Potential LLM task exposure
Eloundou et al., GPTs are GPTs, revised August 2023
The authors rated potential exposure of US occupational tasks to LLMs and complementary software. They estimated that about 80% of workers could have at least 10% of their tasks affected. These are capability estimates, not observed time savings or adoption forecasts.
Findings and scope
Human expertise and GPT-4 classifications informed the exposure rubric.
The estimates distinguish access to a model from additional software built around it.
Exposure does not establish which decisions may be delegated or what a deployed workflow will achieve.
How we use this: identify the records, tools and human decisions a proposed workflow needs, then test it. Exposure ratings do not replace those checks.
Eloundou, Manning, Mishkin and Rock, GPTs are GPTs, revised 21 August 2023Model capability and observed Claude use
Anthropic labor market study, 5 March 2026
Anthropic's U.S. study compares theoretical LLM capability with work-related Claude usage using O*NET. In Computer and Math occupations, estimated theoretical task coverage was 94%, versus 33% observed coverage. These are category-specific estimates, not the share of all occupations or all AI activity.
Findings and scope
Observed exposure weights tasks by time spent and gives automated usage more weight than assistance.
Possible limits include model capabilities, legal restrictions, required software and human verification.
Zero coverage reflects the study's usage threshold; it does not show that a worker is unaffected by AI.
The study does not establish workflow design as the cause of the gap.
How we use this: review workflow responsibilities, constraints and readiness before rollout. Simulations can help test that design; the study does not measure Nuvepro's impact.
Anthropic, Labor market impacts of AI: A new measure and early evidence, 5 March 2026Worker attitudes and access to AI learning
Guild: AI training article and June 2023 member survey
Guild discusses worker attitudes and access to AI learning. Its article describes a June 2023 survey of 355 members and learners, alongside separate programme-application figures. These populations do not establish readiness across the national workforce.
Findings and scope
The member survey examined attitudes to AI and differences between respondent groups.
Application growth measures interest in the programmes offered, rather than skills demonstrated.
Neither a survey response nor an application establishes capability to perform a specific job.
How we use this: ask what people already know and what work they need to perform. Match simulations and support to those needs, then review what they demonstrate.
Guild, AI Training for Employees: How to Get & Stay Ahead of the Curve; June 2023 member surveyEnterprise learning priorities
Everest Group: Getting the L&D Pulse, 2026
Everest Group's public report abstract describes a global survey of learning, HR, technology and business leaders. It examines how enterprise learning is shifting toward capabilities connected to evolving roles, business priorities and work. The full report is available to members. Detailed survey findings are not available in the public material reviewed here.
Findings and scope
The report examines enterprise skill priorities and pain points across learning journeys.
Its stated scope includes L&D budgets, outsourcing objectives and learning delivery models across industries and geographies.
The abstract describes learning becoming more embedded in workflows, supported by digital platforms and data-driven insights.
How we use this: connect preparation to the work people need to perform. Agree the simulation, business environment and evidence for the responsibilities assessed. This is our application of the research, not an Everest endorsement or measured Nuvepro result.
Everest Group, Getting the L&D Pulse: the State of Enterprise Learning in the Age of AI, public abstract, 2026Modelled economic effects of generative AI
Cognizant and Oxford Economics, January 2024
The report models possible US economic effects across roughly 18,000 tasks and 1,000 occupations. Its scenarios vary business adoption and workforce adjustment. They project outcomes through 2032; they do not measure completed enterprise deployments.
Findings and scope
The model distinguishes task exposure from adoption and labour-market adjustment.
Its scenarios depend on assumptions about adoption and workers moving into other work.
The estimates do not isolate the causal effect of a training programme or validate Nuvepro's classification method.
How we use this: examine the actual responsibilities changing in your workflow. Agree preparation and operating checks locally; a national GDP scenario cannot predict a project's outcome.
Cognizant and Oxford Economics, New Work, New World with Generative AI, January 2024AI adoption and employment transitions
Goldman Sachs Research, 13 August 2025
Goldman Sachs Research estimates possible job displacement under wider AI adoption. Its baseline is 6-7% of the US workforce, with outcomes varying by assumptions. Its discussion of temporary unemployment draws on historical technology transitions, rather than a guaranteed recovery period for AI.
Findings and scope
The analysis examines more than 800 occupations.
