeism too, it's just one you can measure.

If you run an office and you're reading this, the number I'd want from you is the one I couldn't find, which is how far apart two of your people are when they handle the same case. I'd bet it's closer to 55% than to 10%, and that you haven't measured it either.

References

Read on 3 November 2026. Quotes are as published.

1. [1]Office for National Statistics, 1 May 2026, "Sickness absence in the UK labour market, 2025". Source for 148.8 million working days lost to sickness or injury in 2025, 4.4 days a worker, a sickness absence rate of 2.0% of working hours unchanged from 2024 and 0.2 days above 2019, and the public sector rate of 2.9% against 1.7% in the private sector. These are official statistics in development. ons.gov.uk
2. [2]HEC Paris, news on Olivier Sibony's work with Daniel Kahneman and Cass Sunstein on their 2021 book "Noise, A Flaw in Human Judgment". Source for the insurance noise audit, in which executives expected two underwriters pricing the same risk to differ by a median of about 10% and the median difference was about 55%, and for the mean difference of 3.5 years between sentences handed down by different judges for the same simplified cases. hec.edu
3. [3]OECD, the annual hours worked per worker dataset (ANHRS), as tabulated on Wikipedia's list of countries by average annual labour hours, read on 3 November 2026. Source for 1,533 hours a year in the United Kingdom and 1,800 in the United States in 2025. The 8,760 hours in a year and the seats arithmetic are mine. en.wikipedia.org
4. [4]US Bureau of Labor Statistics, Current Population Survey, Table 47, "Absences from work of employed full-time wage and salary workers by occupation and industry", 2025 annual averages. Source for an illness or injury absence rate of 2.8% and a lost worktime rate of 1.4% for office and administrative support occupations, against 1.9% and 0.9% for management, professional and related occupations, and 2.2% and 1.1% for all full-time workers. bls.gov
5. [5]Deloitte, 17 May 2024, "Poor mental health costs UK employers £51 billion a year for employees". Source for the £51 billion a year estimate, down from £55 billion in 2021 and up from £45 billion in 2019, and for presenteeism being the largest part at about £24 billion a year. deloitte.com
6. [6]Christopher P. Landrigan and colleagues, 2004, "Effect of reducing interns' work hours on serious medical errors in intensive care units", New England Journal of Medicine 351(18), as summarised by AHRQ PSNet. Source for interns on the traditional schedule, with shifts of over 24 hours, making nearly 36% more serious medical errors and 5.6 times as many serious diagnostic errors than interns on a schedule with shorter shifts. psnet.ahrq.gov
7. [7]Luis C. Torres and James H. Williams, 2022, "Tired Judges? An Examination of the Effect of Decision Fatigue in Bail Proceedings", Criminal Justice and Behavior 49(8). Source for the study of 284 bail hearings in two New Jersey jurisdictions, in which case order and session length limited one judge's engagement and fatigue affected the bail amounts set by both judges. The abstract gives no effect sizes. experts.umn.edu
8. [8]TestMu AI, "OpenAI GDPval Benchmark, Tasks, Grading, Scores and Limits", a write-up of the GDPval paper (Tejal Patwardhan and colleagues, OpenAI, 5 October 2025), read on 3 November 2026. Source for the benchmark's 1,320 tasks across 44 occupations, grading by occupation experts blind to the source, and two experts agreeing with each other 70.8% of the time on the paper's measure, with OpenAI's automated grader agreeing with experts 65.7% of the time. testmuai.com
9. [9]Nikhil Agarwal, Alex Moehring, Pranav Rajpurkar and Tobias Salz, 17 November 2025, "Combining Human Expertise with Artificial Intelligence, Experimental Evidence from Radiology". Source for 227 radiologists, 324 chest X-ray cases and 41,920 diagnostic decisions, a model (CheXpert, AUROC 0.85) more accurate than 78% of the radiologists, and AI assistance not significantly improving average performance, with radiologists moving their probabilities 18.3% of the way towards the model's. economics.mit.edu
10. [10]TechRadar, 29 September 2025, "Claude just beat GPT-5, Gemini, and Grok in real-world job tasks, according to OpenAI's own study". Source for Claude Opus 4.1 winning or tying against industry experts in 47.6% of GDPval tasks and GPT-5 at 38.8%. techradar.com
11. [11]Rajarshi Haldar and Julia Hockenmaier, November 2025, "Rating Roulette, Self-Inconsistency in LLM-As-A-Judge Frameworks", Findings of EMNLP 2025. Source for models used as judges giving different scores to the same input across three runs with settings unchanged, Qwen 3 giving the same score on all three runs in 61.3% of SummEval cases, and majority voting across runs improving balanced accuracy against human labels. aclanthology.org
12. [12]US Bureau of Labor Statistics, 15 May 2026, "Occupational Employment and Wages, May 2025", Table 1. Source for 17,753,430 jobs in office and administrative support occupations. bls.gov
13. [13]Simon Willison, 11 September 2025, a link post on Thinking Machines' "Defeating Nondeterminism in LLM Inference". Source for the argument that floating point non-associativity on concurrent GPUs is not the main cause of a model answering differently to the same prompt, that varying batch size under varying server load is, and for the batch-invariant kernels (batch_invariant_ops) that make Qwen3-8B answer deterministically under vLLM. simonwillison.net
14. [14]IncidentHub, 20 July 2026, updated 7 October 2026, "H1 2026 Cloud and SaaS Reliability Report". Source for 30,246 outages across 1,082 providers from 1 January to 30 June 2026, May 2026 as the busiest month with 6,070, Azure's East US control plane incident of 24 April lasting over 12 hours, AWS's me-central-1 and me-south-1 regions damaged by drone strikes in March with normal operations expected to take months, and the note that Anthropic's suspension of Claude Fable 5 and Mythos 5 from 13 June to 1 July (18 days 19 hours) is outside the window and not counted as an outage. blog.incidenthub.cloud