{
  "name": "Figures on the time AI frees, the time put to use, its cost and its risks, with sources",
  "description": "The figures behind the payback tool, one row per figure. Each row gives the figure, its unit, who or what it covers, what kind of number it is (measured, self-reported, calculated, reported), the line of the result it feeds, the source with a quote, and its caveat.",
  "url": "https://www.ai-first.company/payback/",
  "version": "2026-10-13",
  "licence": "Data: CC BY 4.0, Paweł Szczęsny, ai-first.company",
  "licenceUrl": "https://creativecommons.org/licenses/by/4.0/",
  "count": 88,
  "rows": [
    {
      "id": "acemoglu-share-exposed",
      "url": "https://www.ai-first.company/payback/#source-acemoglu-share-exposed",
      "claim": "About 20% of US work tasks, weighted by wages, are exposed to AI in Acemoglu's calculation.",
      "value": 20,
      "low": null,
      "high": null,
      "unit": "% of tasks (wage-bill weighted)",
      "population": "US occupations and tasks (Eloundou et al. 2023 exposure scores), weighted by 2019-2022 wage bills",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "This calculation implies that 20% of US labor tasks are exposed to AI.",
      "caveat": "Exposure means AI could do or help with the task; it says nothing about whether anyone uses it. It is a US economy-wide share, not a share for one company.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "acemoglu-share-worth-automating",
      "url": "https://www.ai-first.company/payback/#source-acemoglu-share-worth-automating",
      "claim": "Acemoglu assumes that 23% of exposed tasks are worth automating within ten years, a figure he borrows from a study of computer-vision tasks.",
      "value": 23,
      "low": null,
      "high": null,
      "unit": "% of exposed tasks",
      "population": "Extrapolated from US computer-vision tasks to all AI-exposed tasks",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "Svanberg et al.’s base estimates imply that among computer vision-exposed tasks, 23% can be feasibly (and profitably) automated within 10 years.",
      "caveat": "Acemoglu calls the extrapolation from computer vision to all AI 'not an innocuous step' and says he has no other source for it.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "svanberg-23-percent-cost-effective",
      "url": "https://www.ai-first.company/payback/#source-svanberg-23-percent-cost-effective",
      "claim": "A cost model of computer-vision tasks found that only 23% of the exposed wages were worth automating at the firm level, because of large up-front costs.",
      "value": 23,
      "low": null,
      "high": null,
      "unit": "% of exposed compensation",
      "population": "US non-farm business tasks that computer vision could do",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Svanberg, M., Li, W., Fleming, M., Goehring, B. and Thompson, N. (2024). Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision? MIT FutureTech working paper, January 2024 (working paper, not peer-reviewed).",
      "source_url": "https://futuretech-site.s3.us-east-2.amazonaws.com/2024-01-18+Beyond_AI_Exposure.pdf",
      "quote": "We find that only 23% of worker compensation “exposed” to AI computer vision would be cost-effective for firms to automate because of the large upfront costs of AI systems.",
      "caveat": "About custom-built vision systems in 2024, not about off-the-shelf chat assistants, whose up-front cost is far lower.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "acemoglu-share-impacted",
      "url": "https://www.ai-first.company/payback/#source-acemoglu-share-impacted",
      "claim": "Multiplying the two shares, Acemoglu puts the part of the economy that AI will actually change within ten years at 4.6%.",
      "value": 4.6,
      "low": null,
      "high": null,
      "unit": "% of tasks",
      "population": "US economy",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "I arrive at the GDP share of tasks impacted by AI within the next 10 years as 0.23 × 0.200 = 4.6% of all tasks (or occupations).",
      "caveat": "A ten-year, economy-wide figure. A single firm that picks its tasks well can be far above it, and a firm that picks none is at zero.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "acemoglu-labour-cost-saving",
      "url": "https://www.ai-first.company/payback/#source-acemoglu-labour-cost-saving",
      "claim": "On tasks where AI is used, Acemoglu takes the labour-cost saving to be 27%, the average of two experiments.",
      "value": 27,
      "low": null,
      "high": null,
      "unit": "% labour cost on affected tasks",
      "population": "Average of Noy and Zhang (2023, writing tasks) and Brynjolfsson, Li and Raymond (2023, customer support)",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "Under these assumptions, the average labor cost savings are 27% (= 0.27).",
      "caveat": "He treats faster work as cost saved and ignores quality changes; adding the GitHub Copilot experiment (Peng et al.) raises it to 36%.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "acemoglu-overall-cost-saving",
      "url": "https://www.ai-first.company/payback/#source-acemoglu-overall-cost-saving",
      "claim": "Because labour is only part of the cost of a task, Acemoglu converts the 27% labour saving into a 14.4% saving on total cost.",
      "value": 14.4,
      "low": null,
      "high": null,
      "unit": "% total cost on affected tasks",
      "population": "US industries, labour shares from the Bureau of Economic Analysis",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "This gives an average (AI exposure-adjusted) labor share of 0.535, and thus the average (overall) cost savings from AI are about 0.27 × 0.535 = 0.144.",
      "caveat": "",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "acemoglu-tfp-10-years",
      "url": "https://www.ai-first.company/payback/#source-acemoglu-tfp-10-years",
      "claim": "Acemoglu's headline result: AI raises US productivity by no more than about 0.66% in total over ten years.",
      "value": 0.66,
      "low": null,
      "high": null,
      "unit": "% TFP, total over 10 years",
      "population": "US economy, total factor productivity",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "TFP gains over the next 10 years = 0.046 × 0.144 ≃ 0.0066.",
      "caveat": "A total over ten years (about 0.064% a year), and an upper bound in his own words. It is a macro figure and must not be shown as any one company's return.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "acemoglu-hard-tasks-adjustment",
      "url": "https://www.ai-first.company/payback/#source-acemoglu-hard-tasks-adjustment",
      "claim": "For 'hard-to-learn' tasks, where there is no clear right answer to learn from, Acemoglu assumes a 7% labour saving in place of 27%, which lowers his ten-year productivity estimate to about 0.53%.",
      "value": 0.53,
      "low": null,
      "high": null,
      "unit": "% TFP, total over 10 years",
      "population": "US economy; 72.6% of exposed tasks classed as easy, the rest as hard",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "TFP gains over the next 10 years = 0.033 × 0.144 + (0.046 − 0.033) × 0.037 ≃ 0.0053.",
      "caveat": "The 7% is his own assumption ('half of the cost savings' of the customer-support study), and he calls the easy/hard split speculative.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "acemoglu-hard-task-saving",
      "url": "https://www.ai-first.company/payback/",
      "claim": "Acemoglu's assumed saving on judgement-heavy tasks is 7% of labour cost, about a quarter of the saving on well-defined tasks.",
      "value": 7,
      "low": null,
      "high": null,
      "unit": "% labour cost on hard tasks",
      "population": "Assumption, anchored on the 14% gain in the customer-support study",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "This motivates my choice of half of the cost savings of their study, 7% (which is also about a quarter of the baseline 27% cost reduction estimate I used in the previous subsection).",
      "caveat": "An assumption, not a measurement: 'it is impossible to know what their cost savings will be'.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "acemoglu-gdp-10-years",
      "url": "https://www.ai-first.company/payback/#source-acemoglu-gdp-10-years",
      "claim": "In GDP terms Acemoglu's estimate is roughly 0.9% to 1.2% in total over ten years.",
      "value": null,
      "low": 0.93,
      "high": 1.16,
      "unit": "% GDP, total over 10 years",
      "population": "US economy",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time freed",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "GDP is also estimated to grow by 0.93% − 1.16% over the next 10 years.",
      "caveat": "With a larger investment response he gives an upper bound of 1.4% to 1.56%.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "acemoglu-caveat-adjustment-costs",
      "url": "https://www.ai-first.company/payback/#source-acemoglu-caveat-adjustment-costs",
      "claim": "Acemoglu notes that his figures ignore the cost and delay of reorganizing work, so even his small estimate may be too high for the first ten years.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "US economy",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time put to use",
      "source": "Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text).",
      "source_url": "https://www.nber.org/system/files/working_papers/w32487/w32487.pdf",
      "quote": "If so, the 14.4% overall cost reductions may be a significant overestimate for the next 10 years.",
      "caveat": "Direction only; he does not put a number on the adjustment cost.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "goldman-sachs-7-percent",
      "url": "https://www.ai-first.company/payback/#source-goldman-sachs-7-percent",
      "claim": "Goldman Sachs Research forecast that generative AI could raise global GDP by 7% and productivity growth by 1.5 percentage points a year over ten years.",
      "value": 7,
      "low": null,
      "high": null,
      "unit": "% global GDP over 10 years",
      "population": "Global economy; task content of over 900 occupations",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Background",
      "source": "Goldman Sachs (5 April 2023). Generative AI could raise global GDP by 7%. Goldman Sachs Research article summarising Briggs and Kodnani (investment-bank research, not peer-reviewed).",
      "source_url": "https://www.goldmansachs.com/insights/articles/generative-ai-could-raise-global-gdp-by-7-percent",
      "quote": "they could drive a 7% (or almost $7 trillion) increase in global GDP and lift productivity growth by 1.5 percentage points over a 10-year period.",
      "caveat": "The underlying report is not public; this is the bank's own summary page.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "mckinsey-genai-productivity",
      "url": "https://www.ai-first.company/payback/#source-mckinsey-genai-productivity",
      "claim": "McKinsey Global Institute estimated that generative AI could add 0.1 to 0.6 percentage points a year to labour productivity growth through 2040.",
      "value": null,
      "low": 0.1,
      "high": 0.6,
      "unit": "percentage points of labour productivity growth a year",
      "population": "Global economy; 63 use cases and 2,100 work activities",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Background",
      "source": "McKinsey Global Institute (14 June 2023). The economic potential of generative AI: The next productivity frontier (consultancy report, not peer-reviewed).",
      "source_url": "https://web.archive.org/web/2024/https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier",
      "quote": "Generative AI could enable labor productivity growth of 0.1 to 0.6 percent annually through 2040, depending on the rate of technology adoption and redeployment of worker time into other activities.",
      "caveat": "McKinsey's own range depends on 'redeployment of worker time into other activities', which is the gain-kept line. Read from an archived copy; the live page blocked automated access.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "mckinsey-genai-value",
