Work, jobs and the labour market

Hirji, Rahim

Each entry gives the study, how it was done, what it found, what it supports and, as plainly, what it does not. Grades describe what kind of evidence a source is; how to read them is on the main page. Every entry has a permanent link and a copyable citation.

Compiled review#

Barney, J. (1991). Firm Resources and Sustained Competitive Advantage

Journal of Management, 17(1), 99-120, March 1991. DOI 10.1177/014920639101700108

Method: Theoretical article in strategic management. No new data. Builds on the stated assumptions that strategic resources are heterogeneously distributed across firms and that those differences are stable over time.
Finding: Sets out four empirical indicators of the potential of a firm resource to generate sustained competitive advantage: value, rareness, imitability and substitutability. Applies the model to several firm resources and draws out implications for other business disciplines. The founding statement of the resource-based view.
What it supports: That a widely used and long-established test exists for whether a resource can confer advantage, against which a purchasable AI licence can be assessed.
What it does not support: Anything empirical, and nothing about AI, which postdates it by three decades. It is a theory with a substantial critical literature, and applying it to a technology three years into commercial diffusion is an argument rather than a finding. The abstract read at source uses rareness, imitability and substitutability; the VRIN acronym is later shorthand and does not appear in the paper's abstract.

competitive advantage · resource-based view · strategy · if everyone has ai where is the advantage

Peer-reviewed#

Garicano, L. (2000). Hierarchies and the Organization of Knowledge in Production

Journal of Political Economy, 108(5), 874-904. University of Chicago Press

Method: Theoretical model of knowledge acquisition and problem-solving in production, with communication costs and knowledge acquisition costs traded off against each other. No empirical estimation.
Finding: A knowledge-based hierarchy is a natural way to organise the acquisition of knowledge when matching problems with those who know how to solve them is costly. Production workers acquire knowledge of the most common or easiest problems and refer exceptions upward to specialist problem solvers, with problems passed on until somebody solves them or the conditional probability of a solution is too low to justify continuing. Adding layers of problem solvers raises the utilisation rate of knowledge and economises on knowledge acquisition, at the cost of increasing the communication required.
What it supports: That the number of layers in an organisation is a function of two costs rather than of custom, which gives a testable prediction for what happens when either cost falls. It is the model every subsequent claim about AI flattening organisations implicitly relies on.
What it does not support: Anything measured. It is a model, published a quarter of a century before generative AI, and it contains no data about firms, layers or technology adoption.

knowledge hierarchy · organisational design · span of control · problem solving · what happens to work that moves information

Peer-reviewed#

Gathmann, C. and Schoenberg, U. (2010). How General Is Human Capital? A Task-Based Approach

Journal of Labor Economics, 28(1)

Method: German administrative employment panel, using task overlap between occupations to measure task-specific human capital.
Finding: Task-specific human capital accounts for up to 52 per cent of overall wage growth. Workers move to occupations with similar task profiles, and the distance of those moves shrinks with experience.
What it supports: That skill is portable along task lines rather than being either fully general or locked to one job.
What it does not support: That broad general education transfers well. If anything it argues the opposite, which complicates the study-anything advice rather than supporting it.

· what should i tell my children to study

Working paper#

Frey, C. B. and Osborne, M. A. (2013). The Future of Employment: How Susceptible Are Jobs to Computerisation?

Oxford Martin School working paper, 17 September 2013. Later published in Technological Forecasting and Social Change, 114 (2017)

Method: Probability of computerisation estimated for 702 detailed US occupations using a Gaussian process classifier, trained on 70 occupations hand-labelled by machine learning researchers at an Oxford workshop.
Finding: In the authors' own words: 'about 47 percent of total US employment is at risk'. The paper's title asks how SUSCEPTIBLE jobs are, and the estimate is of technical susceptibility to computerisation, not a forecast of job losses.
What it supports: That a large share of US employment sits in occupations whose tasks were, in 2013, judged technically susceptible to computerisation.
What it does not support: That 47 per cent of jobs will be, or have been, lost. It is not a prediction, carries no date attached to any loss, and models whole occupations rather than tasks within them. The routine citation as '47 per cent of jobs will disappear' reverses what the paper claims. The working paper is 2013; the journal version is 2017, and the two dates are frequently confused.

labour market · the most quoted ai statistics checked

Peer-reviewed#

Bloom, N., Garicano, L., Sadun, R. and Van Reenen, J. (2014). The Distinct Effects of Information Technology and Communication Technology on Firm Organization

Management Science, 60(12), 2859-2885. DOI 10.1287/mnsc.2014.2013. Earlier version NBER Working Paper 14975, May 2009

Method: Survey data on worker and plant manager autonomy and span of control, combined with measures of technology adoption, and instrumented using distance from ERP's place of origin and heterogeneous telecommunication costs arising from regulation. The working-paper version describes approximately 1,000 manufacturing firms with 100 to 5,000 employees across the US, France, Germany, Italy, Poland, Portugal, Sweden and the UK, drawn from the CEP double-blind management survey, with technology data from the Harte-Hanks ICT panel.
Finding: Information technology is a decentralising force and communication technology is a centralising force. Better information technologies, ERP for plant managers and computer-assisted design or manufacturing for production workers, are associated with more autonomy and a wider span of control. Technologies that improve communication, such as data intranets, decrease autonomy for workers and plant managers. Instrumenting strengthens the result.
What it supports: That 'technology' has no single organisational effect, and that predicting what a tool does to authority requires knowing whether it lowers the cost of knowing or the cost of telling.
What it does not support: Anything about generative AI, which is both kinds of technology in one interface. Manufacturing plants only, data predating 2009. The sample description above is verified in the working paper; the published article abstract says only 'American and European manufacturing firms', and the typeset article could not be opened at the publisher.

decentralisation · organisational design · span of control · autonomy · what happens to work that moves information

Peer-reviewed#

Bloom, N., Garicano, L., Sadun, R. and Van Reenen, J. (2014). The Distinct Effects of Information Technology and Communication Technology on Firm Organization

Management Science, 60(12), 2859-2885, 2014. DOI 10.1287/mnsc.2014.2013. Circulated as NBER Working Paper 14975, May 2009, DOI 10.3386/w14975. PROVENANCE: the abstract was read at source on the NBER working-paper record on 21 September 2026, which also confirms the published venue. The published article is paywalled and was not opened; the LSE copy returns an error page. No coefficient or sample size is quoted here.

Method: An original dataset of firms in the United States and seven European countries, measuring production-worker autonomy, plant-manager autonomy and spans of control against separate measures of information technology and communication technology. Identification supported by 'exogenous variation in cross-country telecommunication costs arising from differential regulatory regimes'. Sample size not read at source.
Finding: In the authors' words, 'better information technologies (Enterprise Resource Planning for plant managers and CAD/CAM for production workers) are associated with more autonomy and a wider span of control. By contrast, communication technologies (like data networks) decrease autonomy for both workers and plant managers.' Their account is that technologies reducing information costs let lower-level agents acquire more knowledge, and so widen what those agents may settle for themselves, while technologies reducing communication costs 'substitute agent's knowledge for directions from their managers, and lead to centralization'.
What it supports: That decision rights inside firms move when technology changes the cost of information or of communication, and that the direction of the move depends on which of the two costs falls. It is the measured counterpart to the claim that an unallocated decision right drifts rather than staying put.
What it does not support: Anything about AI, which postdates the technologies measured. The results are associations between technology use and organisational form, with one instrument carrying the causal argument, on manufacturing plants. Nothing here shows that an explicit allocation of decision rights produces better decisions.

decision rights · delegation · autonomy · organisational design · centralisation · how should ai decision rights be allocated

Institutional modelling#

United States Bureau of Labor Statistics (2018). Occupational Projections Evaluation, 2006 to 2016

BLS Employment Projections programme

Method: The agency's own retrospective scoring of its 2006 ten-year projections for 840 detailed occupations against actual 2016 outcomes.
Finding: BLS correctly projected whether an occupation would grow or decline 78 per cent of the time, but correctly projected which occupations would grow faster than the economy as a whole only 57 per cent of the time. Projected average occupational growth was 10.4 per cent against an actual 3.6 per cent, the gap driven by a recession the projections could not foresee.
What it supports: That the only organisation which scores its own occupational forecasts gets the useful question, relative growth, barely better than a coin toss, and that its largest errors come from shocks.
What it does not support: That forecasting is worthless. Direction of change at aggregate level was reasonably good. It also predates generative AI, and no equivalent scored record exists for a technological discontinuity.

· what should i tell my children to study

Peer-reviewed#

Acemoglu, D. and Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor

Journal of Economic Perspectives, 33(2), 3-30. DOI 10.1257/jep.33.2.3

Method: Task-based theoretical framework in which production is allocated between capital and labour, with an empirical decomposition of US industry-level data over recent decades.
Finding: Automation shifts the task content of production against labour through a displacement effect, and therefore ALWAYS reduces the labour share in value added, and may reduce labour demand even while raising productivity. The counterweight is the creation of new tasks in which labour has a comparative advantage, which always raises the labour share. Their decomposition attributes slower employment growth over three decades to an accelerating displacement effect, a weaker reinstatement effect and slower productivity growth.
What it supports: That the distributional question is separate from the productivity question, and that a technology can raise output while reducing labour's share of it. This is the framework almost every serious argument in this area now runs through.
What it does not support: That AI specifically will behave this way. The empirical work predates generative AI and concerns industrial automation and robotics.

labour market · who captures the productivity gains from ai

Peer-reviewed#

Brynjolfsson, E., Rock, D. and Syverson, C. (2021). The Productivity J-Curve: How Intangibles Complement General Purpose Technologies

American Economic Journal: Macroeconomics, 13(1), 333-72, January 2021. DOI 10.1257/mac.20180386. Earlier version NBER Working Paper 25148

Method: Theoretical model of general purpose technology adoption with unmeasured intangible complementary investment, applied to US national accounts data on computer hardware and software.
Finding: General purpose technologies enable and require significant complementary investments that are often intangible and poorly measured in national accounts. This produces underestimation of productivity growth in a new technology's early years and overestimation later, when the benefits of the intangible investments are harvested, a pattern the authors name the Productivity J-curve. Adjusting for intangibles related to computer hardware and software yields a total factor productivity level 15.9 per cent higher than official measures by the end of 2017. The authors state the AI-related intangible capital effects on measured productivity are currently small but growing.
What it supports: That an early productivity read on a general purpose technology is biased downwards for a structural and quantified reason, and that the direction of the later bias is upwards. It supplies the argument for staging an AI investment review rather than taking a single reading.
What it does not support: That AI will follow the same curve, or on what timescale. The 15.9 per cent figure is for computer hardware and software to 2017, not for AI, and the authors describe current AI intangible effects as small. It is also a measurement argument rather than a forecast of returns to any individual firm.

productivity · intangible capital · general purpose technology · measurement · how long before you know if an ai investment worked if everyone has ai where is the advantage

Peer-reviewed#

Yang, L., Holtz, D., Jaffe, S., Suri, S., Sinha, S., Weston, J., Joyce, C., Shah, N., Sherman, K., Hecht, B. and Teevan, J. (2022). The effects of remote work on collaboration among information workers

Nature Human Behaviour, 6, 43-54. DOI 10.1038/s41562-021-01196-4, published online 9 September 2021

Method: Observed telemetry on emails, calendars, instant messages, video and audio calls and workweek hours of 61,182 US Microsoft employees over the first six months of 2020, using workers already remote before the pandemic as a comparison to separate firm-wide remote work from other pandemic effects.
Finding: Firm-wide remote work caused the collaboration network to become more static and siloed, with fewer bridges between disparate parts of the organisation, a decrease in synchronous and an increase in asynchronous communication. The authors state these effects may make it harder for employees to acquire and share new information across the network.
What it supports: That changing how information moves reorganises who knows what, without anybody redesigning a role, and that the change is measurable in the network rather than only in self-report.
What it does not support: Anything about AI, and nothing about outcomes. It measures communication structure rather than performance, in one very large technology company, during a pandemic. Full text is paywalled; this entry rests on the published abstract.

collaboration · information flow · organisational network · tacit knowledge · what happens to work that moves information where your ai practice lives

Peer-reviewed#

Acemoglu, D. (2024). The Simple Macroeconomics of AI

NBER Working Paper 32487; published in Economic Policy, 40(121), 2025

Method: Task-based macroeconomic model applying Hulten's theorem to existing estimates of AI task exposure and task-level cost savings.
Finding: Estimates total factor productivity gains of no more than 0.66 percent over ten years, revised to under 0.53 percent once the difficulty of hard-to-learn tasks is accounted for. Argues AI is likely to widen the gap between capital and labour income rather than reduce labour income inequality.
What it supports: That plausible macroeconomic gains are an order of magnitude smaller than the headline value estimates in circulation.
What it does not support: That AI is unimportant. It models productivity through task-level cost savings, and would not capture effects running through new products, new tasks or capability change.

productivity · macroeconomics · total factor productivity · inequality · how should leaders respond to ai ai workforce strategy will ai replace my job ai and human judgement does ai actually make people more productive the cost of delegation

Working paper#

Autor, D. (2024). Applying AI to Rebuild Middle Class Jobs

NBER Working Paper 32140, DOI 10.3386/w32140, issued February 2024. Still a working paper: the NBER record carries no published version as at 15 September 2026. Its own acknowledgements say it was drafted for NOEMA Magazine and appeared there as 'How AI Could Help Rebuild The Middle Class'.

