Two numbers from one Anthropic report are going round: AI agents worked six hours for every human hour, and the share of junior technical hires fell from 27 per cent to 9. Both are real, both come from Anthropic’s own workforce, and at least one widely read write-up quotes them without their definitions. One headline went as far as saying the company had all but stopped hiring junior staff, which the report’s own figures do not say. This is a reader’s guide to what the numbers measure, and to what the outside evidence says beside them.
Preface
Why I am writing this. A company that sells AI has published a careful account of what AI did to its own workforce. That is useful. It is also one entity doing the treating, the measuring and the reporting, with no outside comparison, and two of its headline numbers carry definitions that change what they say. A reader should know those definitions before repeating either number.
Terms. The six-to-one figure is the ratio of employee-directed, interactive agent working time to human working time. The 27-to-9 figure is the share of technical hires who were junior. An outside comparator is a comparison population, outside the firm, that did not get the treatment. The report has internal comparisons, and section VI says what they can and cannot do.
Interest. The interest belongs here, at the start, and it runs both ways. My own frames, the Frozen Workforce (a workforce held in place by AI-driven hiring freezes) and Ghost GDP (output that grows without matching human income), would be helped if these numbers are right, so I tried to read against them. One example: my first draft said the hires line had been cherry-picked at its endpoints, and two reviewers showed me it survives smoothing, so that claim is gone. My employer also sells assurance and governance advice, and it would be a candidate to do the independent measurement I recommend.
Assumptions. I assume the report’s authors did what they say they did. I do not question their honesty, and the report is candid about several limits.
Confidence. High on the definitions, which I read in the report’s own text. Medium on values I read off the two charts, which are approximate. Low on what caused the shift in junior hiring, which the report does not claim to establish.
What I will not claim. That the numbers are wrong. That AI is not reducing entry-level hiring. That Anthropic misled anyone. That one firm’s workforce says nothing about the economy.
What was checked, and how. I read the report, “How employment is changing at Anthropic”, published on 16 October 2026, through a reader proxy on 27 October. Its page metadata shows a later modification date of 23 October, and I found no change note on 3 November, when I rechecked my quotations and chart readings against the current text; I cannot compare it with the 23 October text. I viewed both charts. I read the other studies below at their sources, except one version noted in the table. The reading, the checking and the first drafts were done by AI agents under my rules, and I set the claims, edited and published. Chart values are my readings of lines on a figure and could be off by a point or two. I did not contact Anthropic or the authors.
I. What the two numbers are
The report’s key findings include that employee-directed interactive agent hours outnumbered employees six to one by August 2026, and that the share of junior technical hires fell from 27 per cent to 9. They are from the same report, but they are not about the same people. The first covers hours of work by agents that individual employees direct. The second covers hires into technical departments.
The report links them only by timing: the junior decline, it says, occurred as agent hours relative to employees rapidly expanded. It does not claim that one caused the other.
The numbers have already travelled. The Decoder’s headline on 18 October said Anthropic had nearly stopped hiring junior staff because agents now work six times as many hours. Its text gives the shares, as the report does, but no counts, and does not mention that human hours are assumed. The report says technical hiring continued to grow in absolute terms.
Why this matters: two numbers in one document look like one story, and the report is careful not to say they are.
II. Six to one is a stopwatch
The note under the report’s second chart defines the measure. Agent time is runtime, including inference and active tool use, excluding any wait of more than half an hour. Human time is, in the report’s words, “Human work hours are calculated as headcount and assumed 40 hours per week.”
Each choice pushes the ratio one way or the other, and the net direction is not known.
- Pooling parallel agents, up. The chart’s note says only that subagents are pooled. The arithmetic implies the hours add up: six to one is 240 agent-hours for each employee-week, more than the 168 hours in a week, which is about 1.4 agents running around the clock per employee.
- An assumed 40-hour week, up. If people work longer than forty hours, the true ratio is lower. The human side is assumed, not observed.
- Interactive sessions only, down. The chart leaves out autonomous agents. The report’s text says that all agent hours, autonomous and interactive together, passed human hours in March 2026, while the plotted interactive line crosses parity around May. So six to one is a floor on one axis.
- Active time, not output. The report says sessions spend around 80 per cent of their time waiting for human input. It does not say how that waiting enters the ratio, and I read the figure as active runtime, not a measure of work finished.