Alternative assumptions produce different displacement estimates.
Historical recovery patterns do not guarantee an individual worker's transition or establish the effect of reskilling.
How we use this: define the responsibilities people would take on and assess relevant practice. Preparation can inform assignment decisions; it does not guarantee employment or an economy-wide recovery.
Goldman Sachs Research, How Will AI Affect the Global Workforce?, 13 August 2025Employer expectations for jobs through 2030
World Economic Forum, Future of Jobs Report 2025
The report extrapolates employer expectations and employment data to project job creation and displacement through 2030. It considers technological change, demographics, the green transition and other macrotrends. The totals cover several macrotrends, not AI alone.
Findings and scope
The report covers employer perspectives across 22 industry clusters and 55 economies.
Its projections combine expected growth and decline; they are not observed job counts.
How we use this: use these forecasts to ask which responsibilities may change. Agree work, simulations and evidence with your enterprise; the forecast does not guarantee job creation or financial performance.
World Economic Forum, The Future of Jobs Report 2025, January 2025Productivity and typical-worker compensation
Economic Policy Institute tracker, updated 14 September 2026
EPI compares US productivity with compensation for a typical worker and documents their divergence since the late 1970s. Its methodology specifies the populations, benefits and price adjustments used. This historical series is not a study of AI deployment or the effect of workforce preparation.
Findings and scope
The productivity measure covers the economy; the pay measure starts with private-sector production and nonsupervisory workers.
EPI attributes the divergence in its adjusted series to wage inequality and the shift of income toward capital.
Those measurement choices matter when comparing this tracker with other pay and productivity series.
How we use this: distinguish operating improvements from how gains are shared. Practical preparation does not determine compensation; impact and compensation decisions need separate evidence and ownership.
Economic Policy Institute, The Productivity-Pay Gap, updated 14 September 2026Enterprise AI adoption and workflow constraints
Andreessen Horowitz enterprise AI analysis, 8 April 2026
Kimberly Tan reports that 29% of the Fortune 500 were live, paying customers of a leading AI startup in a16z's analysis. The criteria included a signed top-down contract, converted pilot and live product. This is narrower than a census of all enterprise AI use.
Findings and scope
The analysis combines startup-provided information, public data and a16z conversations.
The article argues that partial task automation may leave bottlenecks, so capability gains need not produce proportional economic value.
It identifies verifiable outputs and workflow constraints as relevant to adoption; it does not establish classification as sufficient for every role.
How we use this: use task analysis to form hypotheses about bottlenecks. Build and test the agreed workflow against its baseline before claiming an operating improvement.
Kimberly Tan, Where Enterprises are Actually Adopting AI, Andreessen Horowitz, 8 April 2026Worker voice in manufacturing
MIT Sloan: worker voice in manufacturing, 4 November 2025
MIT Sloan reports research linking employee input with productivity and pay in US manufacturing. Researchers matched a 2015 survey of 30,000 establishments with administrative earnings and business records. This concerns worker voice in manufacturing, not a test of AI implementation or Nuvepro preparation.
Findings and scope
Productivity was measured as revenue generated per worker, not AI output or agent task completion.
Researchers compared similar establishments using survey and administrative records; this was not a randomized training intervention.
How we use this: involve the people doing the work when defining responsibilities and simulations. Your team builds or configures the agreed work; measure readiness and operating impact separately. Participation or preparation does not guarantee productivity gains.
MIT Sloan, For manufacturers, listening to workers pays off in productivity, 4 November 2025Why Task-Level Analysis
How research informs the approach
Eloundou et al. distinguish model capability alone from capability with complementary software
Your team builds or configures the agreed solution. Nuvepro scopes the simulation, environment and preparation; integration and operating responsibilities are agreed separately
Guild examines worker attitudes and access to learning
Practical capability needs different evidence. Nuvepro scopes hands-on projects for agreed responsibilities; assessment records what participants demonstrate and owners decide what further preparation is needed
a16z argues that partial automation can leave bottlenecks in a workflow
Task analysis helps form hypotheses about bottlenecks. Agree the baseline and checks, then evaluate the changed work before claiming operating impact
Use research to ask better questions about your work.
Bring one workflow, the people and agents involved, and the change your enterprise needs. Agree a bounded project, its business environment and the evidence required. Published study findings do not predict your operating results.