Method: An essay rather than a study, written for a general magazine and circulated through the working-paper series. Argument and synthesis rather than empirical test, drawing on the author's prior work on task structure and labour demand.
Finding: Argues that AI's distinctive opportunity is to extend the reach of expertise, letting a wider set of workers with complementary knowledge perform higher-stakes decision tasks currently reserved to elite experts.
What it supports: That there is a serious, well-argued case for AI as an expertise-widening technology rather than an expertise-replacing one.
What it does not support: That this will happen. The author is explicit that the thesis is an argument about what is possible rather than a forecast, and no evidence yet shows it occurring at scale.

expertise · middle-skill work · complementarity · task structure · what is the judgement premium will ai replace my job staying valuable in the age of ai what happens to work that moves information what becomes more valuable as ai gets cheaper pricing work when production is cheap

Working paper#

Bick, A., Blandin, A. and Deming, D. J. (2024). The Rapid Adoption of Generative AI

NBER Working Paper 32966

Method: Nationally representative US surveys of generative AI use at work and at home.
Finding: By late 2024, nearly 40 percent of US adults aged 18-64 used generative AI and 23 percent of employed respondents had used it for work in the previous week, but only 1 to 5 percent of all work hours were assisted.
What it supports: Enormous reach, thin penetration into actual hours. Adoption is not transformation.
What it does not support: Quality of use. Self-reported use counts any use at all.

AI adoption · ai workforce strategy why learn to prompt is weak career advice

Working paper#

Bonney, K., Breaux, C., Buffington, C., Dinlersoz, E., Foster, L., Goldschlag, N., Haltiwanger, J., Kroff, Z. and Savage, K. (2024). Tracking Firm Use of AI in Real Time: A Snapshot from the Business Trends and Outlook Survey

US Census Bureau, Center for Economic Studies Working Paper CES 24-16, March 2024; also NBER Working Paper 32319. Not peer reviewed

Method: Analysis of the AI questions in the Business Trends and Outlook Survey, a high-frequency nationally representative survey of US firms, over the collection period covered by the paper.
Finding: Bi-weekly estimates of the AI use rate rose from 3.7 to 5.4 per cent, with an expected rate of about 6.6 per cent by early autumn 2024. 94.6 per cent of AI-using businesses reported no net change in employment in the previous six months attributable to AI use. On discontinuation, which the survey does not ask about directly, 67.9 per cent of current AI users expect to use it in future, 14.5 per cent do not, and 17.6 per cent do not know, so about one in seven current users may de-adopt; the authors attribute this to experimentation that does not yield anticipated benefits or organisational synergies. The authors state explicitly that the analysis does not seek to identify a causal link between AI use and firm performance, only whether use is associated with better performance in general, and that causal analysis awaits integrated data and repeat collections after about three to four years.
What it supports: That the body with the best firm-level data in the world declines to make the causal claim that consultancy and vendor reporting makes routinely, and says in its own text what would be required to make it.
What it does not support: Any effect of AI on firm performance, by the authors' own statement. Adoption rates here are not comparable with later Census figures, because the question was broadened in November 2025 from use in producing goods or services to use in any business function. Census working papers carry the standard disclaimer that they have not undergone the review accorded Census Bureau publications.

AI adoption · measurement · causal inference · firm performance · what becomes more valuable as ai gets cheaper which ai investments should we stop

Vendor research#

Cisco Systems (2024). Cisco 2024 Data Privacy Benchmark Study (release: More than 1 in 4 Organizations Banned Use of GenAI Over Privacy and Data Security Risks)

Cisco investor news release, 25 January 2024. Seventh edition of the benchmark. Read at source 5 October 2026

Method: Survey of 2,600 privacy and security professionals across 12 geographies. Self-report. Sampling frame and questionnaire not published with the release.
Finding: 27 per cent said their organisation had banned generative AI at least temporarily. 63 per cent had established limits on what data can be entered and 61 per cent on which tools employees can use. 48 per cent admitted entering non-public company information into generative AI tools and 45 per cent employee information.
What it supports: That among privacy and security professionals in early 2024, bans and data limits on generative AI were common, and that entering company and employee information into the tools was widely admitted even by the people responsible for protecting it.
What it does not support: Prevalence across all organisations or workers: the respondents are privacy and security professionals, not a representative sample of employers. It does not show what happened to the information entered or whether any harm followed. Cisco sells security and privacy products, and the 2024 figures may not describe organisations in 2026.

generative AI ban · shadow AI · data protection · AI policy · what are the risks of using ai at work my company banned ai what should i do what should i never put into ai

Peer-reviewed#

Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs

Science, 384(6702), 1306-1308

Method: Human and model ratings of task exposure across occupational task descriptions.
Finding: Around 80 percent of US workers could have at least 10 percent of tasks affected; about 19 percent could see at least half affected.
What it supports: Where pressure is likely to fall across the occupational structure.
What it does not support: That any job will be lost. This is exposure, not displacement, and the authors say so explicitly. It is the most misquoted number in the field.

task exposure · labour market · will ai replace my job

Vendor research#

Microsoft and LinkedIn (2024). 2024 Work Trend Index Annual Report: AI at work is here. Now comes the hard part

Microsoft WorkLab, 8 May 2024. Read at source 18 September 2026

Method: Survey of 31,000 full-time employed or self-employed knowledge workers across 31 countries, fielded 15 February to 28 March 2024, combined with LinkedIn labour market data and Microsoft 365 telemetry. Published by a company that sells the tools being surveyed.
Finding: 78 per cent of AI users said they were bringing their own AI tools to work, 80 per cent at small and medium-sized companies. 52 per cent of people who use AI at work said they were reluctant to admit using it for their most important tasks, and 53 per cent worried that using it on important tasks made them look replaceable. 60 per cent of leaders worried that their organisation's leadership lacked a plan and vision to implement AI.
What it supports: That reluctance to disclose AI use and fear of appearing replaceable were reported by about half of AI-using knowledge workers across 31 countries in early 2024, two years before the same pattern was measured in the UK.
What it does not support: Any outcome, and the population is knowledge workers who already use AI. Self-report, with a sponsor whose products are the subject, and the questionnaire is not published in full.

BYOAI · shadow AI · AI adoption · knowledge workers · disclosure · what should we tell employees about ai and headcount should i tell my boss i used ai how do we train people to use ai my company banned ai what should i do

Working paper#

Xu, F. F. and others (2024). TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks

arXiv 2412.14161, 18 December 2024; v3 10 September 2025. Read at source 24 October 2026

Method: Simulated software company with tasks across roles, scored for autonomous completion.
Finding: The most competitive agent completed 30 per cent of tasks autonomously.
What it supports: That agents in 2025 failed most simulated office tasks.
What it does not support: Current performance, or performance in a real organisation.

AI agents · benchmark · agent boss · what is an agent boss

Employer policy#

Anthropic (2025). How to collaborate with Claude during our hiring process

Anthropic, candidate guidance page, 'Last updated Jul 10, 2025' on the page. Read at source on 19 October 2026.

Method: An employer's published rule for applicants. Anthropic is also the developer of the Claude models, so the rule is one firm's position on its own product.
Finding: For resumes and application questions: write the first draft yourself, then use Claude to refine it; the experiences must be the candidate's own, and using Claude to create experiences the candidate has not had is not allowed. Take-home assessments are completed without Claude unless the company says otherwise, and live interviews are done without AI assistance unless the company says otherwise. Using Claude to research the company, practise answers and prepare questions is encouraged. The page does not mention cover letters.
What it supports: That one major AI developer, as of July 2025, told applicants to draft the substance themselves and use the tool for refinement and preparation, and to avoid it in assessments and interviews unless told otherwise.
What it does not support: What any other employer requires, whether the rule is enforced or how often it is broken, or what happens to a candidate who disregards it. The page may have changed since it was read.

applications · AI disclosure · assessment · hiring · can i use ai to write a job application

Operator account#

Arvind Krishna, reported by The Wall Street Journal and PYMNTS (2025). IBM CEO: layoffs due to AI led to 'more investment' in other roles

PYMNTS, 6 May 2025, reporting a Wall Street Journal interview of the same day

Method: A chief executive's account of his own company in a newspaper interview. IBM is reported to have replaced a few hundred human resources staff with AI agents doing spreadsheet analysis, research and email drafting; the timeframe is not stated.
Finding: Krishna: 'While we have done a huge amount of work inside IBM on leveraging AI and automation on certain enterprise workflows, our total employment has actually gone up, because what it does is it gives you more investment to put into other areas.' The areas named are software engineering, sales and marketing, which he characterised as work needing critical thinking and dealing with people.
What it supports: That one large employer describes the same decision the other way round from Klarna: automate the routine, then choose where the freed money goes. The choice of where it goes is the leadership act, and IBM's chief executive presents it as one.
What it does not support: That AI raised IBM's headcount; total employment moves for many reasons and no figures separating them were given. 'A few hundred' HR roles is the company's own round number. It is a chief executive's description of his own choices.

ai leadership · ibm · reinvestment · headcount · work design · automation · ai leadership

Peer-reviewed#

Autor, D. and Thompson, N. (2025). Expertise

NBER Working Paper 33941, DOI 10.3386/w33941, issued June 2025; published in the Journal of the European Economic Association, 23(4), August 2025, 1203-1271, DOI 10.1093/jeea/jvaf023. Originally the Joseph R. Schumpeter lecture to the European Economic Association, Rotterdam, 29 August 2024. Both records re-read at source 15 September 2026 and the journal details reproduce exactly.

Method: Four decades of task data across 303 US occupations, 1980-2018, with a novel content-agnostic measure of task expertise.
Finding: Automation that removed the LESS expert tasks raised wages and reduced employment. Automation that removed the EXPERT tasks lowered wages and increased employment.
What it supports: That which tasks are automated matters more than how many, and gives a testable way to ask whether a given role is appreciating or commoditising.
What it does not support: Anything measured about generative AI. The data ends in 2018, so this is a lens, not a forecast.

expertise · task automation · human capability · what is the judgement premium staying valuable in the age of ai will ai replace my job which tasks do workers not want automated what becomes more valuable as ai gets cheaper

Peer-reviewed#

Babina, T., Fedyk, A., He, A. and Hodson, J. (2025). Firm Investments in Artificial Intelligence Technologies and Changes in Workforce Composition

Chapter 3 in Technology, Productivity, and Economic Growth, NBER Studies in Income and Wealth 83, University of Chicago Press, July 2025. Earlier version NBER Working Paper 31325

Method: Worker resume and job-posting datasets combined to measure firm-level AI investment against workforce composition variables including educational attainment, specialisation and hierarchy.
Finding: AI investments are associated with a flattening of firms' hierarchical structure, with significant increases in the share of workers at the junior level and decreases in the shares in middle-management and senior roles.
What it supports: That the composition of the flattening, and not only its existence, is measurable, and that the layer shrinking is the one that has historically developed juniors into seniors.
What it does not support: Any magnitude. The abstract read for this entry states direction and significance and no percentage, so none should be attributed to it. Association rather than causation, and firms that invest in AI differ from those that do not in many other ways.

middle management · workforce composition · flattening · AI investment · what happens to work that moves information

Institutional survey#

CIPD; Jake Young and Derek Tong (2025). CIPD Good Work Index 2025

CIPD, June 2025

Method: CIPD/YouGov UK Working Lives survey of 5,017 UK employees, fielded in January and February 2025 with quotas and weights to ONS figures on gender, working hours, organisation size, sector, industry and age. Annual series since 2018.
Finding: 16% of UK employees say job tasks have been automated by AI, ranging from 25% of 18 to 24 year olds to 6% of those aged 55 and over. Of those, 85% say it improved their performance and 5% say it worsened it. Tasks automated were most often repetitive (69%), complex (45%) or dangerous (17%). Employees with AI-automated tasks report higher job satisfaction (78% versus 53%) and are more likely to say work has a positive effect on mental health (72% versus 39%).
What it supports: That in early 2025 AI had touched the tasks of a minority of UK workers, concentrated among the young, and that those workers report better job quality than others.
What it does not support: The satisfaction gap is not causal; younger, higher-skilled workers in adopting sectors may already have better jobs. All measures are self-reported and the AI questions are a small part of a broader job quality survey.

uk workforce · job quality · task automation · job satisfaction · age differences

Working paper#

Ewens, M. and Giroud, X. (2025). Corporate Hierarchy

NBER Working Paper 34162, issued August 2025, revised October 2025. DOI 10.3386/w34162. Not peer reviewed

Method: A measure of corporate hierarchy for over 3,100 US public firms, built from online resumes of 7 million US-based workers, 2016 to 2023, using a network estimation technique to identify hierarchical layers. AI adoption is proxied by AI job postings following Babina, Fedyk, He and Hodson.
Finding: Firms average ten hierarchical layers and a pyramidal structure, with the average and median number of layers declining across the sample period. More hierarchical firms show a more educated workforce, higher internal promotion rates, longer tenure, higher operating performance and higher administrative costs. Companies flattened their hierarchies following adoption of AI technologies, while pharmaceutical companies added layers after Covid-19. The authors state the AI tests are under-powered and that point estimates are significant at the 10 per cent level regardless of the adoption metric.
What it supports: That the Garicano prediction has now been tested against firm-level data and points in the predicted direction, which is more than the flattening discourse previously had.
What it does not support: That AI causes flattening. The authors call their own tests under-powered at the 10 per cent level, adoption is measured by job postings rather than by use, hierarchy is inferred from self-reported resumes, and the sample is US public firms. Not peer reviewed. Figures of 2,500 firms and 16 million employees circulate from an earlier draft and are wrong for this version.

organisational design · flattening · middle management · AI adoption · what happens to work that moves information

Vendor research#

Faros AI (2025). The AI Productivity Paradox Report 2025

Faros AI, 23 July 2025. Read at source 24 October 2026

Method: Telemetry from more than 10,000 developers in 1,255 teams, comparing high and low AI adoption.
Finding: High-adoption teams completed 21 per cent more tasks and merged 98 per cent more pull requests while review time rose 91 per cent; no significant improvement at company level.
What it supports: That in these teams review time grew with AI-driven output.
What it does not support: Causation; the data are correlational and the vendor sells engineering analytics.

verification bottleneck · code review · software · what is the verification bottleneck

Working paper#

Galdin, A. and Silbert, J. (2025). Making Talk Cheap: Generative AI and Labor Market Signaling

arXiv:2511.08785, v1 submitted 11 November 2025, the only version listed on 19 October 2026. No journal reference, comments field or other publication found. Read in the arXiv abstract page, HTML and PDF.