The chart plots the ratio on a log scale. All employees reach about six times by late August, and a line for research organisations reaches about nineteen. Both are the last point plotted, not averages. The chart begins in February 2026 at about 0.13 for all employees and 0.19 for research organisations.
None of this makes six to one wrong. It is a measure of how much agents run per assumed human hour at one firm, closer to adoption than to productivity.
Why this matters: the figure counts runtime against an assumed work week, with biases in both directions, and it was never a ratio of output.
III. Twenty-seven to nine is a real trend with no counts
The junior hiring chart has three lines for technical departments: hires, headcount and open roles. By my reading, the hires line goes from about 27 per cent in May 2025 to about 9 per cent in June 2026. It is a steady fall of around a point and a half a month with a couple of points of wobble. The average of the first three months is about 25 per cent and of the last three about 13. The fall survives smoothing, and I would not call it noise.
So the complaint is not about volatility. It is about what is missing. The report gives no counts or confidence intervals for the hiring shares, and the last point, 9 per cent, is the least precise. It never defines junior or technical, using phrases such as junior-level and lower band technical positions. A share of hires can fall while the number hired does not, and the report says technical hiring continued to grow in absolute terms. A rough check from three of its own figures: the report says the firm tripled in size, which I read as headcount, the technical share of employment fell from 61 to 56 per cent, and the junior share of technical headcount fell from 21 to 14 per cent. That implies junior technical headcount grew by something like 80 per cent while its share fell, or 65 to 100 per cent if “tripled” means anything from 2.7 to 3.3 times. The windows for the shares also differ. Growth that slow relative to the firm is still a lost opportunity, but it is not a cut.
There is one unexplained step. The junior share of technical headcount drops from about 21 per cent in May 2025 to about 16 in June, in a single month. A stock share does not usually move like that. It looks more like re-banding or a large cohort event, and the report does not say.
On timing, the hires series runs from May 2025 to June 2026 and the agent-hours chart starts in February 2026, so the first nine months of the hires series have no agent series to compare. But the steepest fall in hires, from about 19 per cent in March 2026 to 9 in June, sits inside the window, and the report dates a sharp fall in postings for juniors from December 2025. The two charts also end on different dates. The report’s own account is that the decline occurred as agent hours expanded, and that is where its evidence stops.
The non-technical departments’ junior hire share stays between roughly 22 and 32 per cent and ends at 26, while junior open roles there fall from about 31 to 13 per cent.
Why this matters: the fall is real, and what is missing are the counts, the definitions and an explanation of the one-month step.
IV. What the report says it cannot show
To its credit, the report is candid. It describes one pattern on junior employment as descriptive: junior employment is slowing in jobs that use Claude for the most complex tasks. For the fall in the technical share of employment it offers two alternative explanations: the natural expansion of business functions as the company scales, and coding tools letting non-technical employees build software themselves. The section that asks whether AI is driving productivity answers for productivity, not for hiring.
On the productivity work it notes that tool adoption is not random, so employees who tool up early may be about to do something big, which biases the estimate. The study followed 192 engineers as they adopted coding agents against 74 who did not.
The report’s evidence on why hiring changed is a manager survey, seventeen employee interviews and an analysis of hiring scorecards. A majority of managers said they were hiring for a different set of skills because of AI, and about a third of machine-learning managers and a fifth of software engineering managers said they were hiring less because of it. The report gives no sample size for the survey, and no breakdown of the other managers.
Why this matters: the authors have already said most of what I am saying, and the quotations in circulation will not carry it.
V. What the outside evidence says
Several studies point the same way for young workers entering exposed occupations, and they measure different things in different units.
