Research · AI and the Labour Market, Signal, Noise and Structural Change
Back to researchResearch sweep · deep · 2022 – 2026
AI and the Labour Market, Signal, Noise and Structural Change
Evidence of AI-related labour-market change from November 2022–August 2026: employment, vacancies, wages and productivity; occupational and demographic distribution; job creation, displacement and work redesign; and competing economic interpretations using ONS, BLS, Eurostat, OECD, ILO and IMF evidence.
Synthesised 2026-08-14
AI has changed work faster than it has changed employment
Overview
Published statistics do not yet show a distinct economy-wide AI labour-market effect. For the UK, the evidence is especially weak: the retrieved material contains no ONS series linked to firm-level AI adoption, so movements in employment, vacancies, wages, hours or productivity cannot be attributed credibly to AI. Across the US, EU and wider OECD, aggregate outcomes remain consistent with several competing causes, including weak demand, higher interest rates, post-pandemic normalisation, ageing, migration changes and public-sector restraint.
Sources: OECD (2026) (↗); OECD (2025) (↗)
A narrower verdict is possible. Adoption, task redesign and technical capability have advanced rapidly. Several US and European studies also detect weaker demand for junior workers in highly exposed occupations, particularly administration, software, accounting, professional services and customer support. These findings may be the first labour-market signal, but they do not yet establish net job destruction or prove that AI caused the divergence.
Sources: Bloomberg (2025) (↗); SSRN (2025) (↗); SSRN (2025) (↗)
The defining shift since early 2025 has been from conversational assistance towards systems that write code, use tools and complete multi-step tasks. That expands the range of potentially substitutable work, but capability is not deployment, deployment is not productivity, and productivity is not necessarily lower employment. METR’s finding that experienced developers became 19% slower with early-2025 tools illustrates the distance between benchmark capability and realised production.
Sources: OpenAI (2025) (↗); Anthropic (2025) (↗); Google DeepMind (2025) (↗); METR (2025) (↗)
Three evidentiary categories should remain separate. Measured change includes payroll, vacancies, wages, hours and experimental output. Plausible inference links those outcomes to actual adoption with a credible comparison group. Speculation includes exposure scores, executive intentions, vendor forecasts and scenarios about automatable hours. Much of the public argument still moves silently from the third category to the first.
Sources: arXiv (2023) (↗); IZA Discussion Paper (2024) (↗); McKinsey Global Institute (2023) (↗)
Timeline
- Generative AI becomes a general workplace technology
- Occupational exposure maps appear
- Customer-support experiments show uneven productivity gains
- Enterprise experimentation expands beyond technical teams
- Rapid worker adoption is measured
- EU workplace deployment broadens
- Administrative records find little realised pay or hours impact
- Software surveys detect work redesign
- Tool-using reasoning and coding agents arrive
- Vacancy studies report weaker demand in exposed work
- US junior employment divergence becomes visible
- EU enterprise adoption reaches one firm in five
- Field experiments reveal strongly conditional gains
- Employers begin attributing some restructuring to AI
- OECD evidence still rejects an aggregate displacement verdict
Sources: National Bureau of Economic Research (2023) (↗); National Bureau of Economic Research (2024) (↗); Eurostat (2025) (↗); OECD (2026) (↗)
Key Findings
1. The UK aggregate case remains unproved
The UK evidence in this sweep does not support a claim that AI has raised unemployment, depressed wages or reduced aggregate hours. A recruiter reported a 12% rise in UK financial-sector vacancies during 2025 as demand for AI and technology specialists increased, but one sectoral vacancy measure cannot represent the labour market. It indicates skill reallocation and possible job creation, not a national employment effect.
Sources: Reuters (2026) (↗)
The comparison is similar elsewhere. OECD assessments place AI beside cyclical weakness and demographic, fiscal, trade and migration pressures. Eurostat measures enterprise use, while US studies provide more granular payroll and vacancy signals. None supplies the linked adoption, worker and productivity data needed to identify a UK-wide causal effect.
Sources: Eurostat (2025) (↗); OECD (2026) (↗)
2. Hiring is moving before headline employment
The strongest negative evidence appears in labour demand rather than unemployment. A study of 285 million US job postings reports a relative 9% post-November 2022 decline in postings for occupations where generative AI can substitute for labour, concentrated in administration and professional services. A separate CPS analysis found no corresponding average employment or earnings divergence in highly exposed occupations through early 2025.
Sources: SSRN (2025) (↗); SSRN (2025) (↗)
That combination is economically coherent. Firms can slow recruitment, leave vacancies unfilled or change required skills long before redundancies affect total employment. It also leaves room for a cyclical explanation: a six-country analysis found that software vacancies began contracting after their 2022 peak, well before AI-related demand rose sharply.
Sources: arXiv (2024) (↗); SSRN (2026) (↗)
3. Junior work is the clearest distributional warning
US payroll research reported a 13% employment decline from late 2022 among workers aged 22 to 25 in highly exposed occupations, relative to less-exposed comparisons. Older workers in the same occupational groups performed better. Résumé and posting research covering 65 million US records likewise associates firm-level generative-AI integration with slower junior hiring.
Sources: Bloomberg (2025) (↗); The Wall Street Journal (2025) (↗); SSRN (2025) (↗)
European evidence points in the same direction without settling causality. Flemish vacancy research found an approximately 23% relative fall in entry-level vacancies in highly exposed occupations, with no comparable decline for experienced workers. If this persists, the issue is not simply fewer jobs today but a thinner mechanism for producing experienced workers tomorrow.
Sources: National Bureau of Economic Research (2025) (↗); Gartner (2026) (↗)
The positive counterexample matters. BNY Mellon and IBM described plans to increase junior recruitment because entrants brought AI skills and because roles were being recast around the technology. AI may therefore penalise generic junior production while rewarding entrants who can specify, test and supervise automated work.
Sources: Bloomberg (2026) (↗); Bloomberg (2026) (↗)
4. Productivity gains are real, local and conditional
The best causal evidence comes from bounded workplace experiments. The customer-support study published in the Quarterly Journal of Economics found higher productivity and improved customer outcomes, with the largest benefits for less-experienced and lower-performing staff. AI acted partly as a mechanism for transmitting practices from stronger workers to novices.
Sources: The Quarterly Journal of Economics (2025) (↗)
Other experiments complicate the result. Across 66 firms, a six-month trial reduced email time by about two hours per week but detected no change in the quantity or composition of completed tasks. An Alibaba customer-service experiment found faster service and better customer ratings, particularly among weaker performers, but no improvement in objective quality and worse results among top agents.