Method: Observational data from Freelancer.com, 1 January 2021 to 26 July 2024: about 61,000 job postings, 2.7 million applications (called proposals) and 212,000 workers, split at ChatGPT's release on 30 November 2022. Restricted to fixed-price, English-language coding jobs with budgets of $30 to $250 posted by verified employers. A language model (Llama 4 Maverick) scored how tailored each application was to its posting; all roughly 960,000 pre-release applications and about 39 per cent of later ones were scored. A structural model estimated on pre-release data supplies the counterfactual. A reputation-score dollar result in the paper was not used because its table labels could not be matched.
Finding: A one standard deviation increase in tailoring was worth a $25.67 lower bid to employers before the release and $14.85 in the later sample, in the considered-applications specification (Table 2, text of sections 4.1 and 4.3.2). The abstract says employers paid for tailoring before language models 'but not after'; the introduction says the patterns weaken significantly or disappear completely. In the structural counterfactual where written applications carry no signal, top-quintile workers are hired 19 per cent less often and bottom-quintile workers 14 per cent more often.
What it supports: That on this platform the market price of a tailored written application fell after language models arrived, in the way a collapsing costly signal would predict.
What it does not support: Causation: the authors call parts of the evidence suggestive. The 19 and 14 per cent figures are model output under the authors' assumptions, not observed outcomes. Freelance coding jobs of a narrow price band, a preprint, and a measure of tailoring built by a language model checked against human readers on a small subset only.

costly signal · labour market signalling · cover letters · hiring · can i use ai to write a job application

Institutional survey#

Georgetown University Center on Education and the Workforce (2025). The Major Payoff: Evaluating Earnings and Employment Outcomes Across Bachelor's Degrees

Georgetown CEW

Method: American Community Survey earnings and employment data across 152 majors for prime-age workers and 142 for early career.
Finding: Median prime-age earnings run from 58,000 dollars in education and public service to 98,000 in STEM. Within STEM alone the range is 64,000 to 146,000, and several humanities majors beat the STEM 25th percentile.
What it supports: That the spread within a field is often wider than the gap between fields, which undercuts advice given at the level of STEM against humanities.
What it does not support: Anything about the future. It is a cross-sectional snapshot of people already employed.

· what should i tell my children to study

Working paper#

Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI

NBER Working Paper 33777, revised March 2026

Method: Adoption surveys linked to administrative labour records, roughly 25,000 workers across 7,000 Danish workplaces in 11 exposed occupations.
Finding: Precise null effects on earnings and hours two years after ChatGPT, ruling out effects larger than 2 percent, alongside substantial task reorganisation and new tasks in AI oversight and integration.
What it supports: That the structure of work moves well before earnings do, and that pay is the slowest available indicator.
What it does not support: That the same holds elsewhere. Denmark is high-trust, high-wage and heavily unionised, and two years is early.

labour market · task reorganisation · AI adoption · ai workforce strategy will ai replace my job ai and human judgement is ai already causing job losses

Institutional survey#

Institute of Student Employers (2025). Student Recruitment Survey 2025

Institute of Student Employers, 2025, with 2026 outlook published 7 January 2026

Method: Trade association survey of 155 employers covering over 31,000 student hires from more than 1.8 million applications in the 2024-2025 cycle.
Finding: An average of 89 applications per vacancy. The ISE reports a projected 7 per cent drop in student vacancies for 2026, while 30 per cent of employers increased student hiring.
What it supports: That the contraction is uneven. Nearly a third of employers increased student hiring in a falling market, which cuts against any account of uniform collapse.
What it does not support: Whole-market figures. It is a membership survey, and its members are organisations that run structured student recruitment in the first place.

graduate recruitment · early careers · applications · do apprenticeships still work

Operator account#

Luis von Ahn, reported by Fortune (2025). Duolingo CEO walks back AI-first comments: 'I do not see AI as replacing what our employees do'

Fortune via Yahoo Finance, May 2025, reporting an April 2025 all-hands email and a LinkedIn post of 23 May 2025

Method: Two statements by a chief executive about his own company, a month apart, reported by the press. No data.
Finding: The April email said Duolingo would 'gradually stop using contractors to do work AI can handle', that AI use would count in performance reviews, and that headcount would be added only 'if a team cannot automate more of their work'. After a public backlash, von Ahn wrote on 23 May: 'I do not see AI as replacing what our employees do (we are in fact continuing to hire)' and 'I see it as a tool to accelerate what we do, at the same or better level of quality'.
What it supports: That an AI-first declaration without a written boundary underneath it was reversed in a month under public pressure, with no change in the technology in between. It is the clearest public example of the slogan outrunning the strategy.
What it does not support: That anything operational changed. The retreat is a statement; the contractor line was not withdrawn; contractors are not employees. Both statements are one man's framing of his own company.

ai leadership · ai-first · duolingo · contractors · reversal · is ai first a strategy or a slogan what happened to companies that cut staff for ai

Operator account#

Microsoft (2025). Learn about retention for Copilot and AI apps

Microsoft Learn, Purview documentation, page dated 26 September 2025. Read at source 29 September 2026

Method: Vendor technical documentation describing where Microsoft 365 Copilot prompts and responses are stored and who can search them.
Finding: Data from generative AI messages is stored in a hidden folder in the mailbox of the user who runs the AI app. The folder is not designed to be directly accessible to users or administrators but holds data that compliance administrators can search with eDiscovery tools. Messages remain searchable until permanently deleted, and messages visible in an AI app are not an accurate reflection of whether they are retained or permanently deleted for compliance purposes.
What it supports: That Microsoft documents prompts and responses as stored in a searchable mailbox folder for compliance administrators, and that what a user sees is not a guide to what is kept.
What it does not support: What any organisation has configured or searched, or anything about other vendors' tools. The documentation is the vendor's own and is revised over time.

data retention · eDiscovery · workplace monitoring · can my employer see what i put into ai who owns the capability you built at work where your ai practice lives what it would take to leave

Institutional survey#

Niederhoffer, K., Rosen Kellerman, G., Lee, A., Liebscher, A., Rapuano, K. and Hancock, J. T. (2025). AI-Generated 'Workslop' Is Destroying Productivity

Harvard Business Review, 22 September 2025 (BetterUp Labs and Stanford Social Media Lab); figures from BetterUp's companion post. Read at source 24 October 2026

Method: Online survey of US full-time employees; the sample is given as 1,004 in one place and 1,150 in another.
Finding: 40 per cent had received workslop in the previous month; each incident took 1 hour 51 minutes to deal with, about $186 per employee per month; about half saw senders as less capable and reliable and 42 per cent as less trustworthy.
What it supports: That workers report receiving AI output that shifts work onto them, and that it changes how they judge colleagues.
What it does not support: That AI reduces productivity overall: there is no counterfactual and the value of AI output is not measured (Jason Collins's critique). The survey is by a firm selling coaching.

workslop · AI slop · productivity · trust · rework · what is ai slop

Institutional survey#

OECD (2025). How widespread is algorithmic management in workplaces?

OECD, December 2025. The report's web summary was read at source on 21 September 2026; the PDF would not open through the fetcher, and the country figures were read in The Next Web's report of 20 September 2026 and CNBC's of the same day

Method: An employer survey of 6,047 firms in France, Germany, Italy, Japan, Spain and the United States, fielded by Ipsos between June and August 2024 among mid-level managers at workplaces with 20 or more employees, per The Next Web. Algorithmic management is defined by the OECD as software used to automate or support managerial tasks in how work is instructed, monitored and evaluated.
Finding: Per The Next Web's account, the share of managers reporting at least one algorithmic management tool was 90 per cent in the United States, 81 in France, 78 in Germany and Spain, 76 in Italy and 40 in Japan; 67 per cent of US firms used a tool to sanction poor performance against 4 per cent across the four European countries and 1 per cent in Japan; 91 per cent of managers said workers were made aware, a figure the OECD questioned, and most said workers could not opt out. The OECD's own summary records managers' concerns about unclear accountability for algorithmic decisions, inability to follow the tools' logic and inadequate protection of workers' health.
What it supports: That by mid-2024 software to instruct, monitor or evaluate workers was in use at the large majority of surveyed workplaces in five of six countries, that its use to discipline workers was concentrated in the United States, and that the managers operating it reported not being able to follow its logic or say who was accountable for its decisions.
What it does not support: Anything about generative AI or agents, which the survey predates in most workplaces. Manager self-report, and the country figures here are as reported by a news outlet rather than read in the PDF. It does not measure outcomes for workers or how often a human reviews a sanction.

algorithmic management · monitoring · surveillance · accountability · discipline · OECD · what board oversight of ai looks like

Peer-reviewed#

Reif, J. A., Larrick, R. P. and Soll, J. B. (2025). Evidence of a social evaluation penalty for using AI

Proceedings of the National Academy of Sciences, 122(19), e2426766122, May 2025. Read at source 18 September 2026

Method: Four preregistered online experiments with over 4,400 participants in total, the abstract's own figure, run between March 2024 and February 2025. Participants rated described workers or candidates who used AI, who received equivalent help by other means, or who received none; one study placed participants as hiring managers evaluating candidates who reported different frequencies of AI use.
Finding: People described as using AI for a task were rated as lazier, less competent, less diligent, less independent and less self-assured than people described as receiving comparable non-AI help, with no difference on ambition or dominance. The penalty was largest among evaluators who used AI little themselves and smallest among those who used it weekly or daily; hiring managers who did not use AI preferred candidates who reported no use, while managers who used it regularly preferred candidates who used it daily. The penalty disappeared when the tool's usefulness for the specific task was made explicit. Participants also reported reluctance to disclose their own AI use to managers and colleagues.
What it supports: That a measurable social penalty for visible AI use existed in these samples, that it depends on the evaluator's own familiarity with the tools and on whether the tool's fit to the task is stated, and that the people who anticipate the penalty say they would conceal use.
What it does not support: Behaviour in real organisations: the targets are described, not observed, the participants are online panels, mostly in the United States, and the authors say the penalty is likely to shift as AI use becomes ordinary. It measures perception of users, not the quality of their work.

AI stigma · social evaluation · disclosure · impression management · AI adoption · what should we tell employees about ai and headcount should i tell my boss i used ai can i use ai to write a job application

Working paper#

Salesforce AI Research (2025). CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions

arXiv 2505.18878, 24 May 2025. Read at source 24 October 2026

Method: Benchmark of agents on business tasks in a simulated CRM environment.
Finding: Around 58 per cent single-turn success, falling to about 35 per cent in multi-turn settings, with near-zero inherent confidentiality awareness.
What it supports: That agents were unreliable on multi-step business work and on confidentiality.
What it does not support: Current performance. The authors' employer sells agents.

AI agents · benchmark · confidentiality · what is an agent boss

Peer-reviewed#

Schilke, O. and Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust

Organizational Behavior and Human Decision Processes, 188, article 104405, May 2025. Read at source 30 September 2026

Method: Thirteen experiments with a within-paper meta-analysis, testing how observers' trust in an actor changes when the actor's use of AI is disclosed, across communication, analytic and creative tasks.
Finding: Actors who disclose their AI use are trusted less than those who do not. In the final experiment a tax advisor exposed by a news leak was trusted less than one who disclosed voluntarily, though disclosure still reduced trust against no disclosure. The penalty was smaller among people with positive attitudes to technology and among those who rated the AI as accurate, and did not vanish for them.
What it supports: A consistent trust penalty for disclosed AI use in online experiments, and that being found out is worse than saying so.
What it does not support: What happens to a real firm's customers, contracts or retention. Participants rated described actors, and the paper does not test disclosure wording that states what the human checked.

AI disclosure · trust · transparency · should we tell customers we used ai

Operator account#

Sebastian Siemiatkowski, reported by Bloomberg and CX Dive (2025). Klarna changes its AI tune and again recruits humans for customer service

CX Dive, 9 May 2025, reporting a Bloomberg interview of 8 May 2025

Method: A chief executive's account of his own company, given in an interview, reported by trade press. No data beyond the company's own statements. In February 2024 Klarna had said its AI assistant handled two-thirds of customer service chats, 2.3 million in its first month, doing the work of 700 agents.
Finding: Siemiatkowski: 'As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality. Really investing in the quality of the human support is the way of the future for us.' And: 'I just think it's so critical that you are clear to your customer that there will be always a human if you want.' The company said it would recruit human agents again, keeping the assistant for routine enquiries.
What it supports: That the most-cited corporate case for replacing people with AI was reversed by the man who made it, on the grounds that the decision had been made on cost alone. It is the clearest public example of a leadership decision, not a technology failure, being the thing that went wrong.
What it does not support: How far Klarna reversed, or what the quality gap was; no figures were published for either. It is one company and one chief executive's framing of his own change of mind, and the earlier 700-agent claim was also his.

ai leadership · klarna · customer service · reversal · cost versus quality · human in the loop · ai leadership

Working paper#

Shao, Y., Zope, H., Jiang, Y., Pei, J., Nguyen, D., Brynjolfsson, E. and Yang, D. (2025). Future of work with AI agents: auditing automation and augmentation potential across the US workforce

arXiv preprint

Method: An audit of 844 tasks across 104 occupations, combining what 1,500 domain workers say they want automated with expert assessment of what current systems can do, and scoring each task on a five-level Human Agency Scale from AI operating alone to AI unable to function without continuous human involvement.
Finding: Worker preference and technical capability diverge widely. The audit identifies a set of tasks where systems are capable and workers do not want automation, and another where workers want it and capability is absent.
What it supports: That preference about automation can be measured at the level of the individual task, and that a scale of desired human involvement can be applied consistently across occupations.
What it does not support: What should be automated. The scale records what workers want and what experts judge possible, neither of which is a normative answer. It is also a preprint.

automation · augmentation · delegation · levels of automation · the use or keep test levels of automation