      "url": "https://www.ai-first.company/payback/#source-mckinsey-genai-value",
      "claim": "The same McKinsey report put the value of 63 generative-AI use cases at $2.6 trillion to $4.4 trillion a year.",
      "value": null,
      "low": 2.6,
      "high": 4.4,
      "unit": "trillion USD a year",
      "population": "Global economy; 63 use cases",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Background",
      "source": "McKinsey Global Institute (14 June 2023). The economic potential of generative AI: The next productivity frontier.",
      "source_url": "https://web.archive.org/web/2024/https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier",
      "quote": "generative AI could add the equivalent of $2.6 trillion to $4.4 trillion annually across the 63 use cases we analyzed",
      "caveat": "A potential, built bottom-up from use cases the firm chose; McKinsey sells AI consulting.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "aghion-bunel-068",
      "url": "https://www.ai-first.company/payback/#source-aghion-bunel-068",
      "claim": "Aghion and Bunel reran Acemoglu's formula with their own reading of the evidence and got 0.68 percentage points of extra productivity growth a year, about ten times his figure.",
      "value": 0.68,
      "low": 0.07,
      "high": 1.24,
      "unit": "percentage points of TFP growth a year",
      "population": "US and European economies",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Background",
      "source": "Aghion, P. and Bunel, S. (June 2024). AI and Growth: Where Do We Stand? Note published by the Federal Reserve Bank of San Francisco (not peer-reviewed).",
      "source_url": "https://www.frbsf.org/wp-content/uploads/AI-and-Growth-Aghion-Bunel.pdf",
      "quote": "we obtain a median estimate of 0.68pp additional annual total factor productivity (TFP) growth.",
      "caveat": "Their range runs from 0.07 to 1.24 points a year: the same formula gives a 17-fold spread depending on the inputs chosen.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "oecd-miracle-or-myth",
      "url": "https://www.ai-first.company/payback/#source-oecd-miracle-or-myth",
      "claim": "An OECD paper estimates that AI adds 0.25 to 0.6 percentage points a year to productivity growth over ten years.",
      "value": null,
      "low": 0.25,
      "high": 0.6,
      "unit": "percentage points of TFP growth a year",
      "population": "US calibration, extended to other G7 economies",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Background",
      "source": "Filippucci, F., Gal, P. and Schief, M. (2024). Miracle or Myth? Assessing the macroeconomic productivity gains from Artificial Intelligence. OECD Artificial Intelligence Papers No. 29, November 2024 (working paper).",
      "source_url": "https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/11/miracle-or-myth-assessing-the-macroeconomic-productivity-gains-from-artificial-intelligence_fde2a597/b524a072-en.pdf",
      "quote": "Its main estimates for annual aggregate total-factor productivity growth due to AI range between 0.25-0.6 percentage points (0.4-0.9 pp. for labour productivity).",
      "caveat": "",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "imf-40-percent-exposed",
      "url": "https://www.ai-first.company/payback/#source-imf-40-percent-exposed",
      "claim": "The IMF's headline figure is about exposure, not growth: almost 40% of jobs worldwide, and about 60% in advanced economies, are exposed to AI.",
      "value": 40,
      "low": null,
      "high": null,
      "unit": "% of global employment exposed",
      "population": "Global employment",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Background",
      "source": "Georgieva, K. (14 January 2024). AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity. IMF Blog, summarising IMF Staff Discussion Note SDN/2024/001 (Cazzaniga et al.).",
      "source_url": "https://www.imf.org/en/Blogs/Articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity",
      "quote": "almost 40 percent of global employment is exposed to AI.",
      "caveat": "Exposure only. The IMF's 2025 working paper 'The Global Impact of AI: Mind the Gap' gives no single headline GDP number on its summary page.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "noy-zhang-writing-40",
      "url": "https://www.ai-first.company/evidence/studies/chatgpt-writing-experiment/",
      "claim": "In an experiment on short professional writing tasks, ChatGPT cut the time taken by 40% and raised quality by 18%.",
      "value": 40,
      "low": null,
      "high": null,
      "unit": "% less time on the task",
      "population": "453 college-educated professionals, online, one-off paid writing tasks",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Time freed",
      "source": "Noy, S. and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192 (peer-reviewed).",
      "source_url": "https://doi.org/10.1126/science.adh2586",
      "quote": "Our results show that ChatGPT substantially raised productivity: The average time taken decreased by 40% and output quality rose by 18%.",
      "caveat": "A best case: a task chosen because AI does it well, done once, outside the real job. The quote is the one stored on the site's study page; the journal page could not be re-opened in this session.",
      "study": "https://www.ai-first.company/evidence/studies/chatgpt-writing-experiment/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "brynjolfsson-support-15",
      "url": "https://www.ai-first.company/evidence/studies/generative-ai-at-work/",
      "claim": "In a large customer-support operation, agents given an AI assistant resolved 15% more issues per hour on average.",
      "value": 15,
      "low": null,
      "high": null,
      "unit": "% more issues resolved per hour",
      "population": "5,172 customer-support agents, staggered rollout at one company",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Time freed",
      "source": "Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889-942 (peer-reviewed). Abstract read on arXiv:2304.11771.",
      "source_url": "https://arxiv.org/abs/2304.11771",
      "quote": "Access to AI assistance increases worker productivity, as measured by issues resolved per hour, by 15% on average",
      "caveat": "Gains were concentrated in less experienced agents; the most experienced gained little. The study measures output per hour, not profit.",
      "study": "https://www.ai-first.company/evidence/studies/generative-ai-at-work/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "jagged-frontier-inside-outside",
      "url": "https://www.ai-first.company/evidence/studies/jagged-frontier/",
      "claim": "Consultants using GPT-4 did 12.2% more tasks 25.1% faster on work the AI does well, and were 19 percentage points less likely to be right on a task the AI could not do well.",
      "value": 12.2,
      "low": null,
      "high": null,
      "unit": "% more tasks completed (inside the frontier)",
      "population": "758 consultants at Boston Consulting Group",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Time freed",
      "source": "Dell’Acqua, F. et al. (2026). Navigating the Jagged Technological Frontier. Organization Science, 37(2), 403-423 (peer-reviewed).",
      "source_url": "https://doi.org/10.1287/orsc.2025.21838",
      "quote": "AI assistance improves performance for some tasks but worsens it for others, even within the same knowledge workflow and with a seemingly similar level of difficulty.",
      "caveat": "The quote is the one stored on the site's study page and does not contain the figures; the journal and SSRN pages blocked access in this session, so the figures rest on the site's earlier check.",
      "study": "https://www.ai-first.company/evidence/studies/jagged-frontier/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "cui-developers-26",
      "url": "https://www.ai-first.company/payback/#source-cui-developers-26",
      "claim": "Across three company-run randomised trials, developers given an AI coding assistant completed about 26% more tasks.",
      "value": 26.08,
      "low": null,
      "high": null,
      "unit": "% more completed tasks",
      "population": "4,867 software developers at Microsoft, Accenture and an unnamed Fortune 100 company",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Time freed",
      "source": "Cui, K. Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S. and Salz, T. (February 2025 version). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Working paper (version hosted by MIT Economics; not peer-reviewed at that version).",
      "source_url": "https://economics.mit.edu/sites/default/files/inline-files/draft_copilot_experiments.pdf",
      "quote": "our analysis reveals a 26.08% increase (SE: 10.3%) in completed tasks among developers using the AI tool.",
      "caveat": "Wide uncertainty (standard error 10.3 points), and 'completed tasks' counts pull requests, not their value. Evidence that cuts against the METR slowdown.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "metr-slowdown-19",
      "url": "https://www.ai-first.company/payback/#source-metr-slowdown-19",
      "claim": "In a randomised trial, experienced open-source developers took 19% longer on their own real tasks when allowed to use AI tools.",
      "value": -19,
      "low": null,
      "high": null,
      "unit": "% change in speed (negative = slower)",
      "population": "16 experienced developers, 246 tasks in mature repositories they knew well, February-June 2025",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Time freed",
      "source": "Becker, J., Rush, N., Barnes, B. and Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089v2 (preprint, not peer-reviewed).",
      "source_url": "https://arxiv.org/pdf/2507.09089",
      "quote": "Surprisingly, we find that allowing AI actually increases completion time by 19%—AI tooling slowed developers down.",
      "caveat": "Sixteen people, expert in their own code, early-2025 tools. METR lists what the study does not show, including that AI fails to speed up most developers.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "metr-2026-update",
      "url": "https://www.ai-first.company/payback/#source-metr-2026-update",
      "claim": "METR's follow-up found smaller slowdowns and says its own new data are unreliable, because developers no longer want to do tasks without AI.",
      "value": null,
      "low": -18,
      "high": -4,
      "unit": "% (METR's sign convention as published; both confidence intervals include zero)",
      "population": "57 developers, 143 repositories, 800+ tasks, from August 2025",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Time freed",
      "source": "METR (24 February 2026). We are Changing our Developer Productivity Experiment Design. METR blog.",
      "source_url": "https://metr.org/blog/2026-02-24-uplift-update/",
      "quote": "For the subset of the original developers who participated in the later study, we now estimate a speedup of -18% with a confidence interval between -38% and +9%. Among newly-recruited developers the estimated speedup is -4%, with a confidence interval between -15% and +9%.",
      "caveat": "METR says the estimate is 'likely a bad proxy' and probably a lower bound, and that developers are 'likely' more sped up in early 2026 than in early 2025. Both intervals include zero.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "humlum-time-saved-2-8",
      "url": "https://www.ai-first.company/evidence/studies/chatbots-and-the-labour-market-denmark/",
      "claim": "Across a whole labour market, chatbot users reported saving 2.8% of their working hours on average.",
      "value": 2.8,
      "low": null,
      "high": null,
      "unit": "% of work hours",
      "population": "About 25,000 workers at 7,000 workplaces in 11 exposed occupations, Denmark, surveys in late 2023 and 2024",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time freed",
      "source": "Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56, May 2025 (working paper).",
      "source_url": "https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-3.pdf",