Sources: National Bureau of Economic Research (2025) (↗); arXiv (2026) (↗)
The symmetric implication is important. AI can compress skill differences and lower barriers to competent performance. It can also reduce expert quality, increase verification work or convert saved minutes into slack rather than output. Organisational design determines which result appears.
5. Adoption has outrun measurable economic outcomes
Eurostat found that the share of EU enterprises with at least ten employees using AI rose from 13.5% in 2024 to 20.0% in 2025. US household research also reports rapid generative-AI take-up, while enterprise surveys describe widespread pilots. These are substantial diffusion measures, but they say nothing by themselves about headcount, wages or productivity.
Sources: Eurostat (2025) (↗); National Bureau of Economic Research (2024) (↗); Bain & Company (2024) (↗)
Danish linked survey and administrative records provide a sterner test. Despite widespread workplace adoption, the researchers could rule out average effects on earnings or recorded hours larger than 2% during the first two years after ChatGPT. Adoption was real; a large realised labour-market effect was not.
Sources: IZA Discussion Paper (2024) (↗)
6. Software is the leading laboratory, not the economy in miniature
More than 75% of respondents to DORA’s 2024 technology survey used AI for at least one daily responsibility. Higher reported adoption coincided with better documentation, code quality and review speed, but also with lower delivery throughput and stability. DORA’s 2025 research, based on nearly 5,000 technology professionals, concluded that AI tended to amplify existing organisational strengths and weaknesses.
Sources: Google Cloud Blog, DORA (2024) (↗); DORA (2025) (↗)
Practitioner evidence describes a shift from writing code towards specifying intent, providing context, testing output and governing agents. That may raise the value of architecture, security and domain knowledge while commoditising routine implementation. Yet METR found experienced open-source developers working on familiar repositories became 19% slower, despite expecting to become faster.
Sources: martinfowler.com (2025) (↗); InfoQ (2026) (↗); METR (2025) (↗)
7. Distribution matters more than the net total
The IMF estimates that AI exposes almost 40% of global employment and roughly 60% in advanced economies, but exposure includes both substitution and complementarity. The ILO similarly treats augmentation as more likely than full automation for many jobs while identifying clerical work as unusually exposed. Occupational composition means the incidence can differ by gender, education and income even if aggregate employment barely moves.
Sources: International Labour Organization (2024) (↗); International Monetary Fund (2024) (↗)
Distribution also runs through hiring systems and bargaining power. A controlled résumé-ranking exercise found race-linked and gender-linked variation in GPT recommendations. AI can therefore change access, surveillance and managerial control without eliminating a single position, while productivity gains may accrue to workers, firms or consumers depending on competition and wage-setting institutions.
Sources: Bloomberg (2024) (↗)
8. The economists disagree about the missing second round
Brynjolfsson and colleagues supply direct evidence that AI can augment workers and diffuse expertise inside a firm. Exposure studies and task-based economic models leave open a less benign path in which automation removes routine tasks faster than new work and demand appear. Current data cannot determine which mechanism will dominate at labour-market scale.
Sources: arXiv (2023) (↗); National Bureau of Economic Research (2023) (↗); arXiv (2025) (↗)
The retrieved source set does not contain direct, attributable treatments from David Autor, Daron Acemoglu or Pascual Restrepo, so a precise comparison of their current positions would exceed the evidence assembled here. The underlying disagreement is visible, however: one interpretation expects productivity, new tasks and demand to restore labour demand; another expects cost-saving substitution, weaker worker power and limited productivity spillovers. That is an empirical dispute, not a choice between optimism and pessimism.
Evidence & Data
The most concerning measured figures concern entry rather than exit: a 13% relative employment decline among US workers aged 22 to 25 in exposed occupations, a roughly 23% relative reduction in Flemish entry-level vacancies, and a 9% relative fall in US postings for highly substitutable occupations. Each uses a comparison design, but none fully removes technology-sector contraction, interest rates, post-pandemic hiring excess or other simultaneous shocks.
Sources: Bloomberg (2025) (↗); National Bureau of Economic Research (2025) (↗); SSRN (2025) (↗)
The strongest null result is Denmark’s ability to reject average earnings or hours effects above 2% during the first two years. Gartner offers a different check on corporate rhetoric: it attributed less than 1% of announced first-half 2025 layoffs to AI productivity gains, judging most cuts to be broader strategic repositioning.
Sources: IZA Discussion Paper (2024) (↗); Gartner (2025) (↗)
Adoption and investment data are much larger than outcome estimates. EU enterprise use reached 20% in 2025. CB Insights recorded external AI activity at 349 S&P 500 firms, but five companies accounted for 32% of activity and the median active firm had only three to four documented partnerships. Diffusion is broad but shallow and highly concentrated.
Sources: Eurostat (2025) (↗); CB Insights (2026) (↗)
McKinsey’s estimate that 30% of current US working hours could be automated by 2030 belongs in a different column. It is a scenario conditional on capability, adoption, investment, organisational change and demand, not a measurement of hours already removed. Confusing these quantities turns a planning model into a false statistic.
Sources: McKinsey Global Institute (2023) (↗); McKinsey Global Institute (2024) (↗)
Signals & Tensions
Vacancies may be the canary
Junior and substitutable-job postings are weakening before aggregate employment. That could mark durable labour-saving change, or simply reveal that vacancy data react faster and more violently to the cycle than payrolls do.
Sources: SSRN (2025) (↗); Bharat Chandar on Substack (2025) (↗)
Time saved is not output gained
Customer support produced measurable gains, office workers wrote fewer emails, and experienced developers sometimes slowed down. The underreported variable is the organisation’s ability to convert task-level assistance into better end-to-end output.
Sources: The Quarterly Journal of Economics (2025) (↗); National Bureau of Economic Research (2025) (↗); METR (2025) (↗)
Agents are ahead of the statistics
Frontier systems can complete longer coding and research tasks, but evaluations test bounded assignments rather than maintainable output inside real organisations. Agent capability is a legitimate leading indicator and a poor payroll statistic.
Sources: METR (2025) (↗); METR (2025) (↗); METR (2025) (↗)
Corporate attribution is getting louder
Some employers now cite AI when reducing administration or restraining hiring. Yet Gartner’s layoff analysis and the New York Fed’s business evidence suggest that realised AI-related cuts remain modest relative to restructuring, cost pressure and changing demand.
Sources: Bloomberg (2025) (↗); Bloomberg (2025) (↗); Gartner (2025) (↗)
Junior compression can coexist with inclusion
AI may remove routine apprenticeship tasks while helping novices perform closer to expert level. It may also improve accessibility and lower the cost of expertise. Whether this broadens opportunity or abolishes the training ladder depends on hiring, mentorship and the allocation of saved time.