Vendor research#

Slack Workforce Lab (Salesforce) (2025). The New AI Advantage: Daily AI-Users Feel More Productive, Effective, and Satisfied at Work

Slack, June 2025

Method: Survey of 5,156 desk workers in Australia, France, Germany, Japan, the UK and the US, fielded by Qualtrics from 9 April to 1 May 2025, not targeting Slack or Salesforce employees or customers. Comparisons are to the November 2024 wave of the Workforce Index.
Finding: 60% of desk workers use AI at work and 42% use it at least weekly; daily use is 233% higher than in November 2024. Daily users are 64% more likely to report very good productivity, 58% more likely to report very good focus and 81% more likely to report very good job satisfaction than non-users. 96% of AI users have used it for tasks outside their expertise. 40% have used AI agent chatbots and 23% have assigned an agent to complete work.
What it supports: That among desk workers in six rich economies, self-reported AI use grew fast in 2024 to 2025 and frequent users rate their own work experience more highly.
What it does not support: Cross-sectional self-report cannot separate whether AI improves productivity or whether already productive, satisfied workers adopt AI. Salesforce sells AI agents, and the six-country desk-worker frame excludes frontline and non-desk work.

desk workers · daily ai use · job satisfaction · ai agents · self-reported productivity

Operator account#

Tobi Lütke, reported by TechCrunch (2025). Shopify CEO tells teams to consider using AI before growing headcount

TechCrunch, 7 April 2025, reporting an internal memo Lütke published on X the same day

Method: An internal memo from a chief executive to his company, made public by the author after it leaked. No data; a statement of policy.
Finding: 'Before asking for more headcount and resources, teams must demonstrate why they cannot get what they want done using AI.' The memo also asks teams: 'What would this area look like if autonomous AI agents were already part of the team?'
What it supports: That one large employer made AI the default option before a new hire, in writing, publicly and with a date. It is the clearest published example of an AI-first rule, and the memo frames it as a question about the work rather than a statement about people.
What it does not support: What the rule did. Shopify has not published its effect on hiring, output or quality. A published policy is evidence of intent, not of outcome.

ai leadership · ai-first · headcount · shopify · operating model · is ai first a strategy or a slogan what happened to companies that cut staff for ai

Peer-reviewed#

Wiles, E., Munyikwa, Z. and Horton, J. J. (2025). Algorithmic Writing Assistance on Jobseekers' Resumes Increases Hires

Management Science, 71(12), 10144-10164, December 2025. DOI 10.1287/mnsc.2024.04528, confirmed from Crossref on 19 October 2026. NBER Working Paper 30886 (January 2023, revised October 2023). The arXiv version (2301.08083, January 2023) lists the first author as Emma van Inwegen; the NBER page and the journal list Emma Wiles. Same title, same co-authors.

Method: Field experiment on a large online labour marketplace that the authors do not name. 480,948 new jobseekers who registered between 8 June and 14 July 2021 were randomly allocated, half to have their profile description checked by a commercial writing-assistance service. The service flags spelling and grammar and gives advice on clarity, tone and style. Analysis sample 194,700 jobseekers who were platform-approved with non-empty resumes. Outcomes followed to 14 August 2021. Figures read in the arXiv text; the journal text was not opened.
Finding: Treated jobseekers were 8 per cent more likely to be hired in their first month (95 per cent confidence interval 3 to 13 per cent), from a control base rate of about 3 per cent, and the treatment produced 7.8 per cent more contracts. Hourly wages of treated workers were 8.4 per cent higher, a result conditional on being hired. Employer ratings did not fall (8.84 against 8.835). The authors favour a clarity account, in which better writing lets employers read the same facts more accurately, over a signalling account.
What it supports: That, on this platform in mid-2021, offering new jobseekers a tool that corrected and advised on text they had written raised their hiring rate without lowering employer satisfaction.
What it does not support: That a tool which generates the content has the same effect: this one checked writing the jobseeker had produced, and generative models were not in use. One platform, one month of outcomes, profile descriptions rather than cover letters. The wage result is conditional on hire and may reflect who was hired. The authors say the study cannot measure crowd-out of other workers and that wide use may weaken writing as a signal.

resume · writing assistance · hiring · signalling · can i use ai to write a job application

Undisclosed frame#

Accenture Research (2026). Future of Work: Work That Matters

Accenture, autumn 2026, undated on the page; cites Pulse of Change surveys of July and October 2026. Read at source 24 October 2026

Method: Draws on Accenture Pulse of Change employee and CxO surveys and other analyses; sample sizes and methods are not published on the page.
Finding: Only 39 per cent of C-suite leaders say human accountability is clearly established as agents take on more work; only 45 per cent of organisations have monitoring capable of questioning AI outputs; one in five managers are concerned juniors are missing hands-on experience (source not given).
What it supports: What Accenture reports its surveys found.
What it does not support: Anything that can be checked: the method is not published.

accountability · monitoring · juniors · verification · what is the verification bottleneck missing rungs

Institutional survey#

Allen, J. S. (2026). Monitoring AI Adoption in the U.S. Economy

FEDS Notes, Board of Governors of the Federal Reserve System, 3 April 2026. DOI 10.17016/2380-7172.4032

Method: Comparison of three independent US adoption measures: the Census Business Trends and Outlook Survey (firm-level, around 20,000 responses per wave), the Real-Time Population Survey (individual-level, 5,000 to 6,000 responses) and the Atlanta Fed Survey of Business Uncertainty (senior leaders, 1,032 responses). Four-period moving averages used for all BTOS calculations.
Finding: About 18 per cent of firms had adopted AI as of year-end 2025 on the BTOS. Work-related generative AI adoption in the RPS stood at about 41 per cent of the workforce as of November 2025, with daily use at 12 per cent. The SBU gives an employment-weighted firm adoption rate of about 78 per cent and an LLM adoption rate of about 54 per cent. Allen attributes the variation mainly to differences in sampling distributions and units of analysis, with question framing, the materiality of reported usage, information asymmetries and social desirability bias also contributing, and states senior leaders may face pressure to report AI usage as an efficiency initiative. The Census Bureau broadened its question in November 2025; the do-not-know rate ran at 10 to 11 per cent.
What it supports: That headline AI adoption figures differing by sixty points can all be correct, because they measure different units, and that the choice of measure decides the answer before any analysis begins.
What it does not support: Any productivity or employment effect. The note explicitly does not estimate AI's contribution to output, GDP or productivity, and names those as open questions beyond its scope. It also makes no claim that adoption has plateaued; the only slowdown language is deceleration in the second quarter of 2025.

AI adoption · measurement · survey method · if everyone has ai where is the advantage how long before you know if an ai investment worked

Operator account#

Anthropic Claude Code team (Osmani, A., Lunt, M. W. and Wang, Q.) (2026). Solving code review at Anthropic

Anthropic, Claude Code blog, 15 October 2026. Read at source 23 October 2026

Method: An engineering team's description of its own review system: seven layers of review, most automated, with human approval tiered by the risk of the change. Figures are the company's own with no independent check.
Finding: Claude writes about 90 per cent of Claude Code's code. Human approval is tiered: none for low-risk changes, one person for medium, two for high. 'The change sets the tier, not the author.' 'An agent's approval never counts as a person's, either.' 'We didn't remove review. But we did move most of the reading to machines, optimized for human time and attention, and invested heavily in making it fast and easy to undo mistakes.' People keep deciding whether to build, risky code, architecture others copy, reading the evidence and owning the outcome. Merged changes more than doubled from June to September 2026 while the share reverted fell by over 40 per cent; changes needing five or more review rounds fell from 8.9 to 1.5 per cent.
What it supports: That one team has written down a tiered allocation of human review by risk of change, has placed accountability for outcomes with people while moving most reading to machines, and reports lower revert rates after doing so.
What it does not support: That the design works elsewhere, or what the machine layers miss: the figures are unverified, the codebase and team are unlike most, and no comparison against line-by-line human review is reported.

code review · AI agents · tiered approval · accountability · zero-human merges · can ai check ai what is the judgment class hand head hours heart

Vendor research#

Anthropic Economics team (Wilmers, N., Catipon, R., Massenkoff, M. and McCrory, P.) (2026). How AI is reshaping Anthropic's workforce

Anthropic, research report, 16 October 2026. Read at source 23 October 2026

Method: A study by Anthropic of its own workforce from internal usage data, hiring records and seventeen interviews, using graders built on its own models; the engineering analysis follows 192 engineers who adopted coding agents against 74 who did not. The authors say adoption is not random and describe the employment findings as descriptive.
Finding: By August 2026 agent hours outnumbered employee hours six to one; about 80 per cent of interactive session time was agents waiting for a human; sessions spent building tools for and supervising agents, which the report calls meta-work, rose from 10 to 22 per cent; standing written instructions for agents grew from a few hundred to 8,500 in a year; commits per engineer were about five times the level of six months earlier with no rise in reverted code. The junior share of technical hires fell from 27 to 9 per cent and juniors from 21 to 14 per cent of technical staff since early 2025. Hiring comments mentioning candidates' autonomy and judgement rose most while domain expertise fell as a share. 'This acceleration makes human attention the bottleneck for agent output.'
What it supports: That inside one AI company, by its own account, the volume of agent work has outrun the people available to direct and check it, and that junior hiring has fallen sharply over the same period.
What it does not support: Anything about other employers: the company is unlike most, the data are its own and unverified, the graders are its own models, and the report draws no causal conclusion about junior hiring.

AI agents · workforce · junior hiring · human attention · meta-work · what is judgement allocation capability debt

Vendor research#

Anthropic; Maxim Massenkoff, Eva Lyubich, Szymon Sacher, Zoe Hitzig, Shaoyi Zhang, Ryan Heller, Peter McCrory (2026). Anthropic Economic Index report: Cadences

Anthropic, June 2026

Method: Privacy-preserving analysis of Claude.ai, Claude Desktop, Claude Code and first-party API usage from 10 April to 10 June 2026, classified to O*NET tasks and output types. Paired with an Anthropic Economic Index survey of about 81,000 Claude users (December 2025) and roughly 9,700 respondents linked to their own usage data from mid-May to early June 2026.
Finding: Personal (non-work) conversations rise from around 35% of the total on weekdays to just under 50% at weekends. 93% of conversations produce an identifiable output; the most common are explanations (17%), documents and reports (15%) and guidance (11%). Over a third of surveyed users expect AI to be able to do most or nearly all of their work tasks next year, while 10% rate losing their own job as likely or very likely. Users with more automated usage patterns are more optimistic about AI's effect on pay and job security.
What it supports: That on one frontier model, work use follows the working week and most conversations end in a concrete output; and that heavy users hold both high capability expectations and low personal job risk expectations.
What it does not support: Claude users are self-selected and skew technical; usage patterns on one product are not labour market outcomes. The survey is of the vendor's own customers about the vendor's product, and Anthropic has a commercial interest in the framing.

ai usage data · task classification · augmentation · worker expectations · claude

Institutional survey#

Atlassian (2026). State of Developer Experience Report 2026

Atlassian, 2026, undated on the page; more than 1,500 practitioners surveyed. Read at source 23 October 2026

Method: A vendor's survey of software developers and leaders; sampling, fieldwork dates and regions are not stated on the page read.
Finding: 87 per cent say AI helps them ship faster and 28 per cent report less friction; 49 per cent spend four or more hours a week reviewing AI output and correcting prompts; 75 per cent say they have more responsibility for validating AI output; 17 per cent of developers and 23 per cent of leaders say their organisation has a clearly communicated, actively implemented AI strategy; 64 per cent of leaders say greater AI usage is correlated with better outcomes while 41 per cent of developers say they use AI more than necessary because they feel expected to. 'Engineering is becoming more about judgment than code.'
What it supports: That among the developers surveyed, reviewing AI output is now a large and acknowledged part of the week, and that pressure to use AI is felt separately from its usefulness.
What it does not support: Anything beyond the sample: the sampling is unpublished and the vendor sells the tools whose adoption it reports.

developers · code review · AI output · productivity · survey · what is the judgment class

Institutional survey#

BCG (2026). The Formula for Agentic AI Value (Applied AI Index 2026)

BCG, 30 September 2026. Read at source 24 October 2026

Method: Survey of more than 1,300 C-suite executives and senior leaders across more than 20 sectors; countries and fieldwork dates not stated.
Finding: 42 per cent of companies expect agents to act autonomously by 2030, deciding without human approval; only 5 per cent have the full set of critical controls today; companies with all six controls generate three times as much agentic AI value; 7.5 per cent are 'future-built'.
What it supports: That executives expect agent autonomy well ahead of the controls they report having.
What it does not support: Actual autonomy or value; figures are self-reported by a firm that sells AI transformation.

AI agents · controls · autonomy · agent boss · what is an agent boss how should a board govern ai agents

Argued perspective#

Barr, M. S., Board of Governors of the Federal Reserve System (2026). Economic Conditions and Monetary Policy

Speech at the Detroit Economic Club, 29 September 2026. Read at source 2 October 2026

Method: A central bank governor's speech with one section on AI and the labour market, reporting no new data.
Finding: 'While there are some indications that AI may already be a factor limiting new job opportunities for entry-level workers in sectors heavily exposed to AI, across the economy there is little evidence of significant displacement so far.' 'If labor market changes happen quickly, it will be hard for workers to adjust and dislocations might be large, whereas a more gradual adoption might permit more orderly adjustments.' Barr says 'now is the time for society to begin to consider how to address these potential disruptions, while AI adoption is in its relatively early stages'.
What it supports: How a Federal Reserve governor read the US evidence at the end of September 2026.
What it does not support: Any measurement: it is a judgement on others' data, about the United States, in a speech mainly on monetary policy.

labour market · displacement · central banks · AI and employment · is ai already causing job losses

Argued perspective#

Benedict, G., Reese, H., Richard, M. and Meskarian, M. (PwC) (2026). Three governance shifts to build trust in agentic AI autonomy

PwC, 11 September 2026. Read at source 24 October 2026

Method: Argued perspective.
Finding: Human-in-the-loop supervision 'is no longer sustainable where agents increase in number, pace, and complexity'; every agent needs an accountable business owner; the board should oversee whether governance for agentic AI is effective.
What it supports: That a Big Four firm assigns agents named owners and boards a governance role.
What it does not support: That engineered supervision works better; no evidence is reported.