| Study | Measure | Result, as a relative change | The authors’ own caveat |
|---|---|---|---|
| Stanford Digital Economy Lab, August 2026 update | Payroll records, ages 22 to 25 in the most exposed jobs against similar-aged workers elsewhere, to June 2026 | About 19 per cent below the counterfactual, up from 15; no comparable gap for experienced workers | “descriptive patterns, not causal estimates of the effect of AI” |
| Dallas Fed, January 2026 | Employment share, ages 20 to 24 in the most exposed occupations, 2022 to 2025 | 16.4 to 15.5 per cent of employment, a fall of 0.9 points or about 5.5 per cent in relative terms, mainly fewer entrants | The pattern may not be causal |
| Hosseini and Lichtinger, résumé and posting data, May 2026 version | Junior employment at firms identified as adopting generative AI, against non-adopters | About 9 per cent lower after six quarters, mainly through slower hiring | Adoption is not random; the authors call the break suggestive |
| Anthropic’s March 2026 study, by two of the same authors as the October report | Job-finding rate, ages 22 to 25 entering exposed occupations | About 14 per cent lower than in 2022 | “just barely statistically significant”; not independent of the report |
One study disagrees. A New York Fed note of May 2026 looked at job postings and found “little indication of a distinct AI-driven decline in labor demand”, with junior and senior demand in exposed occupations moving broadly in parallel. It measures postings, not hires, and its event-study data end in 2025.
Read together, the pieces say that young workers entering exposed occupations are doing worse than similar workers elsewhere in most of the payroll and résumé evidence, mostly because fewer are being taken on, and that the size runs from single digits to a fifth depending on the measure, the data and the period. They are not on one scale, so the spread is partly a spread of definitions. None claims to establish cause, and the New York Fed postings evidence does not see a break at all.
That is a different claim from “junior hiring fell from 27 to 9 because of agents”, and the gap between the two is where the argument should happen.
Why this matters: outside evidence supports the direction in most places and disputes the precision, and one of the four studies is not outside.
VI. There is no outside comparator
Here is the structural point. The firm that treated its staff with agents also measured the effect and reported it. That is one entity in three roles, with no comparison population outside it.
The report does have internal comparisons. The non-technical departments’ junior hire share drifted down about five points while the company tripled, though their junior open roles fell much further. Never-adopting engineers serve as controls in the productivity study. And the report compares junior growth across jobs that use Claude more and less. Each is confounded: the firm itself says tool adoption is not random, it tripled in size in a year, and jobs differ in ways that exposure may not capture. They are useful and they are not an outside counterfactual.
I have written before about labs grading their own homework in public, in another context. This report is a good example of the genre because it publishes methods for the productivity work and says what it cannot show. But a report can be careful and still be quoted carelessly, as the Decoder headline shows, and the six-to-one figure may keep being repeated as if it measured output.
Why this matters: a careful self-report is still a self-report, and the quotations will drop the care.
VII. What an independent test would look like
A multi-firm design already exists, in the Hosseini and Lichtinger paper, which compares firms that adopt generative AI with those that do not, and which has its own selection problem. A better test would add four things. Counts of hires with intervals, and a fixed definition of junior. Agent activity measured against something produced, not only against time. The series published in full, so that readers can see the whole line and not its endpoints. And, from Anthropic specifically, the flows behind the shares: hires, promotions out of the junior band, leavers and band definitions.
Exact counts may be sensitive for a company of this size, and individual-level data could identify people. Ranges, bounds or bootstrap intervals would answer most of the question without that risk.
Why this matters: the evidence to settle this exists in pieces, and what is missing is the part only the firm has.
VIII. What would change this view
A prediction first, labelled as one. I expect that by 30 June 2027, if counts are published by Anthropic or others, the trailing six-month junior share of technical hires will be between 8 and 15 per cent, so the fall will have held but not deepened. I hold that loosely, and it is my guess.
I would change my mind if a multi-firm dataset with staggered adoption shows no relative decline in junior hiring at adopting firms, which would suggest the Anthropic pattern is about Anthropic’s growth. I would change it the other way if the same data show a clear gap. If the junior share of technical hires is back above 15 per cent for two consecutive quarters, the endpoint was a dip and not a trend.
Why this matters: the claim should be easy to refute, and one more data release would do it.
IX. The Case Against This Essay
A critique of a careful report deserves its own objections.
First, I am reading values off a chart. They are approximate, and the report’s own text says the junior share fell from 27 to 9.
Second, I am being hard on a report that is unusually open. A reader could fairly say that criticising an honest account discourages the next one. The report does publish methods and sample sizes for its productivity analysis, and Anthropic has published related workforce studies before.
Third, the junior decline may be exactly what it looks like. Most of the outside evidence points the same way, and the report does not claim to have established the cause. My argument is about how the numbers are quoted, and it can be over-read as scepticism about the trend. I think the trend is real.