Sources: The Quarterly Journal of Economics (2025) (↗); Gartner (2026) (↗)
Open Questions
-
UK causal measurement: ONS employment, vacancies, earnings, hours, participation and productivity records need to be linked to representative firm-level adoption data. Without that join, a UK AI effect remains statistically unidentified.
-
The missing career ladder: Researchers need longitudinal evidence showing whether reduced junior recruitment delays progression, raises later skill shortages or merely changes the route through which workers acquire expertise.
-
Net creation: Existing studies measure displaced tasks and altered hiring more readily than new products, firms, occupations and demand. Business-formation, self-employment and occupational-switching records could reveal the positive side earlier.
-
Incidence: Pay, prices, profits and work intensity must be measured together. Productivity gains do not reveal whether workers receive higher wages, firms capture rents, consumers receive lower prices or managers demand more output per hour.
-
Job quality: Representative evidence remains thin on autonomy, monitoring, error responsibility, pace of work and bargaining power. Headcount stability can conceal a material deterioration or improvement in work.
-
Cycle versus structure: A durable AI effect during 2026 to 2028 would show persistent exposure-linked divergence after controlling for sector, age, interest rates, remote work and prior over-hiring. Convergence as demand recovers would favour the cyclical explanation.
-
Policy effectiveness: Retraining, wage insurance, portable benefits, worker consultation, competition enforcement and disclosure rules remain more defensible than an employment tax based on forecasts. Their evaluation should follow measured worker transitions, not exposure rankings.
The decisive evidence will not be another survey asking whether firms use AI. It will be linked administrative data showing what adopters did to jobs, pay, hours and output, and who kept the difference.
Sources
Summary: ↑ Back to summary
Financial Press
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| f1 | AI Makes It Harder for Entry-Level Coders to Find Jobs, Study Says | Bloomberg | 2025-08 | Rachel Metz reports the Stanford payroll-data study finding a 13% fall in employment for young entrants in highly AI-exposed US occupations, making it central evidence on early-career distributional effects. |
| f2 | AI Begins to Disrupt Job Hunters’ Prospects | The Wall Street Journal | 2025-08 | Justin Lahart’s Wall Street Journal report describes the same large payroll-data study, its occupation-by-age design and its contrasting results where AI complements rather than replaces tasks. |
| f3 | Is AI Killing Entry-Level Jobs? Here’s What We Know | Bloomberg | 2025-07 | Jo Constantz separates falling junior vacancies and unemployment from stronger claims of causation, and highlights the career-ladder problem created when routine starter tasks are automated. |
| f4 | New Grads Join Worst Entry-Level Job Market in Years | Bloomberg | 2025-06 | Claire Ballentine documents a weak US graduate hiring market while explicitly identifying hiring freezes and tariff uncertainty as confounders alongside AI substitution of some low-skill roles. |
| f5 | Younger Workers Will Win the AI Economy | Bloomberg | 2025-09 | Walter Frick presents the historical and economic counterargument that younger workers may adapt fastest, cautioning against extrapolating early-career hiring weakness into permanent exclusion. |
| f6 | NY Fed Finds Modest Impact of AI on Jobs Even as Usage Increases | Bloomberg | 2025-09 | Maria Eloisa Capurro reports a New York Fed business survey in which AI use rose substantially but few firms had cut jobs because of it, useful disconfirming evidence although it is self-reported employer data. |
| f7 | US Workers See AI-Induced Productivity Growth, Fed Survey Shows | Bloomberg | 2025-02 | Alexandre Tanzi covers research based on a nationally representative worker survey of generative-AI use and time savings, providing adoption and self-reported productivity evidence rather than employment causality. |
| f8 | Microsoft’s AI Assistants Will Revolutionize the Office, One Day | Bloomberg | 2024-07 | Matt Day’s enterprise-adoption reporting shows that realised workplace gains depend on data preparation, governance and worker training, challenging simple assumptions that licences immediately translate into productivity. |
| f9 | BNY Sees Uptick In Junior Hires For Their AI Skills, CEO Says | Bloomberg | 2026-05 | Yizhu Wang reports Robin Vince’s claim that BNY Mellon has increased analyst and intern hiring, an important firm-level counterexample to the blanket entry-level displacement narrative. |
| f10 | IBM to Triple Entry-Level US Hiring With Roles Recast for AI Era | Bloomberg | 2026-02 | Jo Constantz documents IBM’s announced expansion of entry-level hiring and recasting of roles, evidence of work redesign and complementary demand rather than observed sector-wide employment change. |
| f11 | Companies Are Warming Up to Saying AI Is the Reason for Job Cuts | Bloomberg | 2025-11 | Agnee Ghosh collates corporate announcements linking administrative job cuts and hiring freezes to AI productivity, while the material is best treated as executive attribution rather than independently identified causation. |
| f12 | How AI Has Already Changed My Job | Bloomberg | 2025-05 | Marin Cogan’s cross-occupation reporting illustrates task redesign in nursing, teaching and driving, but is qualitative evidence rather than a measure of employment effects. |
| f13 | OpenAI’s GPT Is a Recruiter’s Dream Tool. Tests Show There’s Racial Bias | Bloomberg | 2024-03 | Bloomberg reporters Leon Yin, Davey Alba and Leonardo Nicoletti conduct a controlled résumé-ranking test that identifies potential racial and gender disparities in AI-assisted recruitment. |
| f14 | AI Forces Return of In-Person Interview | The Wall Street Journal | 2025-08 | Ray A. Smith reports employers restoring face-to-face interviews to counter AI-enabled cheating and impersonation, showing that AI changes recruitment processes as well as labour demand. |
| f15 | AI’s impact on jobs is set to become more pronounced | Financial Times | 2026-01 | Delphine Strauss reports the Financial Times assessment that widespread AI layoffs had not yet materialised by January 2026, while concern focused on graduates and future task reorganisation. |
| f16 | The AI Shift: What millions of job ads reveal about AI displacement | Financial Times | 2026-01 | This Financial Times analysis argues that comparisons of posting trends at adopting and non-adopting firms better fit financial conditions and macroeconomic shocks than an AI-only explanation. |
| f17 | AI Seen Cutting Worker Numbers, Survey By Staffing Company Adecco Shows | Reuters | 2024-04 | Reuters reports an Adecco employer survey on expected five-year staffing reductions, which is relevant as a business-intentions indicator but does not measure realised displacement. |
| f18 | Demand for AI, tech experts pushes UK financial sector vacancies up 12%, recruiter says | Reuters | 2026-01 | Reuters reports recruiter data indicating a 12% increase in UK financial-sector vacancies in 2025 for AI, regulation and data skills, capturing skill-specific demand but not general labour-market impact. |