AI agents · governance · boards · accountability · what is an agent boss how should a board govern ai agents

Working paper#

Bonney, K., Breaux, C., Dinlersoz, E., Foster, L., Haltiwanger, J. and Pande, A. (2026). The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks

US Census Bureau, Center for Economic Studies Working Paper CES-26-25, April 2026

Method: Nationally representative data from the 2026 AI supplement to the US Census Bureau's Business Trends and Outlook Survey, analysed at three layers: overall firm use, deployment across business functions, and worker-task use. Reference period November 2025 to January 2026.
Finding: 18 per cent of firms used AI in a business function, rising to 32 per cent employment-weighted, with adoption expected to reach 22 per cent within six months. Use rates reach 50 to 60 per cent, and 60 to 70 per cent employment-weighted, for very large firms in Information, Professional Services and Finance. Among adopters, 57 per cent integrate AI in three or fewer business functions, most commonly Sales and Marketing (52 per cent), Strategy and Business Development (45 per cent) and IT (41 per cent). Workers use AI in work-related tasks in 23 per cent of firms, 41 per cent employment-weighted, and 65 per cent of firms limit use to three or fewer tasks. Most users, 66 per cent, rely on AI solely to augment tasks, and AI-related employment decreases occur in only 2 per cent of firms. Regression shows a positive correlation between firm commercial performance and the breadth of AI integration, holding across functional deployment, task-level use and operational investment; functional breadth and operational investment are positively associated with employment decreases, while worker-task integration shows no significant link to headcount reduction once the other two are accounted for.
What it supports: That AI diffusion is highly uneven by firm size and sector, and shallow even among adopters, which contradicts the premise that competitors hold equivalent capability.
What it does not support: Causation in either direction between adoption and performance: this is a cross-section of firms, and better-run firms may simply adopt more. Also not a stable time series. The Census Bureau broadened the underlying question in November 2025 from use in producing goods or services to use in any business function, which moved the level. A working paper, not peer reviewed, and US-only.

AI adoption · diffusion · firm size · augmentation · if everyone has ai where is the advantage how long before you know if an ai investment worked which tasks do workers not want automated what happens to work that moves information

Working paper#

Boyd-Swan, C. and Reynolds, C. L. (2026). Early Evidence on Employer Responses to College Student Use of Generative AI

Annenberg Institute EdWorkingPaper 26-1604, October 2026. Abstract and paper read at source 9 October 2026

Method: Survey experiment on about 1,750 US professionals with hiring responsibilities, recruited through Prolific between 21 May and 8 June 2026 and randomised to a control or to one of two framings of student AI use (assisting students, or substituting for their effort and ability) before reviewing two constructed CVs.
Finding: 'Results show that the framing increased respondents' concerns about evaluating job applicants, with a larger effect in the second treatment group.' The substitution framing 'clearly decreased' interview interest and willingness to negotiate starting salaries. Of the information that might allay concern, 'only in-person tests were selected by both treatment groups.'
What it supports: That being reminded students use AI makes people who hire less sure what an application shows, and that an in-person test is the evidence they reach for.
What it does not support: Real hiring: 'Our outcomes are hypothetical evaluations rather than real hiring decisions', the CVs were written by the researchers, and the panel is online and US only. Not peer-reviewed.

hiring · graduates · credentials · proof of work · does ai make juniors better or only look better

Working paper#

Brynjolfsson, E., Chandar, B. and Chen, R. (2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence

Stanford Digital Economy Lab, updated August 2026

Method: ADP payroll microdata covering millions of US workers, comparing employment by age and by occupational AI exposure since the release of ChatGPT.
Finding: No widespread economy-wide displacement. But employment among 22 to 25 year olds in highly AI-exposed occupations sits about 19 percent below where it would be had it tracked similarly aged workers in less-exposed occupations. The underlying levels matter and are easily lost: employment of that age group in the two most exposed quintiles fell about 11 percent between November 2022 and June 2026, while the same age group in the three least exposed quintiles grew about 10 percent. The 19 is the distance between those two, not a fall of 19. The divergence runs through reduced hiring rather than increased separations, and declines concentrate in occupations where AI substitutes for human tasks; where it complements, employment is flat or rising, especially for experienced workers.
What it supports: That the entry-level effect is real, measurable in payroll data rather than inferred, and specific to substitution rather than to AI exposure as such.
What it does not support: Economy-wide job destruction, which the authors explicitly rule out on current evidence. Nor does it establish causation: youth hiring is sensitive to interest rates, cohort size and hiring freezes, and the design is observational. Two further cautions, both from the paper itself. First, the phrase 'below trend' is wrong and is how this figure is usually repeated: the comparison is with same-aged workers in less-exposed occupations, not with a historical trend line and not with older workers. Second, the figures in circulation are not a worsening time series. 13 and 16 percent come from an earlier regression estimator, 15 and 19 from the descriptive kept-pace measure the authors now prefer, so quoting them in sequence as though the effect grew misrepresents a change of method as a change in the world. Note also that the ADP figures widely cited alongside this paper are the same payroll data reported differently, not independent corroboration.

entry-level employment · youth employment · substitution versus complementarity · hiring · will ai replace entry level jobs missing rungs the best writing on ai what is the ai employment gap is ai already causing job losses

Argued perspective#

Catalini, C., Hui, X. and Wu, J. (2026). Some Simple Economics of AGI

arXiv 2602.20946, v1 24 February 2026. Read at source 24 October 2026

Method: An economic framework contrasting the cost to automate with the cost to verify.
Finding: Argues that 'the binding constraint on growth is no longer intelligence but human verification bandwidth', with a falling cost to automate against a biologically bottlenecked cost to verify.
What it supports: That the verification bottleneck can be stated as an economic constraint.
What it does not support: Anything empirical: it is a framework, not a measurement.

verification bottleneck · economics · verification · what is the verification bottleneck

Vendor research#

Dayforce, with Hanover Research (2026). 17th Annual Pulse of Talent: more than one third of organizations are highly fragile

Dayforce, press release and report, 26 October 2026. Read at source 30 October 2026

Method: Online survey of 6,909 workers, managers and executives aged 18 and over at organisations with more than 100 employees in Australia, Canada, Germany, New Zealand, the United Kingdom and the United States, fielded 27 July to 14 August 2026 with Hanover Research. The company sells workforce software and the release introduces a Workforce Fragility Index of its own construction.
Finding: 68 per cent say their organisation is losing important skills or expertise without fully realising it; 73 per cent say lost workforce capabilities are harder to rebuild than organisations expect; 51 per cent say AI is reducing early-career opportunities; 49 per cent say tasks that built work experience are now being handled by AI; 65 per cent say people are losing the ability to perform some tasks themselves; 89 per cent are concerned about AI's effect on workforce development; 54 per cent say important knowledge is becoming concentrated among fewer people; 60 per cent say understaffing has become normalised; 27 per cent of workers and 56 per cent of executives say employees have enough time at work to develop new skills. The release names this 'a growing experience debt, a future capability cost that accumulates when people lose opportunities to learn by doing.' 35 per cent of organisations are classed as high fragility on the company's eleven-measure index.
What it supports: That a large multi-country panel of workers and executives believes their organisations are losing skills they will need, that AI is removing the tasks through which experience was built, and that a vendor has given that belief the name experience debt with a definition.
What it does not support: That any skill has been lost: the survey measures belief, the panel is self-selected, the index is the vendor's own, and the company has a commercial interest in the finding. It reports no measure of capability before and after AI, and no causal link.

experience debt · workforce fragility · early-career opportunities · skills loss · learning by doing · what is experience debt capability debt

Argued perspective#

Deloitte Insights (2026). AI agents in the human workplace

Deloitte Insights, 10 September 2026. Read at source 24 October 2026

Method: Four scenarios built from surveys, executive interviews and horizon scanning; explicitly not forecasts.
Finding: Nearly 70 per cent of respondents to a Deloitte survey named an inability to trust and govern agents as a major barrier; warns oversight can become a rubber stamp when approval speed is the KPI; one scenario names 'preserving the human judgment pipeline before it atrophies'.
What it supports: That a Big Four firm sees accountability and the judgement pipeline as the risks most likely to be overlooked.
What it does not support: Which scenario will occur; the survey behind the 70 per cent is not described on the page.

AI agents · accountability · judgement pipeline · juniors · what is an agent boss missing rungs how should a board govern ai agents

Institutional survey#

Deloitte UK, fieldwork by Ipsos UK (2026). British workers spend one billion pounds of their own money on gen AI for work (UK GenAI Workforce Survey)

Deloitte UK press release, 16 September 2026. Read at source 18 September 2026

Method: Online survey of 25,000 UK workers aged 18 to 70, employed and self-employed, fielded by Ipsos UK between 7 May and 10 June 2026, quota sampled and weighted for gender, age, working status, sector, region, social grade, education and ethnicity. Self-report throughout. The full report and questionnaire were not published with the release.
Finding: 63 per cent of UK working adults say they use generative AI for work and 24 per cent use it every day. Of the users, 31 per cent use it without their employer's knowledge, 23 per cent perceive a stigma around using it at work, and 64 per cent of weekly users worry that their manager will think AI can do their job. About half say they have had no formal training in using it safely and effectively, and 65 per cent report a lack of convincing leadership guidance. 17 per cent pay for at least one tool themselves, which Deloitte scales to £958 million a year. Users report saving about 70 minutes a week. Hayley McKelvey, Deloitte UK's chief AI officer, is quoted: 'Workers who feel stigma around adopting GenAI tools are more likely to conceal use.'
What it supports: That in a large weighted UK sample, concealed use, perceived stigma and fear of being judged replaceable are reported together and at scale, and that most users say they have had neither training nor convincing guidance from leadership.
What it does not support: Any outcome: the time saved, the quality of the work, and the stigma are all self-reported. The 31 per cent measures use the employer does not know about, which is not the same as use a policy forbids. Deloitte advises organisations on the adoption it is measuring, and the questionnaire is not public, so the wording behind each figure cannot be checked.

shadow AI · generative AI · AI adoption · workforce · stigma · AI training · what should we tell employees about ai and headcount what to do when people work around the ai policy should i tell my boss i used ai can my employer see what i put into ai how do we train people to use ai what are the risks of using ai at work my company banned ai what should i do

Argued perspective#

Department for Business and Trade (UK) (2026). Make Work Pay: consultation on workplace monitoring technologies

UK government consultation, issued 8 July 2026, closed 30 September 2026. Read at source 2 October 2026

Method: A government consultation document of 68 pages setting out a definition, eight principles and three options (a statutory code of practice, a legislative requirement to consult and negotiate, and non-statutory guidance) for England, Wales and Scotland.
Finding: Defines workplace monitoring technologies as 'digital tools used by employers to collect, track, analyse or make decisions based on information about workers and their activities'. Cites 'a 2025 survey of UK managers finding that 1 in 3 organisations actively monitor an employee's digital activity, up from 1 in 5 employers in ICO research in 2023'. Under the first option a tribunal would have discretion to adjust compensation by up to 25 per cent where an employer unreasonably failed to follow the code. Under the second, employers' plans to adopt the technology 'would be subject to consultation and negotiation, with a view to agreement of trade unions or elected staff representatives where there is no trade union', and the process 'would not require agreement to be reached in all cases'. The document says the options do 'not indicate a settled government preference'.
What it supports: What the UK government was considering in 2026 and the principles it regards as good practice, including worker engagement and human oversight with a route to challenge.
What it does not support: Any decision: the consultation closed on 30 September 2026 and no response has been published. The one-in-three figure comes from a survey the document does not name. It covers monitoring technology and not AI in general.

workplace monitoring · worker voice · human oversight · consultation · algorithmic management · can an employer use ai to discipline or dismiss someone how should employees be involved in decisions about ai

Institutional survey#

Dixon, J.C. (2026). I surveyed workers to see if AI had caused job losses and was surprised by the findings

The Conversation, 27 August 2026. YouGov survey commissioned by the author, College of the Holy Cross. Described by the author as a study in progress and not peer-reviewed

Method: 1,250 employed US workers surveyed online by YouGov between 30 July and 4 August 2026, 25 questions, weighted to the employed US population on age, gender, race and education. Opt-in panel. The author discloses an AI consulting business.
Finding: About 3 per cent said they had lost a job to AI since 2023, against roughly 6 per cent who said they held a job that did not exist before AI and about 9 per cent reporting an AI-related promotion. Around 95 per cent said no, with 3 to 4 per cent unsure.
What it supports: That self-attributed AI job loss is rare among people currently in work, and that reported AI-created gain in the same sample runs at twice the reported loss. Useful as a counterweight to displacement figures drawn from payroll data, which cannot ask anyone why.
What it does not support: Displacement in the workforce, and the reason is structural rather than a matter of sampling error. Every respondent was employed when surveyed, in Dixon's own words 'none were jobless, whether due to AI or another reason'. Anyone displaced by AI and still out of work is excluded from the numerator and the denominator alike, so the 3 per cent counts only those who lost a job and have since found another. It is a floor among survivors, not an estimate of displacement. It is also not a measure of fear: the question asked what happened, not what respondents expect.

entry-level employment · displacement · self-report · will ai replace entry level jobs what is the ai employment gap

Institutional survey#

EY (2026). EY survey finds that autonomous AI implementation outpaces oversight, yielding an AI governance gap

EY press release, 15 September 2026. Read at source 24 October 2026

Method: Survey of 202 senior decision-makers at US companies with revenue of $1 billion or more.
Finding: 91 per cent use agentic AI; of those, 49 per cent have not updated governance for it, 85 per cent have systems acting without real-time human involvement and 26 per cent cannot detect unauthorised agents; 47 per cent have bypassed their own AI governance for urgent deployments; 36 per cent report an AI incident with materially negative impact.
What it supports: That among the large firms surveyed, agent deployment ran ahead of governance.
What it does not support: Anything beyond large US firms; self-reported by a firm that sells governance services.

