{ "name": "AI Productivity Statistics for 2026", "version": "2026-08-16", "license": "CC BY 4.0", "methodologyUrl": "https://cognautic.com/answers/ai-productivity-statistics#method", "records": [ { "id": "nber-support-issues-per-hour", "statistic": "Customer-support issues resolved per hour with AI assistance", "direction": "increase", "value": "13.8", "unit": "percent", "population": "Customer-support agents at a Fortune 500 enterprise-software company", "task": "Live customer-support conversations", "period": "2020-2021 rollout", "sample": "Roughly 5,000 agents", "comparison": "Agents with access to an assistive generative-AI tool versus agents without access during a staged rollout", "method": "Field study using operational call, quality, and outcome data", "sourceName": "National Bureau of Economic Research", "sourceUrl": "https://www.nber.org/papers/w31161", "limitation": "One company and one assistive tool; effects varied substantially by worker skill and experience and should not be applied to other workflows.", "checkedAt": "2026-08-16" }, { "id": "nber-support-novice-gain", "statistic": "Productivity gain for the lowest-skilled and least-experienced support agents", "direction": "increase", "value": "35", "unit": "percent", "population": "Lowest-skilled and least-experienced agents in the support field study", "task": "Live customer-support conversations", "period": "2020-2021 rollout", "sample": "Subgroup of roughly 5,000 agents", "comparison": "AI-assisted subgroup performance versus the subgroup baseline or control condition", "method": "Heterogeneous-effect analysis within the field study", "sourceName": "National Bureau of Economic Research", "sourceUrl": "https://www.nber.org/papers/w31161", "limitation": "Subgroup result; the most experienced agents saw little benefit and in some cases a small negative effect.", "checkedAt": "2026-08-16" }, { "id": "mit-writing-time", "statistic": "Time required for selected professional writing tasks with ChatGPT", "direction": "decrease", "value": "40", "unit": "percent", "population": "College-educated professionals across writing-intensive occupations", "task": "Occupation-specific, incentivized professional writing assignments", "period": "2023 study", "sample": "444 participants", "comparison": "Random assignment to complete a second task with or without ChatGPT", "method": "Preregistered online randomized experiment", "sourceName": "Massachusetts Institute of Technology", "sourceUrl": "https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf", "limitation": "Short assigned writing tasks; the study did not measure full jobs, downstream acceptance, or later source-validation work.", "checkedAt": "2026-08-16" }, { "id": "mit-writing-quality", "statistic": "Independently rated quality of selected professional writing tasks with ChatGPT", "direction": "increase", "value": "18", "unit": "percent", "population": "College-educated professionals across writing-intensive occupations", "task": "Occupation-specific, incentivized professional writing assignments", "period": "2023 study", "sample": "444 participants", "comparison": "Random assignment to complete a second task with or without ChatGPT", "method": "Preregistered online randomized experiment with blinded output evaluation", "sourceName": "Massachusetts Institute of Technology", "sourceUrl": "https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf", "limitation": "Evaluator-rated assignment quality is not the same as acceptance, accuracy, or business impact in a live workflow.", "checkedAt": "2026-08-16" }, { "id": "hbs-bcg-consulting-speed", "statistic": "Completion speed on consulting tasks inside the tested AI capability frontier", "direction": "increase", "value": "25", "unit": "percent_greater_than", "population": "Boston Consulting Group consultants", "task": "Product innovation and strategy assignments inside the tested model's capabilities", "period": "2023 experiment; paper published later", "sample": "758 consultants", "comparison": "Randomly assigned AI-access groups versus a no-AI control group", "method": "Randomized laboratory-in-the-field experiment", "sourceName": "Harvard Business School", "sourceUrl": "https://www.hbs.edu/ris/download.aspx?name=24-013.pdf", "limitation": "Applies to tasks inside the tested capability frontier; a separate task outside it produced worse results.", "checkedAt": "2026-08-16" }, { "id": "hbs-bcg-consulting-quality", "statistic": "Human-rated quality on consulting tasks inside the tested AI capability frontier", "direction": "increase", "value": "40", "unit": "percent_greater_than", "population": "Boston Consulting Group consultants", "task": "Product innovation and strategy assignments inside the tested model's capabilities", "period": "2023 experiment; paper published later", "sample": "758 consultants", "comparison": "Randomly assigned AI-access groups versus a no-AI control group", "method": "Randomized laboratory-in-the-field experiment with expert-rated outputs", "sourceName": "Harvard Business School", "sourceUrl": "https://www.hbs.edu/ris/download.aspx?name=24-013.pdf", "limitation": "Task-specific rating; it does not support a general quality claim for all consulting or knowledge work.", "checkedAt": "2026-08-16" }, { "id": "hbs-bcg-outside-frontier-correctness", "statistic": "Likelihood