      "quote": "users report average time savings of 2.8% of work hours.",
      "caveat": "From the May 2025 version; the paper was revised as NBER Working Paper 33777 under a new title, and the figure is self-reported.",
      "study": "https://www.ai-first.company/evidence/studies/chatbots-and-the-labour-market-denmark/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "stlouisfed-time-saved",
      "url": "https://www.ai-first.company/payback/#source-stlouisfed-time-saved",
      "claim": "In a US national survey, generative-AI users said it saved them 5.4% of their work hours; across all workers that is 1.4%.",
      "value": 1.4,
      "low": null,
      "high": null,
      "unit": "% of total work hours (all workers); 5.4% among users",
      "population": "Nationally representative US survey of working-age adults, November 2024",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time freed",
      "source": "Bick, A., Blandin, A. and Deming, D. (27 February 2025). The Impact of Generative AI on Work Productivity. Federal Reserve Bank of St. Louis, On the Economy blog.",
      "source_url": "https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity",
      "quote": "When we factor in all workers, including nonusers, workers saved 1.4% of total hours because of generative AI.",
      "caveat": "Self-reported; the authors write that how these gains translate into measured productivity 'remains uncertain'.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "uk-government-copilot-26-minutes",
      "url": "https://www.ai-first.company/evidence/claim-checks/uk-copilot-26-minutes-a-day/",
      "claim": "In the UK government's Copilot trial, users reported saving 26 minutes a day on average.",
      "value": 26,
      "low": null,
      "high": null,
      "unit": "minutes a day per user",
      "population": "20,000 UK civil servants, 30 September to 31 December 2024",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time freed",
      "source": "UK Government Digital Service (2 June 2025). Microsoft 365 Copilot experiment: cross-government findings report.",
      "source_url": "https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html",
      "quote": "Users reported an average of 26 minutes saved per day. However, due to experimental constraints it was not possible to identify how time saved was spent.",
      "caveat": "Self-reported, and the report itself says it could not find out what happened to the time.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "dillon-copilot-email-2-hours",
      "url": "https://www.ai-first.company/evidence/studies/copilot-field-experiment-66-firms/",
      "claim": "In a randomised trial across 66 firms, workers who used an AI assistant in their office software spent two fewer hours a week on email.",
      "value": 2,
      "low": null,
      "high": null,
      "unit": "hours a week on email, among users",
      "population": "7,137 knowledge workers at 66 firms, six-month experiment, measured from software telemetry",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Time freed",
      "source": "Dillon, E. W., Jaffe, S., Immorlica, N. and Stanton, C. T. (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795 (working paper; three authors work at Microsoft Research).",
      "source_url": "https://www.nber.org/system/files/working_papers/w33795/w33795.pdf",
      "quote": "the 80% of treated workers who used this tool spent two fewer hours on email each week and reduced their time working outside of regular hours.",
      "caveat": "Measured, not self-reported, and in a broad rollout: evidence that broad deployment can free real time. The tool is Microsoft's and so are three of the authors.",
      "study": "https://www.ai-first.company/evidence/studies/copilot-field-experiment-66-firms/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "dillon-no-task-shift",
      "url": "https://www.ai-first.company/evidence/studies/copilot-field-experiment-66-firms/",
      "claim": "The same trial found no change in how much work people did or what kind: the time saved did not show up as more output.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "7,137 knowledge workers at 66 firms",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Time put to use",
      "source": "Dillon, E. W., Jaffe, S., Immorlica, N. and Stanton, C. T. (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795.",
      "source_url": "https://www.nber.org/system/files/working_papers/w33795/w33795.pdf",
      "quote": "Apart from these individual time savings, we do not detect shifts in the quantity or composition of workers’ tasks resulting from individual-level AI provision.",
      "caveat": "The tool was given to individuals; the authors' point is that task changes need coordination across people, which the experiment did not provide.",
      "study": "https://www.ai-first.company/evidence/studies/copilot-field-experiment-66-firms/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "humlum-earnings-hours-null",
      "url": "https://www.ai-first.company/evidence/studies/chatbots-and-the-labour-market-denmark/",
      "claim": "Two years after ChatGPT appeared, a Danish study linking surveys to pay records found no effect on earnings or recorded hours, and rules out effects above 2%.",
      "value": 2,
      "low": null,
      "high": null,
      "unit": "% (upper bound on any effect on earnings or hours)",
      "population": "Workers and workplaces in exposed occupations, Denmark; administrative pay and hours records",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Time put to use",
      "source": "Humlum, A. and Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777 (working paper).",
      "source_url": "https://www.nber.org/system/files/working_papers/w33777/w33777.pdf",
      "quote": "we estimate precise null effects on earnings and recorded hours at both the worker and workplace levels, ruling out effects larger than 2% two years after the launch of ChatGPT.",
      "caveat": "Pay and hours are not profit; gains could sit in quality or new tasks. An upper bound, not an estimate of zero.",
      "study": "https://www.ai-first.company/evidence/studies/chatbots-and-the-labour-market-denmark/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "humlum-pass-through-3-7",
      "url": "https://www.ai-first.company/evidence/studies/chatbots-and-the-labour-market-denmark/",
      "claim": "In the Danish study, only 3% to 7% of workers' reported time savings showed up as higher earnings.",
      "value": null,
      "low": 3,
      "high": 7,
      "unit": "% of time savings passed to earnings",
      "population": "About 25,000 workers in 11 occupations, Denmark",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Time put to use",
      "source": "Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56 (May 2025 version).",
      "source_url": "https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-3.pdf",
      "quote": "we estimate that only 3–7% of workers’ productivity gains are passed through to higher earnings, with greater elasticity at firms that encourage chatbot use.",
      "caveat": "This is pass-through to workers' pay, not the share the company keeps; the authors ask readers to treat the exact rates 'with some caution'.",
      "study": "https://www.ai-first.company/evidence/studies/chatbots-and-the-labour-market-denmark/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "humlum-what-freed-time-went-to",
      "url": "https://www.ai-first.company/payback/#source-humlum-what-freed-time-went-to",
      "claim": "Asked what they did with time saved by chatbots, 80% of Danish users said other job tasks, fewer than 10% said breaks or leisure, and 25% said more of the same task.",
      "value": 80,
      "low": null,
      "high": null,
      "unit": "% of users who reallocated saved time to other job tasks",
      "population": "Chatbot users among about 25,000 surveyed workers, Denmark",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time put to use",
      "source": "Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56 (May 2025 version).",
      "source_url": "https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-3.pdf",
      "quote": "the vast majority (80%) reallocate this time to other job tasks, while fewer than 10% use it for additional breaks or leisure. In addition, 25% spend more time on the same tasks they initially saved time on, especially when employers actively encourage chatbot use.",
      "caveat": "Self-reported and multiple answers allowed; 'other job tasks' says nothing about whether those tasks were worth more.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "humlum-encouragement-doubles-adoption",
      "url": "https://www.ai-first.company/payback/#source-humlum-encouragement-doubles-adoption",
      "claim": "Where Danish employers encouraged chatbot use, the share of workers using chatbots nearly doubled, from 47% to 83%, and reported time savings were larger.",
      "value": null,
      "low": 47,
      "high": 83,
      "unit": "% of workers using chatbots (without vs with employer encouragement)",
      "population": "About 25,000 workers at 7,000 workplaces, Denmark",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time freed",
      "source": "Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56 (May 2025 version).",
      "source_url": "https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-3.pdf",
      "quote": "These initiatives greatly boost adoption, nearly doubling take-up rates for the typical worker from 47% to 83%.",
      "caveat": "Cuts against a simple 'mandates are a mistake' toggle: encouragement and training raise use. It is encouragement, not a usage quota, and the comparison is not randomised.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "firm-data-on-ai-no-impact",
      "url": "https://www.ai-first.company/evidence/studies/firm-data-on-ai-four-countries/",
      "claim": "In a survey of nearly 6,000 executives in four countries, 89% said AI had had no effect on their firm's labour productivity over the past three years.",
      "value": 89,
      "low": null,
      "high": null,
      "unit": "% of firms reporting no productivity impact",
      "population": "Nearly 6,000 senior executives (mostly CFOs and CEOs) at US, UK, German and Australian firms",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time put to use",
      "source": "Yotzov, I., Barrero, J. M., Bloom, N., Bunn, P., Davis, S. J. et al. (2026). Firm Data on AI. NBER Working Paper 34836, February 2026, revised March 2026 (working paper).",
      "source_url": "https://www.nber.org/system/files/working_papers/w34836/w34836.pdf",
      "quote": "More than 90% report no impact of AI on own-firm employment over the past three years, and 89% report no impact on labor productivity (sales per employee).",
      "caveat": "Executives' own judgement, not measured productivity; 69% of the same firms use AI.",
      "study": "https://www.ai-first.company/evidence/studies/firm-data-on-ai-four-countries/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "firm-data-on-ai-expected-1-4",
      "url": "https://www.ai-first.company/evidence/studies/firm-data-on-ai-four-countries/",
      "claim": "The same executives expect AI to raise productivity at their firms by 1.4% on average over the next three years.",
      "value": 1.4,
      "low": null,
      "high": null,
      "unit": "% productivity over 3 years (expected)",
      "population": "Nearly 6,000 senior executives in four countries",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time put to use",
      "source": "Yotzov, I., Barrero, J. M., Bloom, N. et al. (2026). Firm Data on AI. NBER Working Paper 34836.",
      "source_url": "https://www.nber.org/system/files/working_papers/w34836/w34836.pdf",
      "quote": "predicting that AI will boost productivity at their firms by an average of 1.4%, raise output 0.8%, and cut employment 0.7%.",
      "caveat": "A forecast by the people responsible for the spending.",
      "study": "https://www.ai-first.company/evidence/studies/firm-data-on-ai-four-countries/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "dbt-no-evidence-of-productivity",