AI agents · governance · boards · incidents · how should a board govern ai agents

Working paper#

Fairlie, R. W. and Wu, J. (2026). The Early Impacts of AI on Employment among Recent College Graduates

NBER Working Paper 35796, September 2026. Abstract read at source 2 October 2026

Method: US Current Population Survey microdata for June, July and August 2026; difference-in-differences and event-study models comparing recent graduates with older graduates and with young workers without a degree; an expanded unemployment measure adding those who report wanting a job; interactions with occupational AI exposure and remote work availability.
Finding: For recent graduates, 'unemployment rates did not spike in summer 2026 relative to summer months in previous years and did not rise in a significant way relative to older college graduates or young workers without a college degree'. The expanded measure 'adds nearly two percentage points' to graduate unemployment and still shows no significant increase. The authors find 'some evidence of a positive relationship with remote work availability'.
What it supports: That three months of US survey data show no jump in unemployment among recent graduates in the first summer the authors think an effect could be detected.
What it does not support: Anything about hiring that did not happen, about employment as distinct from unemployment, about later months or about the UK. The authors take 'an agnostic approach to defining treatment timing'. Not peer-reviewed; abstract only.

entry-level jobs · graduates · unemployment · AI and employment · is ai already causing job losses

Institutional survey#

Federal Reserve Bank of New York (2026). The Labor Market for Recent College Graduates

New York Fed, data through 2026 Q2

Method: Current Population Survey and ACS tracking of unemployment and underemployment by major since 1990.
Finding: As at the second quarter of 2026, unemployment among recent graduates runs around 5.6 per cent and underemployment around 42 per cent, the highest since 2020.
What it supports: That the graduate labour market has tightened measurably, and that underemployment is the larger number.
What it does not support: Attribution to AI. The series is descriptive and the Fed states it is not a forecast.

· what should i tell my children to study

Argued perspective#

Fernandez, J. (2026). Your Talent Strategy Has to Keep Up with Your AI Transformation

Harvard Business Review, 29 June 2026

Method: Practitioner article in a management magazine. The article body is paywalled and was not read; the entry rests on the HBR page summary, read 12 October 2026, and on a secondary summary by Brian Heger dated 7 July 2026. The HBR page names Jenny Fernandez as author; whether there are co-authors is not confirmed.
Finding: The HBR summary proposes that leaders keep their automation strategies and build the talent infrastructure those strategies need by redesigning entry-level roles as capability-building cohorts, building a distributed apprenticeship pipeline, and auditing and repaying the organisation's capability debt.
What it supports: That the phrase capability debt is in use at organisation level in a major management journal, applied to a shrinking early-career pipeline, independently of this research.
What it does not support: Anything measured, and nothing about what the article body argues or evidences. The sense of the term in the body is not verified. It does not establish who used the phrase first, and no priority is claimed for or against anyone.

capability debt · early-career pipeline · apprenticeship · capability debt

Institutional survey#

Frick, J., Doddapaneni, P. and Ganti, A. (Bain & Company) (2026). The Half-Finished Redesign: How AI Reshapes Software Organizations (Technology Report 2026)

Bain & Company, 29 September 2026. Read at source 24 October 2026

Method: Bain AI Org Assessment Survey 2026 of 155 software companies; design not described further.
Finding: Junior engineers fell from about one third of teams to less than a fifth while senior ranks grew; time spent directing AI agents rose from 4 per cent two years ago to 23 per cent, projected 34 per cent in two years.
What it supports: That in the software firms surveyed, junior shares fell and agent direction became a large part of the work.
What it does not support: Anything outside software, or causation.

juniors · missing rungs · AI agents · software · what is an agent boss missing rungs how should a board govern ai agents

Institutional survey#

Gallup; Andy Kemp (2026). Organizational AI Adoption Jumps Six Points

Gallup, July 2026

Method: Gallup Panel survey of 22,573 employed US adults aged 18 and over, working full or part time, fielded 6 to 20 May 2026 with a margin of error of plus or minus 0.9 percentage points at 95% confidence. Quarterly series since Q2 2023; the previous wave (4 to 19 February 2026) had 23,717 respondents.
Finding: 52% of US workers use AI in their role at least a few times a year, 30% a few times a week or more and 15% daily; in Q2 2023 the any-use figure was 21%. 47% say their organisation has integrated AI tools, up from 41% the previous quarter. The most common uses are writing and editing (51%), search or research (49%) and general assistance or problem-solving (39%). Among employees using AI for one or two purposes, 45% report a positive effect on productivity; among those using it for seven or more purposes, 90% do. In Q1 2026, 8% of employees in AI-adopting organisations strongly agreed that AI had transformed how work gets done, and 18% of all employees thought their job was likely to be eliminated within five years.
What it supports: That US workplace AI use has more than doubled in three years on a probability-based panel, and that reported benefit scales with breadth of use.
What it does not support: Self-reported productivity is not measured output, and breadth of use may follow from being in roles where AI helps most. US only. It does not measure quality of work or skill change.

us workforce · ai adoption trend · frequency of use · breadth of use · probability panel

Working paper#

Garicano, L. (2026). The Vanishing Advantage of Specialization: AI, Knowledge Utilization, and the Boundary of the Firm

Brookings Papers on Economic Activity conference draft, summary posted 23 September 2026, presented 25 September 2026. Brookings summary read at source 28 September 2026; the paper itself not read

Method: A knowledge-economics argument that AI lowers the fixed cost of acquiring specialist knowledge, tested against US employment in six occupations covering 5.9 million workers, comparing 2022 to 2025 with the pre-AI trend, with supporting evidence from ChatGPT usage data and corporate legal departments.
Finding: In four of the six occupations, lawyers, accountants, software developers and management analysts, employment in specialised outside firms fell faster between 2022 and 2025 than the pre-AI trend. 'Artificial intelligence reduces the fixed cost and hence benefits those who use this knowledge less often.' Garicano names a 'broken ladder': the less complex jobs in outside firms that trained graduates go first.
What it supports: That the work of outside specialist firms is moving in-house in the occupations most exposed to language models, on US employment data, and that the entry-level rungs in those firms are the first to go.
What it does not support: Causation: the comparison is against a trend rather than a control, and interest rates, offshoring and other causes are not ruled out in the summary read; nothing outside the United States; a conference draft, not peer-reviewed.

AI and the professions · entry-level jobs · boundary of the firm · specialisation · broken ladder · will human judgement protect lawyers from ai

Argued perspective#

Hancock, B. and Seiler, C. (2026). Building expertise in the age of AI: Who trains the next generation?

McKinsey Quarterly, 14 July 2026

Method: Consultancy article by a partner and an associate partner at McKinsey, edited by Barbara Tierney. It combines third-party survey and labour-market figures with a proposed training pattern and a described medical-student pilot. Read at source 12 October 2026. No outcome data for the proposed pattern were found in the text read.
Finding: Proposes the answer-key model: the employee attempts first, the AI grades the attempt, and the employee sits down with a manager who discusses the differences. Contrasts attempt-then-check learning with borrowed competence, a phrase used once and not defined.
What it supports: That a major consultancy is recommending attempt-first sequencing for training with AI, and what the pattern is.
What it does not support: That the pattern works. No outcome measurement for the model itself was found in the text read. The survey figures quoted around it are context. The page title in the browser tag differs from the article headline (Rethinking talent development in the age of AI).

answer-key model · borrowed competence · entry-level work · what is the answer key model

Institutional survey#

High Fliers Research (2026). The Graduate Market in 2026

High Fliers Research, January 2026

Method: Annual survey of graduate recruitment at the UK's 100 leading graduate employers. An employer survey of a selected group, not an official statistic.
Finding: Graduate recruitment fell 5.1 per cent in 2025, after a 14.6 per cent drop in 2024 and a 6.4 per cent decrease in 2023, with a further 0.5 per cent decrease forecast for 2026. Graduate recruitment at these employers has fallen 24.5 per cent since 2022, the lowest level since 2012. Employers received on average 23 per cent more applications in the first half of the 2025-2026 season, with applications roughly doubling since 2023.
What it supports: That the graduate entry route at the UK's largest recruiters has contracted by roughly a quarter in three years while competition for what remains has roughly doubled.
What it does not support: Attribution to AI, which the survey does not establish, and it covers 100 leading employers rather than the whole labour market.

graduate recruitment · early careers · United Kingdom · do apprenticeships still work will ai replace entry level jobs

Institutional survey#

IBM Institute for Business Value, with Oxford Economics (2026). 2026 CHRO Study: Designing the Thinking Organization

IBM Institute for Business Value, global C-suite study series, published 21 September 2026 with a press release datelined Armonk, New York. The report's own page and the press release were read at source on 23 September 2026; the figures were checked against Campus Technology and Virtualization Review (21 September) and Fair Play Talks (23 September). A commercial publication by a vendor of AI software, not peer reviewed; response rates and sampling frames are not published

Method: Two surveys fielded April to June 2026: 1,500 chief human resources officers and senior workforce strategy executives across 21 geographies and 23 industries, in organisations with annual revenue or budget from 250 million to 123.7 billion dollars and from 200 to more than 880,000 employees; and 8,800 full-time employees across 28 countries. Respondents were segmented by the study's own measure of strategic maturity, and the outcome figures are respondents' estimates of their own organisations.
Finding: 71 per cent of executives rank the ability to supervise, validate or override AI output among the most essential skills, against 38 per cent of employees; 57 per cent of executives and 49 per cent of employees put critical thinking and problem framing first; 29 per cent of employees rank judgement among the most important skills for their near future. 26 per cent of organisations clearly define work across human-led, AI-assisted and AI-executed activities, while 52 per cent of employees say their tasks changed in the past year because of AI. 36 per cent of executives say unclear accountability complicates AI deployment; 43 per cent of employees say the blame falls on them when AI fails; 80 per cent of CHROs believe AI creates invisible work and 42 per cent of employees say AI has increased unrecognised work. 60 per cent of employees worry AI is eroding their skills, three quarters of those say it already has, and 46 per cent of executives name skill erosion as a top concern, which the study distinguishes from the skills gap that the 80 per cent of organisations with a reskilling roadmap address. Where judgement is built into the workflow 62 per cent of CHROs report rising employee confidence in AI-enabled decisions; where it is not, 57 per cent report it falling. 46 per cent of organisations do not involve the CHRO when AI strategy is defined. Organisations with clearly defined workflows report 18 per cent lower risk and 20 per cent higher quality.
What it supports: That in a large, multi-country sample the people who set the expectation that AI output will be supervised and overridden and the people expected to do it disagree by 33 points on whether that is part of the job, that most organisations had not written down which work is human-led, and that a large minority of employees believe failures are blamed on them. It is the first large-sample count of the moral crumple zone as employees experience it, and the largest self-report of skill erosion to date.
What it does not support: Any capability. A gap in perceived priority is not a gap in ability to supervise, and a worry about erosion is not a measured loss; nothing in the study tests what anyone can do unaided. All figures are self-reported to a vendor whose products are the subject, the outcome figures are respondents' estimates of their own organisations rather than measured results, no override rate is reported, and the study did not verify any case of blame. 'Thinking Organization' is IBM's own label for its high-maturity segment and carries no independent standing.

human oversight · override · skill erosion · deskilling · moral crumple zone · invisible work · accountability · CHRO · work redesign · do employees know they are expected to override ai what is a moral crumple zone what is deskilling what are ai red lines

Compiled review#

Imas, A. and Schaal, J. (2026). Has AI impacted the labor market yet? Evaluating the state of the evidence

Ghosts of Electricity (Substack), 29 September 2026. Read at source 2 October 2026

Method: A review by two economists of the published studies of AI and employment to September 2026, selecting its own literature.
Finding: Concludes that 'the impact of AI on the overall labor market has been consistently muted'. Reports the payroll dashboard figure of employment among 22 to 25 year olds in high-exposure roles 19 per cent below less-exposed peers by June 2026; Lambert and Schindler's result that when remote work is entered jointly 'the coefficient for AI exposure drops to zero or even flips positive'; Humlum and Vestergaard's 'precise null effects on earnings, recorded hours, and wages' for 25,000 Danish workers; and null results for Finnish and Norwegian youth.
What it supports: That the published studies disagree about young workers in exposed jobs and agree that no aggregate effect has been measured.
What it does not support: Which side is right. An unreviewed essay that selects its own studies, and it gives two different figures for the July 2025 vintage of the payroll study without reconciling them.

labour market · entry-level jobs · evidence review · AI and employment · is ai already causing job losses

Regulator guidance#

Information Commissioner's Office (UK) (2026). Data protection and monitoring workers

ICO guidance, employment practices and data protection. Read at source 29 September 2026; page metadata dated 16 June 2026

Method: Statutory regulator's published guidance on how UK data protection law applies to monitoring workers, using must, should and could to separate legal requirements from good practice.
Finding: Employers must make workers aware of how and what personal information is collected during monitoring, must tell them in an accessible and easy to understand way, and apart from very exceptional circumstances where covert monitoring is justified must inform workers about any monitoring. They must identify a lawful basis, and must carry out a data protection impact assessment before processing likely to cause high risk to workers' interests. The guidance lists tracking of internet activity and keystrokes among monitoring technologies.
What it supports: The UK regulator's stated view of what an employer must do before and while monitoring workers.
What it does not support: Anything about whether a given employer complies, or the position outside the UK. It is guidance on the law and not itself the law, and it does not address AI chat logs by name.

workplace monitoring · transparency · data protection · can my employer see what i put into ai can an employer use ai to discipline or dismiss someone