of a correct answer on a consulting task outside the tested AI frontier", "direction": "decrease", "value": "19", "unit": "percentage_points", "population": "Boston Consulting Group consultants assigned the outside-frontier task", "task": "Difficult business problem requiring integration of quantitative and qualitative evidence", "period": "2023 experiment; paper published later", "sample": "Subset of 758 consultants", "comparison": "AI-access groups versus a no-AI control group", "method": "Randomized laboratory-in-the-field experiment", "sourceName": "Harvard Business School", "sourceUrl": "https://www.hbs.edu/ris/download.aspx?name=24-013.pdf", "limitation": "One deliberately selected outside-frontier task; the result establishes a boundary, not a universal AI error rate.", "checkedAt": "2026-08-16" }, { "id": "github-copilot-javascript-speed", "statistic": "Completion speed for a JavaScript HTTP-server task with GitHub Copilot", "direction": "increase", "value": "55", "unit": "percent", "population": "Professional developers familiar with JavaScript", "task": "Build an HTTP server in JavaScript to a shared specification", "period": "2022 experiment; page updated 2024", "sample": "95 developers", "comparison": "Random assignment to Copilot access or no Copilot access", "method": "Controlled experiment with automated correctness and completeness tests", "sourceName": "GitHub", "sourceUrl": "https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness/", "limitation": "One bounded task; it does not represent architecture, maintenance, debugging, security review, or long-running production delivery.", "checkedAt": "2026-08-16" }, { "id": "microsoft-three-firms-completed-tasks", "statistic": "Completed software-development tasks with access to an AI coding assistant", "direction": "increase", "value": "26.08", "unit": "percent", "population": "Software developers at Microsoft, Accenture, and an anonymous Fortune 100 company", "task": "Ordinary company software-development work", "period": "Field experiments published as a 2025 preprint", "sample": "4,867 developers", "comparison": "Random assignment to AI coding-assistant access or control within each company", "method": "Combined estimate from three randomized field experiments", "sourceName": "Microsoft Research", "sourceUrl": "https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/", "limitation": "Each company-level estimate was noisy; the combined preprint result may change with peer review or different tools and work settings.", "checkedAt": "2026-08-16" }, { "id": "metr-experienced-developers-time", "statistic": "Time required by experienced open-source developers when AI tools were allowed", "direction": "increase", "value": "19", "unit": "percent", "population": "Experienced open-source developers working in repositories they knew well", "task": "Real bug fixes and features in mature open-source repositories", "period": "February-June 2025", "sample": "16 developers completing 246 tasks", "comparison": "Random assignment of each task to AI-allowed or AI-disallowed conditions", "method": "Randomized controlled trial using real repository tasks", "sourceName": "METR", "sourceUrl": "https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/", "limitation": "Small specialist sample and early-2025 tools; many participants had limited experience with the primary AI interface used.", "checkedAt": "2026-08-16" }, { "id": "st-louis-fed-work-hours-saved", "statistic": "Self-reported generative-AI time savings as a share of all U.S. work hours", "direction": "decrease", "value": "1.6", "unit": "percent_of_work_hours", "population": "U.S. workers ages 18-64, including nonusers", "task": "Work respondents reported completing with generative-AI assistance", "period": "February, May, and August 2025 survey waves", "sample": "Pooled Real-Time Population Survey respondents", "comparison": "Respondents' estimate of extra hours needed to produce the same work without generative AI", "method": "National workforce survey and aggregate calculation", "sourceName": "Federal Reserve Bank of St. Louis", "sourceUrl": "https://www.stlouisfed.org/on-the-economy/2025/nov/state-generative-ai-adoption-2025", "limitation": "Self-reported counterfactual time rather than observed time logs; saved time may not become additional output or lower cost.", "checkedAt": "2026-08-16" }, { "id": "anthropic-conversation-estimated-time", "statistic": "Median estimated task-time savings in sampled Claude.ai conversations", "direction": "decrease", "value": "81", "unit": "percent", "population": "Tasks represented in sampled Claude.ai conversations", "task": "A range of tasks mapped to O*NET occupations", "period": "2025 research note", "sample": "100,000 Claude.ai conversations", "comparison": "Claude-estimated professional time without AI versus Claude-estimated time with AI", "method": "Privacy-preserving transcript analysis with model-generated time estimates", "sourceName": "Anthropic", "sourceUrl": "https://www.anthropic.com/research/estimating-productivity-gains", "limitation": "Both time values were estimated by Claude; the method cannot see off-chat review, refinement, rejected outputs, or later work.", "checkedAt": "2026-08-16" } ] }