      "url": "https://www.ai-first.company/evidence/cases/uk-department-for-business-and-trade-copilot-pilot/",
      "claim": "A UK government department that measured its Copilot pilot found time savings on some tasks and no evidence that they led to higher productivity.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "1,000 licence holders at the UK Department for Business and Trade, October-December 2024",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Time put to use",
      "source": "UK Department for Business and Trade (August 2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade.",
      "source_url": "https://assets.publishing.service.gov.uk/media/68adbe409e1cebdd2c96a19d/dbt-microsoft-365-copilot-evaluation.pdf",
      "quote": "The evaluation did not find evidence that time savings have led to improved productivity, and control group participants had not observed productivity improvements from colleagues taking part in the M365 Copilot pilot.",
      "caveat": "The report adds that measuring productivity 'was not a key aim of the evaluation'.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "dta-40-percent-reallocated",
      "url": "https://www.ai-first.company/payback/#source-dta-40-percent-reallocated",
      "claim": "In the Australian government's Copilot trial, about 40% of participants said they were able to move saved time to higher-value work.",
      "value": 40,
      "low": null,
      "high": null,
      "unit": "% of participants reporting they moved time to higher-value work",
      "population": "Over 2,000 survey and consultation participants out of nearly 5,765 licence holders in 56 agencies, January-June 2024",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time put to use",
      "source": "Digital Transformation Agency (October 2024). Evaluation of the whole-of-government trial of Microsoft 365 Copilot: Summary of evaluation findings. Evaluation by Nous Group.",
      "source_url": "https://www.digital.gov.au/sites/default/files/documents/2024-10/Copilot%20Microsoft%20365%20summary%20of%20evaluation%20findings.pdf",
      "quote": "Around 40% of trial participants were able to reallocate their time to higher value activities.",
      "caveat": "Self-reported and from volunteers nominated by their agencies; read from an archived copy of the PDF because the live site timed out. The closest thing found to a 'share of gain kept', and it is a share of people, not of hours.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "dta-hour-a-day-on-few-tasks",
      "url": "https://www.ai-first.company/payback/#source-dta-hour-a-day-on-few-tasks",
      "claim": "The same trial found the perceived saving, about an hour a day, was limited to three tasks: summarising, first drafts and searching for information.",
      "value": 1,
      "low": null,
      "high": null,
      "unit": "hour a day (perceived, on three task types)",
      "population": "Australian Public Service trial participants, 2024",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time freed",
      "source": "Digital Transformation Agency (October 2024). Evaluation of the whole-of-government trial of Microsoft 365 Copilot: Summary of evaluation findings.",
      "source_url": "https://www.digital.gov.au/sites/default/files/documents/2024-10/Copilot%20Microsoft%20365%20summary%20of%20evaluation%20findings.pdf",
      "quote": "Improvements in efficiency and quality were perceived to occur in a few tasks with perceived time savings of around an hour a day for these tasks.",
      "caveat": "Perceived, not timed. Supports targeting: the gain sat in a few named tasks.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "treasury-expectation-gap",
      "url": "https://www.ai-first.company/payback/#source-treasury-expectation-gap",
      "claim": "In the Australian Treasury's Copilot trial, 85% expected the tool to speed up their work beforehand and 49% said afterwards that it had.",
      "value": null,
      "low": 49,
      "high": 85,
      "unit": "% agreeing it speeds up work (after vs expected before)",
      "population": "218 Treasury staff, 20 May to 23 August 2024",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time put to use",
      "source": "Australian Centre for Evaluation, The Treasury (February 2025). Evaluation of a trial of generative artificial intelligence (Copilot) in The Department of the Treasury.",
      "source_url": "https://evaluation.treasury.gov.au/sites/evaluation.treasury.gov.au/files/2025-02/evaluation-generative-artificial-intelligence.pdf",
      "quote": "49 per cent of respondents agreed or strongly agreed that Copilot improved the speed at which they completed tasks (with 22 per cent disagreeing or strongly disagreeing to this statement). While this is lower than the initial expectations reported in the pre-trial survey (85 per cent reported expecting that Copilot would improve the speed at which they complete tasks)",
      "caveat": "Most participants used the tool two or three times a week or less.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "mcelheran-j-curve",
      "url": "https://www.ai-first.company/payback/#source-mcelheran-j-curve",
      "claim": "In US manufacturing, plants that adopted industrial AI first became less productive before gains appeared; the losses fell mainly on older businesses.",
      "value": -1.33,
      "low": null,
      "high": null,
      "unit": "percentage points of productivity per standard deviation of AI use (short run)",
      "population": "US manufacturing establishments, Census Bureau data for 2017 and 2021",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Time put to use",
      "source": "McElheran, K., Yang, M.-J., Kroff, Z. and Brynjolfsson, E. (2025). The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s). US Census Bureau, Center for Economic Studies Working Paper CES 25-27, April 2025 (working paper).",
      "source_url": "https://www2.census.gov/library/working-papers/2025/adrm/ces/CES-WP-25-27.pdf",
      "quote": "a one–standard-deviation increase in our AI index correlates with a 1.33 percentage-point drop in productivity.",
      "caveat": "Industrial AI in factories before 2022, not chat assistants. The paper's causal estimate is far larger (around -60 points) and applies only to a narrow group of plants, so do not use it as a general figure.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "productivity-j-curve",
      "url": "https://www.ai-first.company/payback/#source-productivity-j-curve",
      "claim": "A new general-purpose technology needs unmeasured investment in processes and skills first, so measured productivity dips before it rises: the 'productivity J-curve'.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "US economy; software, hardware and R&D intangibles",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Time put to use",
      "source": "Brynjolfsson, E., Rock, D. and Syverson, C. (2018). The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. NBER Working Paper 25148; published in American Economic Journal: Macroeconomics, 13(1), 2021, 333-372 (peer-reviewed).",
      "source_url": "https://www.nber.org/system/files/working_papers/w25148/w25148.pdf",
      "quote": "Our model generates a Productivity J-Curve that can explain the productivity slowdowns often accompanying the advent of GPTs, as well as the increase in productivity later.",
      "caveat": "A model with historical calibration; it gives the shape, not the length of the dip for a given firm.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "babina-ai-firm-growth",
      "url": "https://www.ai-first.company/payback/#source-babina-ai-firm-growth",
      "claim": "US firms that invested more in AI grew faster in sales, employment and market value, mainly by launching new products.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "US public firms; AI investment measured from employee resumes",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Time put to use",
      "source": "Babina, T., Fedyk, A., He, A. and Hodson, J. (2024). Artificial intelligence, firm growth, and product innovation. Journal of Financial Economics, 151, 103745 (peer-reviewed).",
      "source_url": "https://www.sciencedirect.com/science/article/pii/S0304405X2300185X",
      "quote": "AI-investing firms experience higher growth in sales, employment, and market valuations. This growth comes primarily through increased product innovation.",
      "caveat": "Counter-evidence to 'AI does not pay'. Pre-generative AI, and the growth 'concentrates among larger firms'; the gain came from new products, not cost cutting. Abstract read from an archived copy.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "czarnitzki-ai-productivity-germany",
      "url": "https://www.ai-first.company/payback/#source-czarnitzki-ai-productivity-germany",
      "claim": "German firms that used AI were more productive, and the link held when the authors corrected for which firms choose to adopt.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "German firms, Community Innovation Survey data",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Time put to use",
      "source": "Czarnitzki, D., Fernández, G. P. and Rammer, C. (2023). Artificial intelligence and firm-level productivity. Journal of Economic Behavior & Organization, 211, 188-205 (peer-reviewed, open access).",
      "source_url": "https://www.sciencedirect.com/science/article/pii/S0167268123001531",
      "quote": "We find positive and significant associations between the use of AI and firm productivity.",
      "caveat": "Counter-evidence. Pre-generative AI; the abstract gives no single magnitude.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "aghion-france-employment-sales",
      "url": "https://www.ai-first.company/payback/#source-aghion-france-employment-sales",
      "claim": "French firms that adopted AI saw employment and sales rise afterwards, with different results for different uses.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "French firms, 2017-2020, national ICT survey",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Time put to use",
      "source": "Aghion, P., Bunel, S., Jaravel, X., Mikaelsen, T., Roulet, A. and Søgaard, J. (2025). How Different Uses of AI Shape Labor Demand: Evidence from France. AEA Papers and Proceedings, 115, 62-67 (lightly reviewed conference volume).",
      "source_url": "https://ideas.repec.org/a/aea/apandp/v115y2025p62-67.html",
      "quote": "difference-in-difference estimates reveal an increase in firm-level employment and sales after AI adoption, suggesting that the induced productivity gains allow firms to grow and outweigh potential displacement effects.",
      "caveat": "Counter-evidence. First-generation AI, and adopters were already larger and more productive.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "ai-extended-workday",
      "url": "https://www.ai-first.company/payback/#source-ai-extended-workday",
      "claim": "A US time-diary study found that people in jobs more exposed to AI worked longer hours and had less leisure, not more.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "US individual time-diary data, 2004-2023",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Time put to use",
      "source": "Jiang, W., Park, J., Xiao, R. and Zhang, S. (2025). AI and the Extended Workday: Productivity, Contracting Efficiency, and Distribution of Rents. NBER Working Paper 33536 (working paper).",
      "source_url": "https://www.nber.org/system/files/working_papers/w33536/w33536.pdf",
      "quote": "we find that higher AI exposure—whether stemming from the ChatGPT shock or broader AI evolution—is associated with longer work hours and reduced leisure time",
      "caveat": "An association with occupational exposure, not with measured AI use.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "nanda-95-percent",
      "url": "https://www.ai-first.company/evidence/claim-checks/mit-95-percent-of-ai-pilots-fail/",
      "claim": "The MIT NANDA report's '95%' is the share of organizations in its sample getting no measurable profit-and-loss return from generative AI; it also says only 5% of custom enterprise AI tools reached production.",