Fourth, I did not obtain the underlying data and could not test the report’s methods myself. I read the text and the charts, and I did not read the November 2025 version of the Hosseini and Lichtinger paper.
Fifth, the managers’ own stated reasons are the nearest thing to a mechanism for hiring in the report. They are weak on magnitude and not nothing, and I have not given them their due.
Sixth, the interest runs both ways, as the Preface says, and I cannot prove that I read against it. I can only say which claim I dropped.
Seventh, the check is only partly independent. The agents did the reading and the first reviews, and none of them is a labour economist.
X. Asks
Four audiences, small requests.
Anthropic. Publish the flows behind the hiring shares as ranges or intervals, with a definition of junior and technical and the full monthly series for both charts. If you can, release agent-use data to outside researchers under privacy protections.
Journalists and analysts. When you quote six to one, say it is employee-directed interactive agent runtime against an assumed forty-hour week at one firm. When you quote 27 to 9, say it is the share of technical hires at that firm, with no counts published.
Boards and executives. Do not benchmark your own entry-level plan against one lab’s workforce. Measure your own hires, with counts, before and after you give people agents.
Statistical agencies and ministries. Give researchers a way to track entry-level hiring by occupation exposure, with the definitions published. Singapore’s manpower data is a natural place to start.
Both numbers will be quoted for a year. The least we can do is quote them with their definitions.
Notes and sources
- Anthropic, “How employment is changing at Anthropic”, published 16 October 2026 (page metadata shows a modification on 23 October; I found no change note when I checked again on 3 November), by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory. Primary; read in full through a reader proxy on 27 October and again by a reviewer on 3 November; both charts viewed. Chart values in this essay are my readings and approximate. anthropic.com
- Anthropic, study of AI’s labour-market impacts (the page title uses US spelling), 5 March 2026, by Maxim Massenkoff and Peter McCrory. Primary; two of its authors also wrote the October report. anthropic.com
- Stanford Digital Economy Lab, “Canaries in the Coal Mine” update, 12 August 2026, by Erik Brynjolfsson and colleagues. Primary; ADP payroll data to June 2026. digitaleconomy.stanford.edu
- Federal Reserve Bank of Dallas, “Young workers’ employment drops in occupations with high AI exposure”, 6 January 2026, by Tyler Atkinson and Shane Yamco. Primary. dallasfed.org
- Federal Reserve Bank of New York, “Do Job Postings Show Early Labor-Market Effects of AI?”, Liberty Street Economics, 14 May 2026, by Richard Audoly, Miles Guerin and Giorgio Topa. Primary. libertystreeteconomics.newyorkfed.org
- Seyed Mahdi Hosseini and Guy Lichtinger, “Generative AI as Seniority-Biased Technological Change”, SSRN 5425555, version of 6 May 2026. Primary; read by a reviewer on an accessible copy, because SSRN returned an error. “Six quarters” counts from 2023Q1, not from each firm’s adoption. papers.ssrn.com, raw.githack.com
- The Decoder, article of 18 October 2026 on Anthropic’s junior hiring and agent hours, headline paraphrased in section I. Secondary; cited for the headline. the-decoder.com
- A note on earlier work. I discussed the Anthropic figures briefly in “Receipts, Please” and the wider labour argument in “Paid to Disappear”, both on this site. genesishumanexperience.com
Declarations
Interest. My own frames, the Frozen Workforce and Ghost GDP, would be helped by these numbers being right, and my employer sells assurance and governance advice, which independent measurement of AI’s effects on work would help and which would be a candidate to do it.
Not checked. The report’s underlying data and counts; the definitions of junior and technical; the November 2025 version of the Hosseini and Lichtinger paper; Danish and other studies on earnings and hours; whether coverage of the report carried its qualifiers beyond the examples I saw; anything Anthropic has said about the report since 16 October.
Not contacted. Anthropic, the report’s authors, and the authors of the outside studies. A right of reply would be the first addition.
Dr Luke Soon is an AI Leader at PwC Singapore, covering fourteen Asia Pacific markets, and co-author of Singapore’s Model AI Governance Framework for Agentic AI. He writes at genesishumanexperience.com under the Genesis: Human Experience in the Age of Artificial Intelligence banner.
Leave a Reply