Working paper#

Jadhav, R. and Danve, J. (2026). The AI Skills Shift: Mapping Skill Obsolescence, Emergence, and Transition Pathways in the LLM Era

arXiv 2604.06906, 8 April 2026

Method: Benchmarking of four frontier models (LLaMA 3.3 70B, Mistral Large, Qwen 2.5 72B, Gemini 2.5 Flash) across 263 text-based tasks covering all 35 skills in the US Department of Labor O*NET taxonomy, 1,052 model calls. Cross-referenced against the Anthropic Economic Index. Measures models, not people. Preprint.
Finding: Introduces a Skill Automation Feasibility Index. Mathematics scores 73.2 and programming 71.8 for automation feasibility; active listening scores 42.2 and reading comprehension 45.5. All four models converge to similar skill profiles within a 3.6-point spread. Reports that 78.7 per cent of observed AI interactions are augmentation rather than automation, and a "capability-demand inversion" in which the skills most demanded in AI-exposed jobs are those the models perform least well at.
What it supports: That measured model capability and stated employer demand point in different directions, which is a useful counterweight to displacement forecasts built on exposure scores alone.
What it does not support: What happens to human capability. The authors state their index "measures LLM performance on text-based representations of skills, not full occupational execution". No humans were studied.

skills taxonomy · automation feasibility · O*NET · augmentation · cognitive debt and capability debt

Institutional survey#

Jones, J. M., Gallup (2026). AI Benefits at Work Unevenly Distributed

Gallup, 6 October 2026: first results from the second year of the American Job Quality Study, with Jobs for the Future and The Families & Workers Fund. Read at source 9 October 2026

Method: Web and mail surveys, 26 January to 24 March 2026, of a random sample of 17,874 US working adults aged 18 to 75, drawn from the Gallup Panel and an address-based sample. The article's findings rest on 15,482 employees, of whom 7,845 have used AI in their job.
Finding: 28 per cent of employees use AI at least weekly: 37 per cent of managers against 25 per cent of individual contributors. Among employees who have used AI, 63 per cent say they do their job tasks faster; 46 per cent say they produce higher quality work and 43 per cent say they do not; 31 per cent say their employer has asked them to take on more responsibilities. 52 per cent of all employees report less influence on the adoption of new technology than they would like.
What it supports: That in a large probability sample speed is the commonly reported benefit of AI at work, that views on quality divide almost evenly, and that regular use is higher among managers.
What it does not support: That AI caused any of it: the answers are self-reports, and nothing here measures the quality of the work as others would judge it.

AI at work · productivity · work quality · managers · survey · what is the judgment class

Institutional survey#

KPMG US (2026). Q3 2026 AI Pulse

KPMG press release, 24 September 2026; fielded 24 July to 25 August 2026. Read at source 24 October 2026

Method: Survey of 314 US C-suite and business leaders at organisations with revenue of $1 billion or more.
Finding: Confidence in governance rose from 57 to 73 per cent in a quarter while the share with controls built into agents fell from 43 to 30 per cent over two quarters; 49 per cent had defined high-risk uses where agents cannot decide alone; 62 per cent are building or deploying agents.
What it supports: That confidence and controls moved in opposite directions among the leaders surveyed.
What it does not support: Anything beyond large US firms, or the quality of the controls.

AI agents · governance · controls · verification · what is the verification bottleneck how should a board govern ai agents

Working paper#

Kochan, T., Kelly, E. and Minster, A. (2026). Negotiating to Work in Partnership on AI: The 2025 Agreement Between Kaiser Permanente and the Alliance of Health Care Unions

MIT Sloan School of Management report, 2026, as summarised by Betsy Vereckey, MIT Sloan Ideas Made to Matter, 28 September 2026. The summary was read at source on 2 October 2026; the report itself was not read

Method: A case study of one negotiation by researchers at MIT Sloan and San Francisco State University.
Finding: The 2025 agreement between Kaiser Permanente and an alliance of unions representing 62,000 healthcare workers incorporates 'worker input into every stage of decision-making around how AI will be developed and used'. An AI task force 'comprises 10 senior leadership members (five from labor and five from KP) who have direct decision-making authority over technology investments'; more than 3,500 unit-based teams look for uses; peer advisers are trained to help colleagues. About 300 union and 80 company leaders met before bargaining in February 2025.
What it supports: That a large employer and its unions have written worker involvement into AI decisions, with joint authority over investment, and how they got there.
What it does not support: That it works: the summary reports no outcomes. One US healthcare employer with a long-standing labour partnership. The report was read only through MIT Sloan's own summary.

worker voice · collective bargaining · AI governance · decision rights · how should employees be involved in decisions about ai

Working paper#

Kropp, B., Bedard, J., Wiles, E., Hsu, A. and Krayer, L. (2026). Research: Why You Shouldn't Treat AI Agents Like Employees

Harvard Business Review, 6 May 2026. Abstract read at source 24 October 2026

Method: Experiment reported in a practitioner journal; full method not read.
Finding: Anthropomorphising AI agents reduced individual accountability, increased unnecessary escalation and lowered review quality.
What it supports: That in the experiment, treating agents like staff weakened oversight.
What it does not support: Effect sizes or generality; only the summary was read.

AI agents · accountability · agent boss · review quality · what is an agent boss

Working paper#

Lambert, P. J. and Schindler, Y. (2026). The Broken Ladder: AI, Remote Work, and Early-Career Hiring

Discussion paper 228/26 (hosted at rfberlin.com, dated May 2026 on its masthead, series date September 2026); also on SSRN, abstract 6787638. Not peer-reviewed. Read at source 3 October 2026 through a summary of the PDF; the SSRN page returned a rate-limit error

Method: Two microdata sources across the US, UK, Canada and Australia, 2017 to 2025: 243 million employer-employee linked records of new hires from Revelio Labs resume data and 407 million online job vacancy postings from Lightcast. Occupation-region and firm-level analysis comparing remote-work (WFH) exposure and generative-AI exposure as predictors of the junior share of hiring.
Finding: By 2025 a two-standard-deviation increase in WFH exposure predicted a fall of around 5 percentage points in the junior share of new hires and around 3 points in the share of job ads requiring limited experience. Estimated jointly, the WFH effect remains while the GenAI coefficient attenuates heavily and often becomes statistically insignificant.
What it supports: That in these data the early-career hiring fall is associated with remote-work exposure at least as strongly as with AI exposure, and that the AI association weakens when both are entered.
What it does not support: That AI has no effect. The authors say they do not interpret the evidence as ruling out strong impacts of GenAI on labour markets, that it bears only on relative junior and senior hiring up to 2025, and that a longer post-period is needed. Remote-work and AI exposure are strongly correlated across occupations, which limits what separating them can show. Unrefereed. Author affiliations as stated on the paper: Warwick and LSE; Ellison Institute of Technology, Oxford.

early careers · remote work · entry-level hiring · missing rungs · is it harder to get a first job now

Institutional survey#

Makridis, C. (Gallup) (2026). Using AI More Does Not Reassure Workers, Managers Do

Gallup Workplace, 9 September 2026. Read at source, including the Survey Methods section, 28 September 2026

Method: Gallup Panel, probability-based recruitment; self-administered web surveys in four waves from 2023 to the first quarter of 2026 among employed US adults aged 18 and over; nearly 30,000 worker observations, about 30 per cent of respondents observed in two or more waves; weighted to Current Population Survey targets. Workers rated how likely their current job was to be eliminated within five years because of new technology, automation, robots or artificial intelligence, and how often they used AI. Comparisons within broad occupation, with demographic controls, and within the same worker over time, netting out occupation and industry shifts.
Finding: Workers using AI daily or several times a week were more than twice as likely to say their job was very likely to be eliminated within five years as those using it a few times a month or year (6.3 against 3.09 per cent in Q1 2026). Roughly 19 per cent of all workers said elimination was somewhat or very likely. Among frequent users, feeling respected, and feeling that the organisation cared about their wellbeing, went with a smaller association between use and fear (6.8 and 11.1 points respectively); among infrequent users management made little difference. Workers who feared elimination scored lower on engagement and job satisfaction and were more likely to be job-hunting.
What it supports: That in a large US panel, heavier AI use goes with more fear of displacement rather than less, including when the same worker is followed over time, and that the gap is smaller where workers feel respected and cared for.
What it does not support: That using AI causes the fear, or that management causes the reduction: the design is observational, and Gallup says the strongest within-worker comparisons rest on the smaller multi-wave group and are less precise. Expectations, not jobs actually lost. The item names technology, automation and robots as well as AI. US only.

job displacement · AI adoption · management · worker wellbeing · does using ai make you less worried about your job

Vendor research#

Massenkoff, M. and McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence

Anthropic, 5 March 2026. Corrected 8 March 2026. Read at source 5 September 2026

Method: A new occupational exposure measure, observed exposure, built from three inputs: O*NET task lists for roughly 800 US occupations, Eloundou and colleagues' theoretical LLM task-exposure scores, and Anthropic's own Claude usage data from the Anthropic Economic Index, weighting automated over augmentative use. Employment outcomes come from the Current Population Survey, in a difference-in-differences comparison of the top exposure quartile against the 30 per cent of workers with zero measured exposure.
Finding: No systematic increase in unemployment for highly exposed workers since late 2022; the pooled difference-in-differences estimate is small and indistinguishable from zero. On hiring, the monthly job-finding rate for 22 to 25 year olds entering the most exposed occupations fell by about 14 per cent in the post-ChatGPT period against 2022. The stable comparator runs at about 2 per cent per month in less exposed occupations, and entry into the most exposed jobs falls by roughly half a percentage point. The authors' own qualification, in their words, is that this is 'just barely statistically significant'. No such decrease appears for workers over 25.
What it supports: That a slowdown in youth hiring into exposed occupations shows up in a second dataset and a second method, alongside Brynjolfsson and colleagues on ADP payroll data. Also that the harder claim, a rise in unemployment among exposed workers, does not appear: the authors estimate they could detect a differential increase of about one percentage point and see nothing.
What it does not support: The 14 per cent is routinely restated as a fall against low-exposure peers. It is not. The paper's own words are 'compared to that in 2022 in the exposed occupations', so it is a change over time within one group, which is a weaker claim than the comparison usually reported. Nor does it establish cause: the authors list three benign readings, that unhired young workers may be staying in existing jobs, taking different ones, or returning to study, and note that survey-measured job transitions are prone to mismeasurement. TWO FURTHER CAUTIONS. The treatment variable is built from Anthropic's own product telemetry and cannot be reconstructed from outside the company. And on 8 March 2026 Anthropic corrected Figure 7, the job-finding figure this number is taken from, which had reversed the labels between the top-quartile and zero-exposure groups. Anyone citing a version of this chart captured between 5 and 8 March 2026 is citing the reversed one.

early careers · jobs · exposure · hiring · will ai replace entry level jobs should juniors use ai the most quoted ai statistics checked what is the ai employment gap

Operator account#

Microsoft (2026). Audit logs for Copilot and AI applications

Microsoft Learn, Purview documentation, page dated 26 August 2026. Read at source 29 September 2026

Method: Vendor technical documentation listing the properties written to the audit log for Copilot and AI applications.
Finding: An audit record typically contains a prompt and response pair, flags whether a message is a prompt, flags a detected jailbreak attempt, and references the files, sites and other resources Copilot or other AI applications accessed to answer the prompt.
What it supports: What kinds of event the audit log is documented to record about an AI interaction.
What it does not support: That the audit record stores the wording of a prompt: the page does not say so, and that question is answered, for the mailbox copy, by the retention documentation. Nothing about how any organisation uses the log.

audit log · workplace monitoring · can my employer see what i put into ai

Argued perspective#

Nudelman, S., Kleppe, A., Bickford, J., Bajic-Bizumic, B., Katsov, K. and Yehuda, E. (BCG) (2026). The Authorization Gap: Why Yesterday's Controls Won't Work with Today's Agents

BCG, 22 September 2026. Read at source 24 October 2026

Method: Argued perspective drawing on BCG adoption data.
Finding: 35 per cent of organisations have agentic AI in production and 44 per cent are planning deployments; more than a third expect agents to hold independent decision rights within three years; 'A reviewer who receives a vague request after the agent has done the hard work is a rubber stamp.'
What it supports: That a major consultancy treats late, vague human review of agents as no control at all.
What it does not support: Anything measured about review quality.

AI agents · authorisation · rubber stamp · oversight · what is an agent boss how should a board govern ai agents

Operator account#

OpenAI (2026). Data access for your managed ChatGPT account

OpenAI Help Center article 20001067. Read at source 29 September 2026

Method: Vendor help documentation describing what an administrator of a managed ChatGPT account may access. Not a study: a statement by the operator of the system about its own product's capability.
Finding: Administrators may access content a user submits (prompts, uploaded files and outputs), conversation history, usage metadata and security settings, and can export, audit, retain, delete and opt in to share data tied to the account. Users are given a Managed ChatGPT Account Notice at set-up. A managed account and a personal account signed in together remain separate.
What it supports: What the vendor says an employer administrator is able to reach in a managed account, and that the vendor says users are notified of this at set-up.
What it does not support: How often any employer reads staff conversations, what any particular employer has switched on, or what happens on a personal account, a company device or a network. Capability as stated by the vendor, with no independent verification, and the page can change.

managed account · workplace monitoring · shadow AI · can my employer see what i put into ai who owns the capability you built at work what it would take to leave

Operator account#

OpenAI (2026). Enterprise privacy at OpenAI

OpenAI, enterprise privacy page. Read at source 29 September 2026

Method: Vendor privacy documentation for its business products. A statement of policy and capability by the operator, not a measurement.
Finding: Within an organisation end users can view their own conversations, and workspace administrators can access an audit log of conversations and GPTs through the Enterprise Compliance API. Administrators control how long data is retained. Deleted conversations are removed from OpenAI's systems within 30 days unless legally required to be retained. By default business data is not used to train models.
What it supports: The vendor's stated position on administrator access, retention and training use for its business products.
What it does not support: Any practice at any employer, or the position for personal accounts. It describes OpenAI's own handling, and says nothing about what an employer does with data once it has been exported.

managed account · workplace monitoring · data retention · can my employer see what i put into ai who owns the capability you built at work what it would take to leave

Operator account#

OpenAI (2026). How your data is used to improve model performance

OpenAI Help Centre article 5722486. The page carries a relative 'Updated: 3 hours ago' note and no calendar date. Read at source on 19 October 2026.