      "value": 95,
      "low": null,
      "high": null,
      "unit": "% of organizations with no measurable P&L return (interview-based)",
      "population": "52 organizations interviewed, 153 senior leaders surveyed at four conferences, over 300 public AI initiatives reviewed",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Time put to use",
      "source": "Challapally, A., Pease, C., Raskar, R. and Chari, P. (July 2025). The GenAI Divide: State of AI in Business 2025. MIT NANDA (project report, not peer-reviewed; copy hosted by MLQ.ai).",
      "source_url": "https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf",
      "quote": "Just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact.",
      "caveat": "The report calls its figures 'directionally accurate based on individual interviews rather than official company reporting' and says 'success definitions may differ across organizations'; success was judged six months after a pilot. Do not present 95% as a measured failure rate.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "nanda-success-definition",
      "url": "https://www.ai-first.company/evidence/claim-checks/mit-95-percent-of-ai-pilots-fail/",
      "claim": "The same report defines success as a tool that users or executives 'remarked' had a marked and lasting effect, which is a judgement and not a measurement.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "52 organizations interviewed",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Background",
      "source": "Challapally, A. et al. (July 2025). The GenAI Divide: State of AI in Business 2025. MIT NANDA.",
      "source_url": "https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf",
      "quote": "We define successfully implemented for task-specific GenAI tools as ones users or executives have remarked as causing a marked and sustained productivity and/or P&L impact",
      "caveat": "",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "sp-global-42-percent",
      "url": "https://www.ai-first.company/payback/#source-sp-global-42-percent",
      "claim": "In S&P Global's 2025 survey, 42% of companies said they had abandoned most of their AI initiatives before production, up from 17% a year earlier.",
      "value": 42,
      "low": null,
      "high": null,
      "unit": "% of companies abandoning most AI initiatives",
      "population": "1,006 mid-level and senior IT and business professionals in North America and Europe",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time put to use",
      "source": "S&P Global Market Intelligence (30 May 2025). Generative AI experiences rapid adoption, but with mixed outcomes: Highlights from VotE: AI & Machine Learning. Summary of 'Voice of the Enterprise: AI & Machine Learning, Use Cases 2025' (vendor research; full report behind a paywall).",
      "source_url": "https://www.spglobal.com/market-intelligence/en/news-insights/research/ai-experiences-rapid-adoption-but-with-mixed-outcomes-highlights-from-vote-ai-machine-learning",
      "quote": "The percentage of companies abandoning the majority of their AI initiatives before they reach production has surged from 17% to 42% year over year, with organizations on average reporting that 46% of projects are scrapped between proof of concept and broad adoption.",
      "caveat": "Read from an archived copy of S&P's own summary page; the questionnaire and the base for each figure are in the paid report. In the same survey most respondents still reported positive impact from generative AI.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "gartner-30-percent-prediction",
      "url": "https://www.ai-first.company/payback/#source-gartner-30-percent-prediction",
      "claim": "Gartner predicted in 2024 that at least 30% of generative-AI projects would be abandoned after proof of concept by the end of 2025.",
      "value": 30,
      "low": null,
      "high": null,
      "unit": "% of projects (predicted)",
      "population": "Analyst prediction",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Time put to use",
      "source": "Gartner (29 July 2024). Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025. Press release.",
      "source_url": "https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025",
      "quote": "At least 30% of generative AI (GenAI) projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value",
      "caveat": "A forecast with no published method, not a count. Read from an archived copy.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "rand-why-ai-projects-fail",
      "url": "https://www.ai-first.company/payback/#source-rand-why-ai-projects-fail",
      "claim": "RAND interviewed 65 data scientists and engineers about failed AI projects; the first cause they named was misunderstanding the problem, including models 'optimized for the wrong metrics'.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "65 experienced data scientists and engineers, interviews",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Background",
      "source": "Ryseff, J., De Bruhl, B. F. and Newberry, S. J. (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Corporation, RR-A2680-1.",
      "source_url": "https://www.rand.org/content/dam/rand/pubs/research_reports/RRA2600/RRA2680-1/RAND_RRA2680-1.pdf",
      "quote": "Too often, trained AI models are deployed that have been optimized for the wrong metrics or do not fit into the overall business workflow and context.",
      "caveat": "The '80 percent of AI projects fail' line often credited to RAND is introduced there as 'By some estimates'; RAND did not measure it. Leave the 80% out.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "mckinsey-2025-high-performers",
      "url": "https://www.ai-first.company/evidence/studies/state-of-ai-2025-survey/",
      "claim": "In McKinsey's 2025 survey about 6% of organizations credited AI with 5% or more of operating profit, and they were nearly three times as likely to have redesigned workflows.",
      "value": 6,
      "low": null,
      "high": null,
      "unit": "% of organizations",
      "population": "1,993 respondents in 105 countries",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Time put to use",
      "source": "McKinsey & Company (5 November 2025). The State of AI: Global Survey 2025 (consultancy survey).",
      "source_url": "https://web.archive.org/web/20251215/https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai",
      "quote": "88% of organizations use AI in at least one business function, but only about 6% are “high performers” that credit AI with 5% or more of their operating profit.",
      "caveat": "The quote is the headline stored on the site's study page (the site's own summary of the survey), not text copied from McKinsey in this session. Self-reported profit, and an association, not a cause.",
      "study": "https://www.ai-first.company/evidence/studies/state-of-ai-2025-survey/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "workday-40-percent-rework",
      "url": "https://www.ai-first.company/payback/#source-workday-40-percent-rework",
      "claim": "In a vendor survey, employees said nearly 40% of the time AI saved them was lost to fixing and checking its output.",
      "value": 40,
      "low": null,
      "high": null,
      "unit": "% of reported time savings lost to rework",
      "population": "3,200 full-time employees who use AI, at organizations with $100M+ revenue, North America, EMEA and APAC, November 2025",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Running cost",
      "source": "Workday (14 January 2026). New Workday Research: Companies Are Leaving AI Gains on the Table. Press release for 'Beyond Productivity: Measuring the Real Value of AI', fielded by Hanover Research (vendor survey).",
      "source_url": "https://newsroom.workday.com/2026-01-14-New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table",
      "quote": "Nearly 40% of AI time savings are lost to rework, including correcting errors, rewriting content, and verifying outputs from one-size-fits-all AI tools.",
      "caveat": "Self-reported, from a company that sells workplace software; the press release does not show how the share was calculated. The only number found for rework as a share of the gain: usable as an illustrative range at most.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "workslop-40-percent",
      "url": "https://www.ai-first.company/payback/#source-workslop-40-percent",
      "claim": "In a BetterUp and Stanford survey, 40% of US desk workers said they had received 'workslop', polished-looking AI output that leaves the real work to the receiver, in the past month.",
      "value": 40,
      "low": null,
      "high": null,
      "unit": "% of desk workers who received workslop last month",
      "population": "1,150 full-time US desk workers, online survey, September 2025",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Running cost",
      "source": "BetterUp Labs and Stanford Social Media Lab (2025). Workslop: The Hidden Cost of AI-Generated Busywork. Reported in Niederhoffer, K., Kellerman, G. R., Lee, A., Liebscher, A., Rapuano, K. and Hancock, J. T. (22 September 2025). AI-Generated “Workslop” Is Destroying Productivity. Harvard Business Review (vendor survey; magazine article, not peer-reviewed).",
      "source_url": "https://www.betterup.com/workslop",
      "quote": "Insights are based on an online survey of 1,150 full-time U.S. desk workers conducted in September 2025 by BetterUp in partnership with the Stanford Social Media Lab.",
      "caveat": "Figures read from BetterUp's own page; the HBR article body was behind a paywall. BetterUp sells coaching.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "workslop-cost-186",
      "url": "https://www.ai-first.company/payback/",
      "claim": "The same survey puts the cost of workslop at about two hours per incident and $186 per employee per month, from respondents' own estimates.",
      "value": 186,
      "low": null,
      "high": null,
      "unit": "USD per employee per month (self-estimated)",
      "population": "1,150 full-time US desk workers, September 2025",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Running cost",
      "source": "BetterUp Labs and Stanford Social Media Lab (2025). Workslop: The Hidden Cost of AI-Generated Busywork.",
      "source_url": "https://www.betterup.com/workslop",
      "quote": "$186 Monthly cost per employee caused by these incidents.",
      "caveat": "Built from self-estimated time multiplied by self-reported pay; the $9 million a year for a 10,000-person company is an extrapolation of the same figure. The method detail (1 hour 56 minutes per incident) is in the paywalled HBR text and was not verified.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "dbt-hallucinations-22",
      "url": "https://www.ai-first.company/evidence/cases/uk-department-for-business-and-trade-copilot-pilot/",
      "claim": "In the UK department's Copilot pilot, 22% of diary respondents had spotted made-up content; checking was inconsistent.",
      "value": 22,
      "low": null,
      "high": null,
      "unit": "% of respondents who identified a hallucination",
      "population": "About 300 diary respondents among 1,000 licence holders",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Running cost",
      "source": "UK Department for Business and Trade (August 2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade.",
      "source_url": "https://assets.publishing.service.gov.uk/media/68adbe409e1cebdd2c96a19d/dbt-microsoft-365-copilot-evaluation.pdf",
      "quote": "When asked to recall hallucinations, 43% of respondents reported they did not identify a hallucination at any point throughout the pilot. However, 22% of respondents did identify hallucinations",
      "caveat": "People who did not notice an error cannot report it, so this is a floor.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "dbt-tasks-that-took-longer",
      "url": "https://www.ai-first.company/evidence/cases/uk-department-for-business-and-trade-copilot-pilot/",
      "claim": "Some tasks took longer with Copilot: scheduling cost 0.6 hours more per task and generating images 0.5 hours more, by users' diaries.",