| f19 | 20% of EU enterprises use AI technologies | Eurostat | 2025-12 | Eurostat provides administrative survey evidence that 20.0% of EU enterprises with at least ten workers used AI in 2025, up from 13.5% in 2024, and specifies coverage, technologies and firm-size limits. |
| f20 | OECD Employment Outlook 2026 | OECD | 2026-07 | The OECD’s latest Employment Outlook provides cross-country labour-market context, notes potential youth exposure, and foregrounds demographic and cyclical forces that complicate attribution to AI. |
| f21 | OECD Employment Outlook 2025 | OECD | 2025-07 | The OECD’s annual assessment documents resilient employment and low unemployment alongside an emerging slowdown, giving a macro baseline against which AI-specific claims must be tested. |
| f22 | What is the possible effect of generative AI on employment? | International Labour Organization | 2024-09 | The ILO applies occupational exposure scores to labour-force-survey structures in more than 140 countries, clearly distinguishing prospective exposure, augmentation and automation from observed job losses. |
| f23 | AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity. | International Monetary Fund | 2024-01 | Kristalina Georgieva outlines the IMF’s estimate that AI could affect roughly 40% of global employment and its distributional risks, but it is a scenario-based exposure assessment rather than realised employment evidence. |
| f24 | Generative AI at Work | The Quarterly Journal of Economics | 2025-05 | Erik Brynjolfsson, Danielle Li and Lindsey Raymond use a real customer-service workplace deployment and difference-in-differences design to show productivity and quality gains, concentrated among less experienced workers. |
| f25 | The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market | Organization Science | 2024 | Xiang Hui, Oren Reshef and Luofeng Zhou study freelancers on a large online platform after major generative-AI releases, offering one of the few direct measures of short-run task-market displacement and adaptation. |
Frontier Lab & Model News
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| t1 | Introducing OpenAI o3 and o4-mini | OpenAI | 2025-04 | OpenAI’s April 2025 release describes reasoning models with integrated web, code, file and image tools, marking a move from answer generation towards agentic execution of professional tasks. |
| t2 | Introducing GPT-5 for developers | OpenAI | 2025-08 | This API announcement frames GPT-5 as OpenAI’s model for coding and agentic tasks and documents new controls for tool calling, reasoning effort and cost-performance choices. |
| t3 | GPT-5 System Card | OpenAI | 2025-08 | OpenAI’s system card records the GPT-5 model family, its routing design, safety evaluations and the decision to apply high biological and chemical capability safeguards to GPT-5-thinking. |
| t4 | Introducing GPT-5.2 | OpenAI | 2025-12 | OpenAI presents GPT-5.2 as a model for professional work and long-running agents, including performance claims on GDPval, coding and tool use that are relevant to task-level workplace substitution claims. |
| t5 | Introducing Claude 4 | Anthropic | 2025-05 | Anthropic’s Claude 4 announcement documents Opus 4 and Sonnet 4 as coding and long-running agent models, including tool use during extended reasoning and persistent-memory features. |
| t6 | Claude 4 System Card | Anthropic | 2025-05 | Anthropic’s primary technical safety document reports evaluations, threat models and Responsible Scaling Policy treatment for Claude Opus 4 and Sonnet 4. |
| t7 | Anthropic Economic Index: Insights from Claude 3.7 Sonnet | Anthropic | 2025-03 | Using observed Claude.ai interactions after the Claude 3.7 release, this report finds increasing coding, education, science and healthcare use, but measures product usage rather than labour-market outcomes. |
| t8 | Anthropic Economic Index: AI’s impact on software development | Anthropic | 2025-04 | This analysis of 500,000 coding-related Claude interactions distinguishes ordinary Claude.ai use from Claude Code agent use and provides a direct leading indicator for work redesign in software occupations. |
| t9 | Estimating AI productivity gains from Claude conversations | Anthropic | 2025-11 | Anthropic extrapolates from 100,000 sampled conversations and model-estimated task times to a possible 1.8 percentage point annual productivity-growth increase, while explicitly stating that this is not a forecast of realised productivity. |
| t10 | Anthropic Economic Index report: Economic primitives | Anthropic | 2026-01 | This report adds task complexity, skill, purpose, autonomy and success measures to transcript analysis, finding concentrated use in high-human-capital tasks but acknowledging that observed conversations do not straightforwardly map to real-world job change. |
| t11 | Anthropic Economic Index report: Cadences | Anthropic | 2026-06 | Anthropic’s June 2026 report combines usage measures with a user survey and explicitly documents severe occupational selection, with computer and mathematical workers heavily over-represented among respondents. |
| t12 | Gemini 2.5: Our most intelligent AI model | Google DeepMind | 2025-03 | Google’s Gemini 2.5 announcement presents a reasoning-first model with claimed advances in coding, mathematics and science, showing the rapid spread of test-time reasoning across frontier systems. |
| t13 | Model cards | Google DeepMind | 2026-07 | Google DeepMind’s model-card index provides a structured record of Gemini releases and updates, including Gemini 2.5 Pro, Deep Think and Computer Use, useful for tracking capability diffusion into agentic workflows. |
| t14 | Advancing Gemini’s security safeguards | Google DeepMind | 2025-05 | This technical safety post identifies indirect prompt injection as a practical risk for agents that access workplace emails, documents and websites, a constraint on autonomous work deployment. |
| t15 | The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation | Meta AI | 2025-04 | Meta introduces open-weight Llama 4 Scout and Maverick as multimodal mixture-of-experts models with long context, expanding the set of systems organisations can run or customise outside closed-model APIs. |
| t16 | Everything we announced at our first-ever LlamaCon | Meta AI | 2025-04 | Meta’s LlamaCon announcement adds a limited-preview Llama API, customisation tools and security evaluation tools, indicating a broader enterprise deployment pathway for open-weight models. |
| t17 | Au Large | Mistral AI | 2024-02 | Mistral’s Mistral Large launch is an early European frontier-model release focused on multilingual reasoning, code generation and API distribution, relevant to the widening supplier base after 2022. |
| t18 | Devstral | Mistral AI | 2025-05 | Mistral and All Hands AI describe Devstral as an Apache-licensed agentic model trained on real GitHub issues, directly targeting software-engineering work rather than isolated code completion. |
| t19 | Latest news | Mistral AI | 2025-07 | Mistral’s primary news archive records the rapid 2025 rollout of agents, coding models, enterprise products, OCR and reasoning systems, helping establish the pace and breadth of deployment-oriented releases. |
| t20 | Grok 4 | xAI | 2025-07 | xAI’s Grok 4 release claims improved reasoning, native tool use and agentic benchmark performance, including a high-cost Grok 4 Heavy configuration based on parallel test-time compute. |