Method: A vendor's statement of its own data policy. Describes settings, measures nothing.
Finding: Content from the Free, Plus and Pro plans may be used to train models unless the user opts out, by turning off 'Improve the model for everyone' under Data controls or using the Privacy Portal. Temporary Chats are not used for training while they stay temporary. Business, Enterprise and Edu plans and the API are excluded by default.
What it supports: OpenAI's stated default and opt-out for training use of conversations, by plan, on the date read.
What it does not support: What any other provider does, how the policy has changed, or that it is followed. The page is undated and may change.

training data · data controls · AI privacy · what should i never put into ai

Institutional survey#

PwC (2026). Global Workforce Hopes and Fears Survey 2026

PwC, 29 September 2026; fielded May and June 2026. Read at source 24 October 2026

Method: Survey of 49,364 workers in 48 countries and regions and 29 sectors.
Finding: 64 per cent used AI at work in the past year, up ten points; daily GenAI use rose from 14 to 22 per cent; access to learning and development fell from 59 to 51 per cent; 31 per cent of daily users name accuracy or quality as their top barrier; 34 per cent of daily users say AI helps with the most complex parts of their job.
What it supports: That use is rising while access to development and confidence in quality lag.
What it does not support: Actual output quality or skill change.

AI use · learning and development · quality · AI slop · what is the verification bottleneck what is ai slop missing rungs

Working paper#

Rohde, W. (AiSuNe Foundation) (2026). Short-Term Gain, Long-Term Fragility: AI Labor Substitution and the Erosion of Sustainable Capability

SSRN abstract 6577818, written 20 April 2026, revised 27 April 2026; also arXiv 2605.27399

Method: Sole-authored conceptual synthesis, 19 pages, no new empirical data. The author's own words: "This paper is a conceptual synthesis rather than a new empirical study" and "the evidentiary strategy is selective and scoped". Preprint, not peer reviewed, no journal reference. Dates verified at the SSRN record on 28 August 2026; re-checked 12 October 2026: arXiv 2605.27399 shows v1 submitted 23 April 2026, and the text states no journal or venue. The phrase 'capability debt' is absent from the abstract but appears five times in the body (sections 1, 4, 5, 10 and 11), so a check of the abstract alone gives the wrong answer.
Finding: Develops a mechanism of capability masking followed by capability erosion: AI output creates a persuasive appearance that organisational capability has been replaced while dependence on skilled human labour remains, supporting hiring restraint and deferred structural reform while costs accumulate. Frames the result as a stack of deferred obligations: technical debt in artifacts and systems, capability debt in the human layer that maintains them, and institutional debt in the wider structures that reproduce skill and resilience.
What it supports: That the capability-erosion argument has been reached independently, from software engineering and political economy rather than from organisational research, and that the masking-before-erosion sequence is not unique to this research's account.
What it does not support: Anything measured. It is a preprint that argues from other people's empirical work rather than presenting its own, and its societal-scale claims about fragility and concentration of power are inference rather than finding. It is also the reason no claim of first use is made for the term capability debt on this site. Checked at source on 28 August 2026: Rohde does not claim to have coined capability debt either. The paper contains no claiming language for it, and what he does claim is a mechanism, "to identify and formalize a mechanism of capability masking and capability erosion".

capability debt · deskilling · labour substitution · capability debt cognitive debt and capability debt

Institutional survey#

Saad, L. (Gallup) (2026). More U.S. Workers Fear Losing Their Jobs to Technology

Gallup News, 15 September 2026. Read at source 28 September 2026

Method: Gallup's annual Work and Education telephone poll, 3 to 24 August 2026: 1,200 US adults, of whom 507 were employed. Margin of error plus or minus 4 points nationally and 6 points for the employed sample.
Finding: 27 per cent of US workers worried that technology could make their job obsolete, a new high, up from 20 per cent in 2025 and 13 per cent when the question was first asked in 2017. 34 per cent of workers aged 18 to 44 against 19 per cent of those 45 and over.
What it supports: That worry about technological obsolescence among US workers rose sharply between 2025 and 2026 on a consistent question.
What it does not support: Anything about AI use: the release carries no breakdown by how often workers use AI. A different instrument, mode and question from the Gallup Panel study, so the two figures cannot be combined. Small employed sample, plus or minus 6 points.

job displacement · worker attitudes · does using ai make you less worried about your job

Working paper#

Shao, Y., Zope, H., Jiang, Y., Pei, J., Nguyen, D., Brynjolfsson, E. and Yang, D. (2026). Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce

arXiv:2506.06576, v3 revised 1 February 2026

Method: Audio-enhanced mini-interviews with 1,500 US domain workers across 104 occupations, covering 844 tasks drawn from O*NET, paired with capability assessments from AI experts. Introduces the Human Agency Scale, H1 to H5, and sorts tasks into four zones by desire against capability.
Finding: Worker preferences diverge sharply from technical capability. Tasks fall into an Automation Green Light Zone, an Automation Red Light Zone where capability exists and workers do not want it used, an R&D Opportunity Zone and a Low Priority Zone. Human Agency Scale profiles vary widely by occupation, and the authors report early signals of core competencies shifting from information-focused skills towards interpersonal ones.
What it supports: That the automate-or-not framing is too coarse, and that there is a measurable, occupation-specific preferred level of human involvement which does not track what the technology can do.
What it does not support: Nothing about what happens to capability when a task is automated. It measures what workers WANT and what experts think is POSSIBLE, which are both stated positions rather than outcomes. A preprint, not peer reviewed, and US-only.

human agency scale · augmentation · automation · what stays human questions which tasks do workers not want automated what becomes more valuable as ai gets cheaper

Institutional survey#

Stack Overflow (2026). 2026 Developer Survey: AI

Stack Overflow, survey results page and press release of 6 October 2026. Both read at source 16 October 2026

Method: Annual self-selected online survey of developers, in its sixteenth year: over 30,000 responses from 169 countries across 103 questions. Per-question respondent counts vary and are shown on the results page.
Finding: 'When asked when it's acceptable to trust AI, 48% indicate they trust AI when they can easily validate the answers'; 6.6% 'are prepared to trust it when it comes to important decisions'; 78.7% say source attribution is important or very important. Coding assistants or agents are used by 66%; 17% of those who avoid AI cite fear of skill atrophy or self-replacement; 24% say their organisation publishes approved tools or an AI policy. 'That's why human-in-the-loop is the dominant mode of working with these tools.'
What it supports: That among the developers who answered, trust in AI output is conditional on being able to check it, and very few would trust it for important decisions.
What it does not support: Anything about developers in general: respondents self-select from Stack Overflow's audience, and the survey is run by a company whose product is a source of checkable answers.

developers · trust · verification · human-in-the-loop · skill atrophy · what is the judgment class

Institutional survey#

WTW (2026). 2026 Employee Experience Global Market Study: uncertainty about future skills is emerging as the next workforce risk

WTW press release via GlobeNewswire, New York, 7 October 2026, reported by HR Dive on 13 October 2026. Release read at source 16 October 2026; the study itself was not read

Method: Responses from 549 employer organisations representing 5.6 million employees, and from 7,000 employees across the United States, United Kingdom, Japan, France and Saudi Arabia. Fieldwork dates not stated.
Finding: 'The proportion of employees who say their manager helps develop their abilities has shifted from 84% to 78% over the past four years.' 'Only about 30% of organizations have established processes for identifying, assessing and tracking workforce skills.' 'This mismatch risks creating "capability anxiety": uncertainty among employees about whether their skills will remain relevant, what capabilities will matter most and where they should focus their development.' 45% of employees say AI is already important to their daily job.
What it supports: That in a large multi-country survey, fewer employees report their manager developing their abilities than four years ago, and most employers do not track skills.
What it does not support: Any link between the fall and AI, which the release does not claim; WTW is a consultancy that sells employee-experience and skills services, and the method is not published in the release.

skills · managers · capability anxiety · employee experience · development · capability debt

Institutional modelling#

Waltmann, B. (2026). New Estimates of the Impact of Undergraduate Degrees on Lifetime Earnings

Institute for Fiscal Studies, commissioned by the Department for Education

Method: Administrative linkage of school, university and tax records for the whole 2002 English GCSE cohort, tracked to age 37 with earnings simulated to 67.
Finding: Average net lifetime return to a degree is around 100,000 pounds, with very large variation by subject. Medicine and economics exceed 400,000 pounds on average. Creative arts, philosophy and languages show low or negative average returns. Around 20 per cent of women and 30 per cent of men are projected to see a negative net return.
What it supports: That subject choice carries a far larger financial spread than the decision to attend at all.
What it does not support: What any subject will return to someone choosing today. The authors explicitly decline to model structural change including AI.

· what should i tell my children to study

Institutional synthesis#

World Economic Forum in collaboration with PwC (2026). Artificial intelligence and the future of entry-level work: a framework for safeguarding and reinventing early career pathways

World Economic Forum white paper, 22 June 2026. Publication page read at source 26 September 2026; the full PDF was not machine readable from this environment

Method: A framework developed from consultation with global stakeholders and workforce leaders across industries, organised on four dimensions: job access, job design, talent pipelines and education system alignment.
Finding: States that more than one in three young workers are employed in occupations with medium to high exposure to AI-driven task change, and proposes that organisations establish clear models for AI-enabled job redesign that use entry-level roles to keep building critical skills rather than layering AI onto existing ways of working.
What it supports: That a standing international body has framed the entry-level problem as one of job design rather than only of hiring volume, and has published a vocabulary for it that organisations are likely to encounter.
What it does not support: Any rate. The exposure figure is an exposure estimate rather than a measurement of jobs lost or roles changed, and the four dimensions are a consultative framework reporting no new primary data. The full report was read only through its publication page.

entry-level work · missing rungs · job design · early careers · missing rungs how do juniors become senior

Qualitative field study#

Ye, X. M. and Ranganathan, A. (UC Berkeley Haas) (2026). AI Doesn't Reduce Work, It Intensifies It

Harvard Business Review, 9 February 2026, based on in-progress research described in the UC Berkeley Haas newsroom announcement of 18 February 2026. The HBR article is paywalled beyond its introduction; graded from the Haas announcement, read in full at source and consistent with the accessible HBR text

Method: Eight-month ethnographic study inside a single 200-person US technology company: real-time workplace observation, meeting attendance, informal conversations and more than 40 semi-structured interviews across functional groups. Described by the authors as in-progress research; no percentages, no control condition and no claim to generalise beyond the one firm studied.
Finding: Rather than reducing work, generative AI intensified it. Employees worked at a faster pace, took on a broader scope of tasks, and extended work into more hours of the day, often without being asked to. Three mechanisms recurred: job scope expanded, tasks bled into non-work time such as lunch and evenings, and people ran AI processes in parallel with other tasks rather than being freed by them.
What it supports: That inside at least one organisation, the time AI saves on a task is not experienced by the people doing the work as time returned; it is absorbed into more tasks, a longer day, or both, through three specific, observed mechanisms.
What it does not support: A rate, a percentage, or anything about organisations generally. One firm, one method, one still-unpublished study: strong on mechanism, silent on how common the pattern is.

work intensification · time savings · task creep · generative AI adoption · what should happen to the time ai saves

Undisclosed frame#

unitQ (2026). New study: Professionals say AI is already reshaping research hiring, and early-career roles are the most exposed

PR Newswire, San Francisco, 3 September 2026. A press release from a vendor that sells the research platform the study was run on. Read in full at source 20 September 2026.

Method: Not stated. The release reports a study of '160 research, design and insights professionals' and gives no field dates, no recruitment source, no screening criteria, no geography, no margin of error and no base size for any subgroup figure. Its two body links go to unitQ's home page and to its product page; there is no report, topline or methodology page to follow. The release says of itself that 'the method was published alongside the finding'. The instrument and the product are the same thing: 'unitQ conducted the study using unitQ Research, an AI research platform for teams', available with a free 30-day trial.
Finding: Verbatim: 42 per cent 'said AI had reduced or slowed hiring or contributed to roles going unfilled'; 28 per cent 'said early-career researchers will feel the impact before anyone else'; 'more than half pointed to execution-heavy work as the first thing to be handed off'. Two further figures are qualified 'Among research and insights professionals', a subgroup of the 160 whose size is never given: 40 per cent 'warned of deskilling or of shallow, homogenized output' and 51 per cent 'said the people who thrive alongside AI are the ones fluent enough to direct it and to recognize when it's wrong'. An unnamed respondent is quoted calling the result an 'expertise drought'.
What it supports: That a vendor published these numbers on this date, and that its chief executive states the ladder-removal argument in public. Nothing else. It is carried in this base as a worked example of the third way a sampling frame fails, not as a measurement.
What it does not support: Any rate in any population. With no frame disclosed, the 42 and 28 per cent have no stated base at all, and the 40 and 51 per cent are percentages of an unstated subgroup that excludes the designers included in the 160. Writing '40 per cent of 160 professionals' would misstate the source. The release's own claim that a method was published is contradicted by the document containing it, which is the reason this entry exists.

sampling frame · deskilling · entry-level · vendor research · early careers · what is the illusion of competence

Where this evidence is used

The research that rests most on this section:

The other sections

The whole base is also machine-readable at evidence.json under CC BY 4.0.

Reference · SS-2026-448 · How I use AI here

Cite this page

Hirji, R. (2026). Work, jobs and the labour market. The SuperSkills evidence base, SS-2026-448. https://thesuperskills.com/research/evidence/jobs-and-the-labour-market. Last reviewed 3 November 2026.

An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.

How citations and IDs work