      "value": null,
      "low": -0.6,
      "high": -0.5,
      "unit": "hours per task (negative = slower)",
      "population": "Diary study, UK Department for Business and Trade",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Running cost",
      "source": "UK Department for Business and Trade (August 2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade.",
      "source_url": "https://assets.publishing.service.gov.uk/media/68adbe409e1cebdd2c96a19d/dbt-microsoft-365-copilot-evaluation.pdf",
      "quote": "“scheduling” resulted in mean time savings of -0.6 hours per task, and “generating images” resulted in mean time savings of -0.5 hours per task, meaning they both took longer when M365 Copilot were used.",
      "caveat": "",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "humlum-new-tasks-8-4",
      "url": "https://www.ai-first.company/evidence/studies/chatbots-and-the-labour-market-denmark/",
      "claim": "Chatbots created new job tasks for 8.4% of workers in the Danish study, including people who do not use the tools themselves.",
      "value": 8.4,
      "low": null,
      "high": null,
      "unit": "% of workers with new tasks",
      "population": "About 25,000 workers in 11 occupations, Denmark",
      "kind": "Self-reported",
      "design": "survey",
      "line": "Running cost",
      "source": "Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56 (May 2025 version).",
      "source_url": "https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-3.pdf",
      "quote": "AI chatbots have created new job tasks for 8.4% of workers, including some who do not use the tools themselves.",
      "caveat": "New tasks include checking AI output and fitting AI into systems; the figure is a share of workers, not hours.",
      "study": "https://www.ai-first.company/evidence/studies/chatbots-and-the-labour-market-denmark/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "treasury-13-minutes-break-even",
      "url": "https://www.ai-first.company/payback/#source-treasury-13-minutes-break-even",
      "claim": "The Australian Treasury worked out that a mid-level official needs to move about 13 minutes a week to higher-value work to cover the Copilot licence.",
      "value": 13,
      "low": null,
      "high": null,
      "unit": "minutes a week per user to cover the licence",
      "population": "Australian Public Service level 6 salary, Copilot licence price in 2024",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Running cost",
      "source": "Australian Centre for Evaluation, The Treasury (February 2025). Evaluation of a trial of generative artificial intelligence (Copilot) in The Department of the Treasury.",
      "source_url": "https://evaluation.treasury.gov.au/sites/evaluation.treasury.gov.au/files/2025-02/evaluation-generative-artificial-intelligence.pdf",
      "quote": "ACE estimates that a staff member on the current APS6 salary would need to redirect approximately 13 minutes of time per week to higher value tasks for the current licence cost to be offset.",
      "caveat": "Covers the licence only; the report says training and onboarding costs come on top and that the trial 'did not quantify time savings'. The formula (licence price divided by hourly pay) can be reused with the reader's own figures.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "price-microsoft-365-copilot-business",
      "url": "https://www.ai-first.company/payback/#source-price-microsoft-365-copilot-business",
      "claim": "Microsoft lists its Copilot add-on for small and mid-sized businesses at $18 per user a month paid yearly (shown as reduced from $21), on top of a Microsoft 365 plan.",
      "value": 18,
      "low": null,
      "high": null,
      "unit": "USD per user per month, paid yearly",
      "population": "Vendor list price, US pricing page, checked 13 October 2026",
      "kind": "Vendor list price",
      "design": "report",
      "line": "Running cost",
      "source": "Microsoft. Microsoft 365 Copilot pricing page (vendor price list), checked 13 October 2026.",
      "source_url": "https://www.microsoft.com/en-us/microsoft-365-copilot/pricing",
      "quote": "Microsoft 365 Copilot Business ... Originally starting from $21.00 now starting from $18.00 ... user/month, paid yearly",
      "caveat": "List price on the day checked, before tax; shown as a promotion and requires an existing Microsoft 365 plan. The bundle 'Microsoft 365 Business Standard with Copilot' was listed at $23.50 per user a month paid yearly.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "price-claude-team",
      "url": "https://www.ai-first.company/payback/#source-price-claude-team",
      "claim": "Anthropic lists its Claude Team plan at $20 per seat a month billed annually, or $25 billed monthly, with a premium seat at $100.",
      "value": 20,
      "low": null,
      "high": null,
      "unit": "USD per seat per month, billed annually",
      "population": "Vendor list price, checked 13 October 2026",
      "kind": "Vendor list price",
      "design": "report",
      "line": "Running cost",
      "source": "Anthropic. Claude pricing page (vendor price list), checked 13 October 2026.",
      "source_url": "https://claude.com/pricing",
      "quote": "Standard seat ... $20 Per seat / month if billed annually. $25 if billed monthly.",
      "caveat": "List price on the day checked, before tax; for teams of 2 to 150.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "price-google-workspace-standard",
      "url": "https://www.ai-first.company/payback/",
      "claim": "Google bundles its Gemini assistant into Workspace; the Business Standard plan was listed at zł 63.00 per user a month on the page served in Poland.",
      "value": 63,
      "low": null,
      "high": null,
      "unit": "PLN per user per month (list, before tax)",
      "population": "Vendor list price, pricing page as served to a Polish connection, checked 13 October 2026",
      "kind": "Vendor list price",
      "design": "report",
      "line": "Running cost",
      "source": "Google. Google Workspace pricing page (vendor price list), checked 13 October 2026.",
      "source_url": "https://workspace.google.com/pricing",
      "quote": "Standard ... zł 63.00 ... Price is per user/month and does not include tax",
      "caveat": "The page localises currency and showed temporary discounts; the US dollar price was not seen in this session, so convert or re-check before showing. It is the price of the whole office suite, not of the AI alone.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "svanberg-implementation-cost",
      "url": "https://www.ai-first.company/payback/",
      "claim": "A cost model for a custom-built AI vision system assumes $1.77 million to build and $243,000 a year to maintain, or $165,000 and $123,000 for a bare-bones version.",
      "value": null,
      "low": 165000,
      "high": 1765000,
      "unit": "USD up-front implementation (bare-bones vs full build)",
      "population": "Engineering cost assumptions taken from one deep-learning case study (Thompson et al. 2021)",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Running cost",
      "source": "Svanberg, M., Li, W., Fleming, M., Goehring, B. and Thompson, N. (2024). Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision? MIT FutureTech working paper.",
      "source_url": "https://futuretech-site.s3.us-east-2.amazonaws.com/2024-01-18+Beyond_AI_Exposure.pdf",
      "quote": "The study reports an upfront implementation cost of C eng,imp = $1, 765, 000 for a 6-month project and a yearly maintenance cost of C eng,m = $242, 840.",
      "caveat": "For a custom-built system, from a single case study. Shows that maintenance is a recurring cost of the same order as a team, not that any reader's project costs this.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "gartner-5-20-million",
      "url": "https://www.ai-first.company/payback/",
      "claim": "Gartner put the up-front cost of the generative-AI deployment approaches it examined at $5 million to $20 million.",
      "value": null,
      "low": 5,
      "high": 20,
      "unit": "million USD up-front",
      "population": "Analyst estimate for large enterprises",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Running cost",
      "source": "Gartner (29 July 2024). Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025. Press release.",
      "source_url": "https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025",
      "quote": "these deployment approaches come with significant costs, ranging from $5 million to $20 million",
      "caveat": "No published method, and sized for large enterprises; not usable for a small or mid-sized company. Read from an archived copy.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "bcg-10-20-70",
      "url": "https://www.ai-first.company/payback/",
      "claim": "BCG's 10-20-70 rule is an assertion about where effort goes: 10% on the model, 20% on data, 70% on changing business processes.",
      "value": 70,
      "low": null,
      "high": null,
      "unit": "% of effort on process change (rule of thumb)",
      "population": "Consultancy rule of thumb; no sample or method stated",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Running cost",
      "source": "Boston Consulting Group (2022). 5 Rules for Fixing AI and Machine Learning for Your Business.",
      "source_url": "https://www.bcg.com/publications/2022/5-rules-for-fixing-ai-and-machine-learning-for-your-business",
      "quote": "10% of the effort will lie in building an adequate AI model. 20% will involve high-quality data. 70% will focus on developing new business processes or transforming the way business functions operate.",
      "caveat": "It speaks of effort, not money, and the page gives no data behind it. Read from an archived copy.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "sam-set-up-and-running-cost",
      "url": "https://www.ai-first.company/evidence/cases/sam-retirement-advisory-ai-agents/",
      "claim": "A two-person French advisory firm reports what its AI set-up cost: about €3,000 for a server, nearly €650 a month in licences, and about 60 hours of set-up work.",
      "value": 650,
      "low": null,
      "high": null,
      "unit": "EUR a month in licences (plus about 3,000 EUR and 60 hours to set up)",
      "population": "One two-person firm, self-reported to a government support programme",
      "kind": "Reported, not measured",
      "design": "case",
      "line": "Running cost",
      "source": "France Num (French government). Case page on SAM, a retirement advisory firm that automated its client files with AI.",
      "source_url": "https://www.francenum.gouv.fr/guides-et-conseils/intelligence-artificielle/automatisation-et-optimisation-des-processus/dans-les",
      "quote": "Le coût de mise en place du serveur dédié est d'environ 3 000 euros et l'entreprise dépense près de 650 euros par mois pour les licences des solutions IA. A ces coûts, il faut ajouter, un investissement d’intégration initial d’environ 60 heures",
      "caveat": "One firm's own account; the only small-company case on the site with all three cost lines stated.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "meta-token-leaderboard-cost",
      "url": "https://www.ai-first.company/payback/#source-meta-token-leaderboard-cost",
      "claim": "An employee-built leaderboard at Meta ranked staff by AI tokens used; Fortune calculated that the top user's 30-day total could have cost more than $1.4 million at list prices.",
      "value": 1.4,
      "low": null,
      "high": null,
      "unit": "million USD (journalist's estimate for one user, 30 days)",
      "population": "Meta, more than 85,000 employees; press report based on The Information",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Running cost",
      "source": "Fortune (9 April 2026). Report on Meta's 'Claudeonomics' employee token leaderboard and its removal, citing The Information and a Meta statement.",
      "source_url": "https://www.fortune.com/2026/04/09/meta-killed-employee-ai-token-dashboard",