| t21 | Grok 4 Model Card | xAI | 2025-08 | xAI’s model card is the primary safety and capability documentation for Grok 4, covering concerning propensities, dual-use evaluations and deployment contexts. |
| t22 | Grok 4 Fast | xAI | 2025-09 | This release shows the commercial push to lower the cost of reasoning and search-enabled agents, a key condition for broad workplace adoption even when absolute model capability changes little. |
| t23 | Measuring AI Ability to Complete Long Tasks | METR | 2025-03 | METR’s reproducible time-horizon evaluation estimates that the duration of software tasks frontier agents can complete with 50% reliability has doubled roughly every seven months, but its extrapolation remains a forecast rather than labour-market evidence. |
| t24 | Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity | METR | 2025-07 | METR’s randomised controlled trial of 16 experienced developers and 246 tasks found AI access made work 19% slower on average, providing an important counterweight to vendor productivity claims. |
| t25 | Frontier Risk Report (February to March 2026) | METR | 2026-06 | METR’s external risk report records evaluations of frontier systems and their performance under substantial agentic task budgets, offering independent evidence on capability progression and remaining limits. |
Academic & arXiv
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| a1 | GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models | arXiv | 2023-03 | Eloundou, Manning, Mishkin and Rock provide the foundational task-exposure framework, estimating potential LLM effects across US occupations rather than realised employment outcomes. |
| a2 | Generative AI at Work | National Bureau of Economic Research | 2023-04 | Brynjolfsson, Li and Raymond use a staggered workplace rollout among 5,179 customer-support agents to estimate a causal 14% productivity increase, concentrated among novice and lower-skilled workers. |
| a3 | The Adoption of ChatGPT | IZA Discussion Paper | 2024-05 | Humlum and Vestergaard link a large Danish survey experiment to register data, documenting adoption patterns and employer restrictions among workers in 11 exposed occupations. |
| a4 | The Rapid Adoption of Generative AI | National Bureau of Economic Research | 2024-09 | Bick, Blandin and Deming provide nationally representative US survey evidence on work and household use, including estimates of work hours assisted and self-reported time savings. |
| a5 | Generative AI Impact on Labor Market: Analyzing ChatGPT's Demand in Job Advertisements | arXiv | 2024-12 | Ahmadi, Khosh Kheslat and Akintomide use US advertisements from May to December 2023 to identify ChatGPT-related skill clusters, measuring employer demand for AI skills rather than realised jobs. |
| a6 | Augmenting or Automating Labor? The Effect of AI Development on New Work, Employment, and Wages | arXiv | 2025-03 | This historical US study distinguishes labour-augmenting from automating AI exposure between 2015 and 2022, offering a pre-ChatGPT baseline for interpreting wage inequality and new-work creation. |
| a7 | Generative AI Adoption and Higher Order Skills | arXiv | 2025-03 | Gulati, Marchetti, Puranam and Sevcenko analyse postings at 378 US public firms recruiting for GenAI skills, finding higher advertised cognitive and social-skill requirements in GenAI roles. |
| a8 | Shifting Work Patterns with Generative AI | National Bureau of Economic Research | 2025-05 | Dillon, Jaffe, Immorlica and Stanton report a randomised experiment across 66 firms and 7,137 knowledge workers, finding time savings in email but no detected changes in task volume or composition. |
| a9 | Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI | National Bureau of Economic Research | 2025-05 | Humlum and Vestergaard combine Danish adoption surveys with administrative worker and workplace records, finding precise null effects on earnings and hours while documenting task reorganisation and occupational switching. |
| a10 | Tracking Employment Changes in AI-Exposed Jobs | SSRN | 2025-06 | Chandar uses US CPS data from Q4 2022 to Q1 2025 and finds no average employment or earnings-growth gap by GenAI exposure, while showing heterogeneity between software and customer-service occupations. |
| a11 | Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data | SSRN | 2025-08 | Hosseini Maasoum and Lichtinger analyse 65 million US résumés across more than 280,000 firms and associate GenAI-integrator hiring with lower junior employment driven by slower hiring. |
| a12 | Labor Demand in the Shadow of Generative AI: Evidence from the U.S. Job Posting Data | SSRN | 2025-09 | Liu, Wang and Yu apply difference-in-differences and event studies to 285 million US postings from 2018Q1 to 2025Q2, reporting relative declines in high-substitution occupations while controlling for interest-rate changes. |
| a13 | Generative AI and Firm Productivity: Field Experiments in Online Retail | arXiv | 2025-10 | Fang, Yuan, Zhang, Donati and Sarvary report large-scale randomised experiments across seven retail workflows, providing causal evidence on firm-level productivity and heterogeneous gains for smaller sellers and less experienced consumers. |
| a14 | Generative AI in Action: Field Experimental Evidence from Alibaba's Customer Service Operations | arXiv | 2026-02 | Ni and colleagues use a large Alibaba customer-service experiment to separate access from use, finding faster service and heterogeneous quality effects across worker performance levels. |
| a15 | Hiring Up, Not Down: Generative AI and the Composition of Labor Demand | SSRN | 2026-05 | Mahieu uses administrative vacancy data for Flanders, 2021 to 2025, and estimates a roughly 23% lower entry-level vacancy rate between the 25th and 75th exposure percentiles at peak adoption. |
| a16 | Generative AI and the Reorganization of Labor Demand | arXiv | 2026-05 | Wang, Wei and Wang use nationwide US posting data to decompose labour-demand adjustment into hiring reallocation and within-job task redesign, with distinct patterns by seniority. |
| a17 | Human Capital, AI, and Labor Commoditization | arXiv | 2026-06 | Siddiq and Zhang analyse Upwork data around ChatGPT's release and report that, in more AI-exposed categories, labour demand places less weight on human-capital signals and more on price. |
| a18 | Generative AI and the Informational Value of Educational Credentials: Evidence from the Master's Margin | SSRN | 2026-07 | Cortes, Dellarocas and Wang use Revelio worker-flow data from 2018 to 2025 to examine whether AI exposure changes employers' use of master's credentials in hiring. |
| a19 | Junior in Title, Senior in Practice: Job Posting Evidence on Entry-Level Software Hiring in the Age of Generative AI | SSRN | 2026-07 | Biswas offers a disconfirming descriptive study across six national software-posting series, finding that the hiring decline predated late-2024 AI-demand growth and warning against simple AI attribution. |
| a20 | HCAST: Human-Calibrated Autonomy Software Tasks | METR | 2025-03 | METR's HCAST supplies human-calibrated measures of agents' autonomous performance on more than 180 software, cybersecurity, machine-learning and reasoning tasks, informing capability rather than labour-market impact. |
| a21 | Evaluating frontier AI R&D capabilities of language model agents against human experts | METR | 2024-11 | METR's RE-Bench work compares frontier agents with human experts on ML research-engineering tasks, giving a bounded leading indicator for possible work substitution in technical research. |
| a22 | Research Update: Algorithmic vs. Holistic Evaluation | METR | 2025-08 | METR connects its developer productivity randomised trial with agent-benchmark results, showing why automatic task success need not yield mergeable work in realistic software development. |
| a23 | How Does Time Horizon Vary Across Domains? | METR | 2025-07 | Kwa and Cheng report METR estimates of frontier-model task time horizons across HCAST, RE-Bench and related suites, while explicitly limiting inference to software and research domains. |
| a24 | Generative AI at Work: From Exposure to Adoption across 35 European Countries | arXiv | 2026-04 | This study uses the 2024 European Working Conditions Survey of more than 36,600 workers to separate occupational exposure from self-reported adoption and early task restructuring across Europe. |
VC & Analyst Reports
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| v1 | Generative AI and the future of work in America | McKinsey Global Institute | 2023-07 | McKinsey Global Institute's July 2023 US scenario model estimates that activities accounting for up to 30% of current work hours could be automated by 2030, while stressing that it models changes in labour-demand mix rather than aggregate employment. |
| v2 | A new future of work: The race to deploy AI and raise skills in Europe and beyond | McKinsey Global Institute | 2024-05 | This 2024 McKinsey Global Institute report compares Europe and the United States, projects 27% and 30% of hours respectively as automatable by 2030, and quantifies disproportionate projected occupational transitions among lower-wage workers. |
| v3 | Enterprise technology’s next chapter: Four gen AI shifts that will reshape business technology | McKinsey | 2024-10 | McKinsey argues that AI-human collaboration could flatten IT organisations, especially operations and help desks, and thin junior roles, making it a useful strategic hypothesis about career-ladder risks rather than an observed labour-market result. |
| v4 | The State of Organizations 2026 | McKinsey | 2026 | McKinsey's survey of 7,904 respondents reports managerial expectations for AI capabilities, including 55% citing exponential productivity gains, but measures beliefs and organisational plans rather than realised output or payroll outcomes. |
| v5 | AI Survey: Four Themes Emerging | Bain & Company | 2024 | Bain's early-2024 enterprise survey found 87% of respondents developing, piloting or deploying genAI, with activity concentrated in code development, customer service, marketing and sales, providing adoption-intention context with an undisclosed sampling frame in the public summary. |
| v6 | Labor 2030: The Collision of Demographics, Automation and Inequality | Bain & Company | 2024 | Bain frames automation alongside demographic scarcity and inequality, estimating potential displacement and wage effects under scenarios, which helps prevent attribution of all workforce adjustment to generative AI alone. |
| v7 | AI Recruits a New Hybrid Workforce | Sequoia Capital | 2023-01 | Sequoia's named investment thesis presents AI as a productivity technology that creates human-software teams, while acknowledging quality control, hallucination and managerial overhead as limits on straightforward labour substitution. |
| v8 | Developer Tools 2.0 | Sequoia Capital | 2023-03 | Sequoia identifies software engineering as an early leading adoption domain and cites GitHub Copilot's early user growth, but its claims are venture-market framing rather than an independent measure of developer employment or productivity. |
| v9 | AI 50: Companies of the Future | Sequoia Capital | 2024-04 | Sequoia's 2024 market map documents enterprise AI applications and cites examples such as ServiceNow's nearly 20% case-avoidance rate, offering company-level operational indicators rather than economy-wide evidence. |
| v10 | Services: The New Software | Sequoia Capital | 2026 | Sequoia's 2026 copilot-to-autopilot thesis argues that agentic systems can sell completed service work rather than software seats, while identifying software engineering as the leading use case and implying future substitution in standardised, verifiable workflows. |
| v11 | Building Tomorrow’s Transformational Companies | Sequoia Capital | 2025 | Sequoia's 2025 fund thesis counters pure displacement narratives by arguing that deploying AI in production expands demand for forward-deployed service workers, although the quoted $500 billion market estimate is an investor assertion. |
| v12 | Future of the workforce: How AI agents will transform enterprise workflows | CB Insights | 2024-07 | CB Insights maps agent deployment opportunities in customer support, sales and engineering and explicitly flags agent reliability as a constraint, useful for distinguishing aspirational autonomous-work claims from deployable workflows. |
| v13 | The Future of the Enterprise AI Buildout | CB Insights | 2026-03 | CB Insights' 2026 analysis of S&P 500 partnerships, investments, acquisitions and hiring finds nearly 70% had documented external AI activity in 2023-25 but that five companies accounted for 32%, evidencing concentrated corporate activity rather than broad labour-market adoption. |
| v14 | Gartner Survey Finds 38% of HR Leaders Reported They Are Piloting, Planning Implementation, or Have Already Implemented Generative AI | Gartner | 2024-02 | Gartner's January 2024 survey of 179 HR leaders shows reported HR-function adoption or planning rose from 19% to 38%, while expected headcount reduction fell from 6.7% to 5.1%, making it a useful but small intentions survey. |
| v15 | Gartner Survey Shows Supply Chain GenAI Productivity Gains at Individual Level, While Creating New Complications for Organizations | Gartner | 2025-02 | Gartner's August 2024 survey of 265 global supply-chain respondents found 4.11 weekly hours saved for desk workers but only 1.5 hours per team member, with no association with improved team output or quality. |
| v16 | AI's Impact on Productivity and Headcount | Gartner | 2025-02 | Gartner's February 2025 CFO-oriented research is a direct warning that many firms have not converted AI investment into material workforce productivity or headcount changes, providing a counterweight to productivity forecasts. |
| v17 | Forget Layoffs: AI Is Coming for Inefficiency, Not People | Gartner | 2025-12 | Gartner's December 2025 review reports that less than 1% of announced first-half 2025 layoffs were attributable to AI productivity gains, while cautioning that this is an analysis of announced layoffs rather than a complete causal employment dataset. |
| v18 | The Hidden Workforce Costs of AI | Gartner | 2026-06 | Gartner's June 2026 workforce-cost thesis predicts that up to 30% of AI-displaced roles could be rehired by 2029 and highlights junior-pipeline and AI-talent-premium costs, but these are forecasts rather than observed outcomes. |
| v19 | Forrester's 2023 Generative AI Jobs Impact Forecast, US | Forrester | 2023-08 | Forrester's US forecast makes the explicit claim that generative AI will reshape substantially more jobs than it eliminates, a useful forecast benchmark that must not be confused with realised employment change. |