      "quote": "In a 30-day period, total employee usage on the dashboard exceeded 60 trillion tokens, and the highest-ranked individual user averaged 281 billion tokens. Using the least expensive version of Claude Opus 4.6, which costs $5 for every million tokens, that one user alone could have cost Meta more than $1.4 million.",
      "caveat": "Press report; the cost is Fortune's arithmetic at list price, not Meta's bill. Meta told Fortune the employee took the dashboard down and that it did not ask for this.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "amazon-kirorank",
      "url": "https://www.ai-first.company/payback/#source-amazon-kirorank",
      "claim": "Amazon dropped an internal AI-usage ranking after employees ran AI agents on needless tasks to climb it, according to a Financial Times report.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "Amazon developers; press report citing insiders",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Running cost",
      "source": "heise online (29 May 2026). Too much tokenmaxxing? Amazon stops internal AI ranking. Summary of a Financial Times report, quoting Amazon senior vice-president Dave Treadwell.",
      "source_url": "https://www.heise.de/en/news/Too-much-tokenmaxxing-Amazon-stops-internal-AI-ranking-11311902.html",
      "quote": "the result was additional costs for Amazon because employees had artificially inflated their AI token consumption. \"Please don’t use AI just for the sake of using AI,\" he reportedly said.",
      "caveat": "Secondary report of a paywalled FT article, which was not opened; no cost figure given. Amazon is reported to have switched to a measure of meaningful code deployed.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "metr-forecast-vs-measured",
      "url": "https://www.ai-first.company/payback/#source-metr-forecast-vs-measured",
      "claim": "The developers in the METR trial expected AI to make them 24% faster, and afterwards believed it had made them 20% faster, while the measured result was 19% slower.",
      "value": null,
      "low": -19,
      "high": 20,
      "unit": "% speed change (measured vs self-estimated)",
      "population": "16 experienced open-source developers, 246 tasks",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Background",
      "source": "Becker, J., Rush, N., Barnes, B. and Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089v2.",
      "source_url": "https://arxiv.org/pdf/2507.09089",
      "quote": "Before starting tasks, developers forecast that allowing AI will reduce completion time by 24%. After completing the study, developers estimate that allowing AI reduced completion time by 20%.",
      "caveat": "The clearest single case of belief and measurement pointing in opposite directions; sixteen people.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "efficiency-gain-illusion",
      "url": "https://www.ai-first.company/payback/#source-efficiency-gain-illusion",
      "claim": "In three pre-registered experiments on simple tasks, people expected AI to save 55.7 seconds and it saved 7.5.",
      "value": null,
      "low": 7.5,
      "high": 55.7,
      "unit": "seconds saved (measured vs predicted)",
      "population": "2,691 online participants, simple tasks such as arithmetic and spell-checking",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Background",
      "source": "Yu, S., Cheng, M., Jabbar, A., Sucholutsky, I., Collins, K. M., Jurafsky, D. and Hawkins, R. D. (2026). The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks. arXiv:2605.22687 (preprint, not peer-reviewed).",
      "source_url": "https://arxiv.org/pdf/2605.22687",
      "quote": "On average, people predicted AI assistance to save time by 55.7 seconds when it only saved 7.5 seconds.",
      "caveat": "Very short lab tasks, not workplace work.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "dbt-diary-vs-timed-task",
      "url": "https://www.ai-first.company/evidence/cases/uk-department-for-business-and-trade-copilot-pilot/",
      "claim": "In the UK department's pilot, users' diaries reported time saved on data analysis, while a timed test found Copilot users did Excel analysis more slowly and worse.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "Diary study and observed task sessions, UK Department for Business and Trade",
      "kind": "Measured in a study",
      "design": "field study",
      "line": "Background",
      "source": "UK Department for Business and Trade (August 2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade.",
      "source_url": "https://assets.publishing.service.gov.uk/media/68adbe409e1cebdd2c96a19d/dbt-microsoft-365-copilot-evaluation.pdf",
      "quote": "M365 Copilot users completed Excel data analysis more slowly and to a worse quality and accuracy than non-users, conflicting time savings reported in the diary study for data analysis.",
      "caveat": "The same report says that for most tasks the timed sessions showed 'similar trends' to the diaries, 'suggesting bias had small impacts on time saving findings'. Self-reports were wrong on one task, not on all.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "coinbase-ai-mandate",
      "url": "https://www.ai-first.company/payback/#source-coinbase-ai-mandate",
      "claim": "Coinbase's chief executive said he gave engineers a week to start using AI coding tools and fired some who did not.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "Coinbase engineers; the chief executive's own account on a podcast",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Background",
      "source": "Fortune (25 August 2025). Report on Coinbase chief executive Brian Armstrong's AI coding mandate, quoting his remarks on the Cheeky Pint podcast.",
      "source_url": "https://fortune.com/2025/08/25/coinbase-ceo-brian-armstrong-ai-coding-assistants-mandate-tech",
      "quote": "“Some of them had a good reason, because they were just getting back from some trip or something,” Armstrong said. “Some of them didn’t, and they got fired.”",
      "caveat": "Reported, not measured: no outcome data. Armstrong claims about a third of the company's code is written by AI; that is his claim.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "shopify-ai-in-performance-reviews",
      "url": "https://www.ai-first.company/payback/#source-shopify-ai-in-performance-reviews",
      "claim": "Shopify's chief executive told staff that AI use would be part of performance reviews and that teams must show AI cannot do a job before asking for more people.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "Shopify, about 8,100 employees; chief executive's memo",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Background",
      "source": "CNBC (7 April 2025). Shopify CEO says staffers need to prove jobs can't be done by AI before asking for more headcount. Report on Tobi Lütke's memo, which he posted on X.",
      "source_url": "https://www.cnbc.com/2025/04/07/shopify-ceo-prove-ai-cant-do-jobs-before-asking-for-more-headcount.html",
      "quote": "The company, which sells web-based software that helps online retailers manage sales and run their operations, will factor AI usage into performance reviews, he added.",
      "caveat": "Reported, not measured: no published result of the policy.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "duolingo-memo-walk-back",
      "url": "https://www.ai-first.company/evidence/cases/duolingo-ai-course-content/",
      "claim": "Duolingo's chief executive said months later that his 'AI-first' memo 'did not give enough context'.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "Duolingo; chief executive's statements to the New York Times, reported by Fortune",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Background",
      "source": "Fortune (18 August 2025). Duolingo CEO admits his controversial AI memo ‘did not give enough context’ and insists the company never laid off full-time employees.",
      "source_url": "https://fortune.com/2025/08/18/duolingo-ceo-admits-controversial-ai-memo-did-not-give-enough-context-insists-company-never-laid-off-full-time-employees",
      "quote": "“This was on me. I did not give enough context,” von Ahn told the New York Times in a recent interview when asked about the controversial memo.",
      "caveat": "A statement about communication; the company kept the policy.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "klarna-cost-as-the-objective",
      "url": "https://www.ai-first.company/evidence/cases/klarna-ai-customer-service/",
      "claim": "Klarna's chief executive said the company had weighed cost too heavily in customer service and got lower quality; it began hiring human agents again.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "Klarna; chief executive's statement to Bloomberg",
      "kind": "Reported, not measured",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "Customer Experience Dive (9 May 2025). Report on Klarna reinvesting in human customer-service staff, quoting chief executive Sebastian Siemiatkowski's remarks to Bloomberg.",
      "source_url": "https://www.customerexperiencedive.com/news/klarna-reinvests-human-talent-customer-service-AI-chatbot/747586/",
      "quote": "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this ... what you end up having is lower quality.",
      "caveat": "No cost of the reversal was published, and Klarna kept the assistant and went on reporting savings from it.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "holmstrom-milgrom-multitask",
      "url": "https://www.ai-first.company/payback/#source-holmstrom-milgrom-multitask",
      "claim": "The standard economic result on measuring the wrong thing: when only part of a job is measured and rewarded, effort moves to the measured part.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "Economic theory",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Background",
      "source": "Holmström, B. and Milgrom, P. (1991). Multitask Principal-Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design. The Journal of Law, Economics, and Organization, 7 (special issue), 24-52 (peer-reviewed).",
      "source_url": "https://econpapers.repec.org/article/oupjleorg/v_3a7_3ay_3a1991_3ai_3a0_3ap_3a24-52.htm",
      "quote": "Multitask Principal-Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design",
      "caveat": "Only the bibliographic record was opened (the journal page was blocked), so the quote is the title and the claim is the paper's well-known result stated from general knowledge.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "so-so-technologies",
      "url": "https://www.ai-first.company/payback/",
      "claim": "Acemoglu and Restrepo call automation that replaces people while saving little a 'so-so technology': it cuts demand for labour without much productivity gain. Their example is automated customer service.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "Economic theory, with US historical data",
      "kind": "Calculated by the study’s authors",
      "design": "model",
      "line": "Background",
      "source": "Acemoglu, D. and Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2), 3-30 (peer-reviewed); quoted from NBER Working Paper 25684.",
      "source_url": "https://www.nber.org/system/files/working_papers/w25684/w25684.pdf",
      "quote": "it is not the “brilliant” automation technologies that threaten employment and wages, but “so-so technologies” that generate small productivity improvements.",
      "caveat": "",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "pro-worker-ai",
      "url": "https://www.ai-first.company/payback/",
      "claim": "Acemoglu, Autor and Johnson argue that AI's use as a collaborator, extending judgement and creating new tasks, is under-used compared with its use to automate.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "Policy paper with worked examples",
      "kind": "Reported, not measured",
      "design": "report",
      "line": "Background",
      "source": "Acemoglu, D., Autor, D. and Johnson, S. (2026). Building Pro-Worker Artificial Intelligence. NBER Working Paper 34854, February 2026.",
      "source_url": "https://www.nber.org/papers/w34854",