| v20 | How To Drive Employee Productivity With Generative AI | Forrester | 2024-03 | Forrester's 2024 report argues that productivity gains vary by role and task and depend on training and upskilling, offering an organisational mechanism for why exposure estimates may not translate into measured output. |
| v21 | Your Employees Aren't Ready For Generative AI Tools | Forrester | 2024-11 | Forrester's November 2024 global benchmark identifies training, understanding and ethical-awareness gaps as barriers to tangible outcomes, reinforcing that provision of AI tools is not equivalent to effective adoption. |
| v22 | Predictions 2026: The Future Of Work | Forrester | 2025-10 | Forrester's 2026 outlook reports that firms chasing AI efficiencies can end up rehiring terminated roles, a strategic forecast that provides a disconfirming mechanism to simple permanent-headcount-reduction claims. |
| v23 | AI Is Rewriting Software Work: What It Means For Your Team | Forrester | 2025-11 | Forrester's 2025 software-development analysis combines practitioner interviews with a claimed quantitative study and identifies reduced entry-level task availability, role convergence and a shift towards orchestration, governance and systems thinking. |
Blogs & Independent Thinkers
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| b1 | A data-driven case that AI has already changed the U.S. labor market | Forked Lightning | 2024-10 | Harvard economist David Deming argues from occupational churn and US employment composition that labour-market change predates generative AI, while setting out why recent shifts require more granular evidence than aggregate employment figures. |
| b2 | The Age of AI, an update: A Quick Q&A with … economist Michael Strain | Faster, Please! | 2025-02 | James Pethokoukis's interview with AEI economist Michael Strain presents a cautiously optimistic view: productivity and invention effects may be substantial, but organisational adaptation and worker skill change are likely to precede large near-term employment shifts. |
| b3 | Generative Artificial Intelligence and the Workforce | Heldrich Center for Workforce Development on Medium | 2025-04 | Jessica Starace and Liana Lin at Rutgers' Heldrich Center map the early literature and emphasise that observed effects are likely to vary by sector and by whether firms use AI to augment, reorganise or replace tasks. |
| b4 | Study: Supplemental GenAI May Soften the Blow to Workers | MIT Initiative on the Digital Economy on Medium | 2025-01 | MIT IDE summarises evidence linking firms' generative-AI exposure to vacancies and wages, with the key interpretation that complementary uses can share productivity gains with workers rather than mechanically displacing them. |
| b5 | Employment and AI | Work in Progress | 2025-06 | Thomas Otter's analysis, cited by other workforce writers, argues that measuring AI success as headcount reduction is a poor organisational objective and distinguishes early software and customer-support signals from slower, constrained adoption in complex sectors such as healthcare. |
| b6 | What can we learn from AI exposure measures? | Justified Posteriors | 2025-07 | Andrey Fradkin and Seth Benzell's interview with Wharton economist Daniel Rock is a useful methodological corrective, explaining why occupational exposure scores should not be read as estimates of job loss or automation. |
| b7 | A Primer on “Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence” | Bharat Chandar on Substack | 2025-08 | Bharat Chandar explains the Stanford Digital Economy Lab and ADP payroll study, which finds relative employment declines for 22-to-25-year-olds in highly AI-exposed occupations but not broad labour-market disruption. |
| b8 | A Note on AI and the Job Market | Jack Meyer on Substack | 2025-11 | Jack Meyer synthesises recent papers on jobs, wages, productivity and hiring, arguing that the clearer early effect may be degraded labour-market signalling and matching rather than widespread replacement of workers. |
| b9 | AI and the Future of Work | Center for Humane Technology | 2025-12 | This Centre for Humane Technology interview with Brookings' Molly Kinder explicitly tests the Stanford payroll findings against competing explanations, cautioning that interest rates, tariffs, pandemic over-hiring and cyclical conditions remain difficult to separate from AI. |
| b10 | Decades of Fearing Automation but Hoping for Augmentation | Tom Davenport on Substack | 2026-03 | Thomas H. Davenport argues that forecasts have repeatedly outrun measurement, accepts the Stanford entry-level result as an important warning, and warns against generalising it to every junior occupation. |
| b11 | AI hiring, slowing, shifting, and still unequal | Bright Data for Journalists | 2026-03 | Jennifer Burns reports a short-run LinkedIn-posting analysis that finds AI vacancies concentrated in software, IT and finance, with mid-senior roles dominant and pronounced gender imbalance, but its small and non-representative sample limits inference about total employment. |
| b12 | AI in Economics | Frankly, the Counterfactual | 2026-04 | Indiana University economist Kosali Simon's research roundup highlights emerging work on AI adoption, productivity and genuinely new work, while separating newly circulated papers from settled empirical conclusions. |
| b13 | Humans in the Loop | Generation AI | 2026-04 | MIT Work of the Future draws on research across more than twenty companies and finds that job-quality outcomes depend on implementation choices, including whether firms preserve learning, judgement and worker discretion rather than merely offload tasks. |
| b14 | How the WFH factor complicates the AI jobs story | Centre for British Progress | 2026-06 | Pedro Serôdio for the Centre for British Progress supplies important disconfirming UK evidence: exposed occupations have not shown a sharp aggregate disruption, and remote-work exposure and pre-existing wage trends may explain part of the apparent deterioration in junior hiring. |
| b15 | Is AI Productivity Real? From 15% Workplace Gains to Macro Diffusion | Korea Invest Insights | 2026-06 | This synthesis usefully separates task-level evidence, including customer-support productivity gains, from uneven and still-unproven aggregate productivity effects, while clearly identifying its dependence on Federal Reserve and academic sources. |
| b16 | Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity | METR | 2026-05 | METR reports technical workers' self-assessed productivity gains but openly distinguishes perceived speed from economic value and notes substantial response variation, making it a useful example of evidence that should not be mistaken for measured firm output or wage effects. |
| b17 | Hiring Without Apprenticeship | Workforce Training Executive Intelligence | 2026-03 | The Intelligence Council develops the thesis that AI may shrink junior hiring mainly through lower recruitment rather than layoffs, but readers should treat its causal claims as a commentary-based interpretation of payroll and labour-force evidence rather than an independent study. |
| b18 | LinkedIn's 2026 Labor Market Report: The Robots Aren't Coming for Your Job (Interest Rates Are) | The Random Recruiter | 2026-02 | This recruiter commentary supplies a competing interpretation of weak entry-level technology hiring, attributing it primarily to interest rates, post-pandemic over-hiring and rising computer-science graduate supply rather than AI-specific substitution. |
Tech Industry & Practitioner