      "quote": "Only the last category is unambiguously pro-worker",
      "caveat": "An argument about direction, not a tested estimate; the paper states no measured effect size. Quote is a short extract of the abstract returned by the fetch tool.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "vaccaro-human-plus-ai",
      "url": "https://www.ai-first.company/evidence/studies/human-ai-combinations-meta-analysis/",
      "claim": "Across 106 experiments, a human working with AI did worse on average than the better of the two alone.",
      "value": null,
      "low": null,
      "high": null,
      "unit": "",
      "population": "106 experiments, 370 comparisons, published 2020 to mid-2023",
      "kind": "Measured in a study",
      "design": "experiment",
      "line": "Background",
      "source": "Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293-2303 (peer-reviewed).",
      "source_url": "https://doi.org/10.1038/s41562-024-02024-1",
      "quote": "on average, human-AI combinations performed significantly worse than the best of humans or AI alone (Hedges’ g = −0.23, 95% confidence interval −0.39 to −0.07).",
      "caveat": "A standardised effect size, not a percentage of output; the comparison is with the better of the two, and creation tasks gained.",
      "study": "https://www.ai-first.company/evidence/studies/human-ai-combinations-meta-analysis/",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-air-canada",
      "url": "https://www.ai-first.company/evidence/cases/air-canada-chatbot-ruling/",
      "claim": "A Canadian tribunal ordered Air Canada to pay a customer C$812.02 after its website chatbot gave wrong information about fares.",
      "value": 812.02,
      "low": null,
      "high": null,
      "unit": "CAD",
      "population": "One tribunal decision, British Columbia, February 2024",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "Moffatt v. Air Canada, 2024 BCCRT 149. British Columbia Civil Resolution Tribunal.",
      "source_url": "https://decisions.civilresolutionbc.ca/crt/crtd/en/item/525448/index.do",
      "quote": "they paid $1,630.36, I find they are entitled to damages of $650.88.",
      "caveat": "Small sum; the cost that matters is the precedent that the company answers for its chatbot. Total ordered was $812.02 with interest and fees.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-mata-avianca",
      "url": "https://www.ai-first.company/evidence/cases/mata-v-avianca-fake-citations/",
      "claim": "A New York court fined a law firm and two lawyers $5,000 for filing six court decisions that ChatGPT had made up.",
      "value": 5000,
      "low": null,
      "high": null,
      "unit": "USD",
      "population": "One federal court sanctions order, June 2023",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 22 June 2023). Opinion and order on sanctions.",
      "source_url": "https://storage.courtlistener.com/recap/gov.uscourts.nysd.575368/gov.uscourts.nysd.575368.54.0_3.pdf",
      "quote": "A penalty of $5,000 is jointly and severally imposed",
      "caveat": "",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-morgan-and-morgan",
      "url": "https://www.ai-first.company/evidence/cases/morgan-and-morgan-in-house-ai-fake-cases/",
      "claim": "A Wyoming court fined three lawyers $5,000 in total ($3,000 and two fines of $1,000) over eight made-up cases from their firm's own AI tool.",
      "value": 5000,
      "low": null,
      "high": null,
      "unit": "USD (3,000 + 1,000 + 1,000)",
      "population": "One federal court sanctions order, February 2025",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "Wadsworth v. Walmart Inc., No. 2:23-cv-118 (D. Wyo. 2025). Order on sanctions.",
      "source_url": "https://storage.courtlistener.com/recap/gov.uscourts.wyd.64014/gov.uscourts.wyd.64014.181.0.pdf",
      "quote": "the Court will order Mr. Ayala to pay a three-thousand-dollar ($3,000) fine.",
      "caveat": "The lawyers also paid the other side's fees (amount not stated on the site); openness reduced the penalty.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-butler-snow-rework",
      "url": "https://www.ai-first.company/evidence/cases/butler-snow-ai-policy-ignored/",
      "claim": "After five made-up citations, a US law firm paid another firm's 28 lawyers to check more than 2,400 citations in 330 filings.",
      "value": 2400,
      "low": null,
      "high": null,
      "unit": "citations re-checked (rework, cost not published)",
      "population": "One federal court sanctions order, Alabama, July 2025",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "Johnson v. Dunn, No. 2:21-cv-1701 (N.D. Ala. 23 July 2025). Sanctions order.",
      "source_url": "https://storage.courtlistener.com/recap/gov.uscourts.alnd.179677/gov.uscourts.alnd.179677.204.0.pdf",
      "quote": "“In all, Morgan Lewis reviewed more than 2,400 separate legal citations across 330 filings.”",
      "caveat": "The cost in money is not public; the measure is rework. Three lawyers were removed from the case.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-ayinde-wasted-costs",
      "url": "https://www.ai-first.company/evidence/cases/primus-solicitors-fake-citations/",
      "claim": "In an English case heard with Primus Solicitors', a law centre and a barrister were each ordered to pay £2,000 in wasted costs over five non-existent cases; the Primus filing had 18.",
      "value": 2000,
      "low": null,
      "high": null,
      "unit": "GBP each (second case in the judgment)",
      "population": "One High Court judgment covering two cases, June 2025",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin).",
      "source_url": "https://www.judiciary.uk/wp-content/uploads/2025/06/Ayinde-v-London-Borough-of-Haringey-and-Al-Haroun-v-Qatar-National-Bank.pdf",
      "quote": "In eighteen instances, the case cited does not exist.",
      "caveat": "The £2,000 orders were in the second case, where the court did not decide whether AI was the source. The Primus solicitor was referred to his regulator; no sum.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-accessibe",
      "url": "https://www.ai-first.company/evidence/cases/accessibe-ai-accessibility-claims/",
      "claim": "The US consumer regulator ordered accessiBe to pay $1 million over claims that its AI plug-in made any website accessible.",
      "value": 1000000,
      "low": null,
      "high": null,
      "unit": "USD",
      "population": "One FTC order, final April 2025",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "US Federal Trade Commission (22 April 2025). FTC approves final order requiring accessiBe to pay $1 million. Press release.",
      "source_url": "https://www.ftc.gov/news-events/news/press-releases/2025/04/ftc-approves-final-order-requiring-accessibe-pay-1-million",
      "quote": "The order prohibits accessiBe from making misleading claims and requires the company to pay $1 million.",
      "caveat": "",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-donotpay",
      "url": "https://www.ai-first.company/evidence/cases/donotpay-robot-lawyer/",
      "claim": "The US consumer regulator ordered DoNotPay to pay $193,000 for selling an untested 'robot lawyer'.",
      "value": 193000,
      "low": null,
      "high": null,
      "unit": "USD",
      "population": "One FTC order, final February 2025",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "US Federal Trade Commission (February 2025). FTC finalizes order with DoNotPay that prohibits deceptive ‘AI lawyer’ claims, imposes monetary relief. Press release.",
      "source_url": "https://www.ftc.gov/news-events/news/press-releases/2025/02/ftc-finalizes-order-donotpay-prohibits-deceptive-ai-lawyer-claims-imposes-monetary-relief-requires",
      "quote": "The final order requires DoNotPay to pay $193,000 in monetary relief",
      "caveat": "",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-delphia-global-predictions",
      "url": "https://www.ai-first.company/evidence/cases/delphia-global-predictions-ai-claims/",
      "claim": "The US securities regulator fined two investment advisers $225,000 and $175,000 for advertising AI they did not have.",
      "value": 400000,
      "low": null,
      "high": null,
      "unit": "USD in total (225,000 + 175,000)",
      "population": "Two SEC settled orders, March 2024",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "US Securities and Exchange Commission (March 2024). SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence. Press release 2024-36.",
      "source_url": "https://www.sec.gov/newsroom/press-releases/2024-36",
      "quote": "Delphia agreed to pay a civil penalty of $225,000, and Global Predictions agreed to pay a civil penalty of $175,000.",
      "caveat": "",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-rite-aid",
      "url": "https://www.ai-first.company/evidence/cases/rite-aid-facial-recognition/",
      "claim": "Rite Aid was banned from using facial recognition for surveillance for five years after deploying it without testing its accuracy.",
      "value": 5,
      "low": null,
      "high": null,
      "unit": "years banned from the technology",
      "population": "One FTC order, December 2023",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "US Federal Trade Commission (19 December 2023). Rite Aid banned from using AI facial recognition after FTC says retailer deployed technology without reasonable safeguards. Press release.",
      "source_url": "https://www.ftc.gov/news-events/news/press-releases/2023/12/rite-aid-banned-using-ai-facial-recognition-after-ftc-says-retailer-deployed-technology-without",
      "quote": "Rite Aid will be prohibited from using facial recognition technology for surveillance purposes for five years",
      "caveat": "A ban, not a sum: no money figure exists for this case.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-victoria-child-protection",
      "url": "https://www.ai-first.company/evidence/cases/victoria-child-protection-chatgpt/",
      "claim": "After a child-protection worker used ChatGPT for a court report, the regulator found nearly 900 staff had opened the site and ordered the department to block such tools.",
      "value": 900,
      "low": null,
      "high": null,
      "unit": "staff who had accessed the tool before any rule existed",
      "population": "One regulator investigation, Victoria, Australia, published September 2024",
      "kind": "Court or regulator record",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "Office of the Victorian Information Commissioner (September 2024). Investigation into the use of ChatGPT by a Child Protection worker.",
      "source_url": "https://ovic.vic.gov.au/regulatory-action/investigation-into-the-use-of-chatgpt-by-a-child-protection-worker/",
      "quote": "Within the period of July to December 2023, nearly 900 employees across DFFH had accessed",
      "caveat": "No fine; the cost was a two-year block on the tools and compulsory reporting.",
      "study": "",
      "reviewed": "2026-10-13"
    },
    {
      "id": "exposure-wpp-investment-without-return",
      "url": "https://www.ai-first.company/evidence/cases/wpp-open-ai-platform/",
      "claim": "WPP raised spending on its AI platform to £300 million a year; its 2025 income still fell and it cut its forecast to a fall of 5.5% to 6.0%.",
      "value": 300,
      "low": null,
      "high": null,
      "unit": "million GBP a year invested (site figure)",
      "population": "One listed company; annual report and trading update",
      "kind": "Reported, not measured",
      "design": "case",
      "line": "Risk: what it cost others",
      "source": "WPP plc (30 October 2025). Third Quarter Trading Update 2025; and WPP Annual Report on Form 20-F for 2024.",
      "source_url": "https://www.wpp.com/en/news/2025/10/third-quarter-trading-update-2025",
      "quote": "FY organic growth guidance revised to -5.5% to -6.0%",
      "caveat": "The £300 million figure is from the site's case page, which cites the 20-F; it was not re-located in the filing text in this session. WPP did not blame the decline on AI and has not published a return on the platform.",
      "study": "",
      "reviewed": "2026-10-13"
    }
  ]
}