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| p1 | Highlights from the 10th DORA report | Google Cloud Blog, DORA | 2024-10 | DORA's 2024 annual survey of technical professionals reports widespread daily AI use and models associations between AI adoption, individual productivity factors, software delivery throughput and stability, making its distinction between local gains and delivery outcomes central. |
| p2 | DORA 2025 State of AI-assisted Software Development Report | DORA | 2025 | The official DORA report page presents the 2025 finding that AI acts as an organisational amplifier rather than an automatic delivery-performance improvement. |
| p3 | DORA 2025 State of AI-assisted Software Development Report | Google Research | 2025 | Google Research identifies the named authors and documents the underlying design, nearly 5,000 global technology-professional survey responses plus more than 100 hours of qualitative data, which clarifies that this is practitioner evidence rather than labour-market administration data. |
| p4 | From adoption to impact: Putting the DORA AI Capabilities Model to work | Google Cloud Blog, DORA | 2025-12 | Nathen Harvey and Allison Park report DORA's cluster analysis and seven capability model, framing AI effects as contingent on organisational conditions and value-stream constraints. |
| p5 | A framework for adopting Gemini Code Assist and measuring its impact | Google Cloud Blog | 2025-03 | DORA lead Nathen Harvey and product manager S. Bogdan explicitly separate adoption and trust metrics from acceleration and business-impact measures, a useful methodological guardrail against treating accepted suggestions or AI-generated code volume as productivity. |
| p6 | 2024 Developer Survey Insights for AI/ML | Stack Overflow Blog | 2024-07 | Stack Overflow's 2024 developer survey provides role, experience and country splits for AI use, trust and perceived job threat, while remaining a self-selected perception survey rather than an observed employment dataset. |
| p7 | AI | 2025 Stack Overflow Developer Survey | Stack Overflow Developer Survey | 2025 | The 2025 survey supplies a large respondent base and reports that 52% of developers saw a positive productivity effect, while a majority either did not use agents or used simpler AI tools, limiting claims of universal agent-driven work replacement. |
| p8 | Survey reveals AI's impact on the developer experience | GitHub Blog | 2023-09 | GitHub's U.S. large-company developer survey documents high tool use and perceived benefits, but its sampling frame and vendor sponsorship make it evidence of adoption and sentiment rather than independent productivity or employment effects. |
| p9 | Survey: The AI wave continues to grow on software development teams | GitHub Blog | 2024-08 | GitHub's updated enterprise team survey tracks organisational promotion of AI tools and toolchain simplicity, useful for understanding management and workflow conditions that mediate adoption. |
| p10 | Thoughtworks Technology Radar Volume 30 | Thoughtworks | 2024-04 | The April 2024 Radar records frontline consultancy observations on AI-assisted software development teams, treating emerging techniques as adoption signals rather than outcome estimates. |
| p11 | Technology Radar Volume 32 | Thoughtworks | 2025-04 | Thoughtworks' April 2025 Radar captures practical movement towards generative-AI techniques, data management and observability in client work, showing how engineering roles and controls are being redesigned. |
| p12 | Volume 33 | Thoughtworks | 2025-11 | The November 2025 Technology Radar documents AI coding workflows, shared instructions and context supply, stressing that teams need collective knowledge diffusion rather than isolated individual automation. |
| p13 | The role of developer skills in agentic coding | martinfowler.com | 2025-03 | Thoughtworks Distinguished Engineer Birgitta Böckeler draws on hands-on use of Cursor, Windsurf and Cline in existing codebases, arguing that developer judgement and skill remain decisive as agentic tools widen the impact radius of mistakes. |
| p14 | How far can we push AI autonomy in code generation? | martinfowler.com | 2025-08 | Birgitta Böckeler's end-to-end experiment found that agents could build simple applications but made unrequested changes, shifting assumptions and declared success with failing tests as complexity increased, supporting continued human supervision. |
| p15 | Coding Assistants Threaten the Software Supply Chain | martinfowler.com | 2025-05 | Jim Gumbley and Lilly Ryan explain how agentic assistants expand the privileges and attack surface of developer environments, indicating why security review and governance work may grow even when coding tasks are automated. |
| p16 | AI in the Trenches: How Developers Are Rewriting the Software Process | InfoQ | 2026-01 | InfoQ's 2026 practitioner panel identifies a shift from code authorship to orchestration, and cautions that activity metrics can increase without corresponding gains in stability, incident performance or code quality. |
| p17 | AI Coding Assistants Haven't Sped up Delivery Because Coding Was Never the Bottleneck | InfoQ | 2026-03 | InfoQ reports practitioner and telemetry claims that AI can move the constraint to specification, review and verification, a plausible competing explanation for strong individual-output measures but modest end-to-end delivery effects. |
| p18 | LLMs and Agents as Team Enablers | InfoQ | 2024-08 | InfoQ records practitioner experiments where LLM agents gave useful but imperfect codebase guidance and frequently failed autonomous tasks, providing disconfirming evidence against claims of reliable replacement. |
| p19 | Virtual Panel: Increasing Engineering Productivity, Develop Software Fast and in a Sustainable Way | InfoQ | 2025-03 | A panel including Google, Slack and Microsoft practitioners argues that sustainable performance relies on feedback loops, quality and reduced friction, placing AI automation within wider organisational design rather than as a standalone labour-saving intervention. |
| p20 | KubeCon + CloudNativeCon North America 2024 day one: keynotes, sessions, announcements, and more | Cloud Native Computing Foundation | 2024-11 | CNCF reports Lunar's operational case study, where AI handled more than 60% of customer text communications and was said to reduce support-resolution time, but it is a single firm example and not evidence of net employment loss. |
| p21 | Emerging trends in the cloud native ecosystem | Cloud Native Computing Foundation | 2024-11 | CNCF describes internal developer platforms as a response to tooling friction and cites a single insurance implementation with faster onboarding and deployments, illustrating complementary investment in platform roles and infrastructure. |
| p22 | Turning legacy to leverage: building developer platforms in brownfield environments | Cloud Native Computing Foundation | 2024-09 | CNCF's practitioner guidance stresses that successful platforms require ongoing user research, trust and support for live systems, showing why AI-era productivity depends on complementary organisational work. |
| p23 | Exploring Generative AI | martinfowler.com | 2025-07 | Birgitta Böckeler's curated Thoughtworks series provides a dated record of hands-on experiments from 2023 through 2026, including codebase onboarding, agent autonomy, context engineering and local coding models. |