{"key":"ai-roi-calculator","updated":"2026-11-16","count_label":"7 peer-reviewed or working-paper studies of AI productivity","data_note":"The calculator applies the time saving measured in one published study to your inputs. Where a study reports more output per hour rather than time saved, we convert it (time saved = 1 minus 1 \/ (1 + gain)); the conversion is shown for every study. Studies measured specific tasks with specific tools, so treat the result as a scenario, not a forecast. Two studies (St. Louis Fed, METR) give a share of all hours or a slowdown; we use them as stated.","glance":[{"label":"Studies in the calculator","value":"7","sub":"2023 to 2025"},{"label":"Largest time saving","value":"55.8%","sub":"GitHub Copilot RCT (2023)"},{"label":"Smallest result","value":"-19.0%","sub":"Experienced developers (METR 2025): AI made tasks slower"},{"label":"Randomised trials","value":"5","sub":"of 7 studies"}],"charts":[{"id":"saving","type":"hbars","title":"Time saved by study","label":"Share of task time saved, %","data":[["Copilot RCT",55.8],["Writing tasks",40],["BCG consultants",25.1],["Dev field trials",20.7],["Support agents",12.3],["US workers (all)",5.4],["METR developers",-19]],"fmt":"n","source_label":"the linked studies; conversions by i10X"}],"sections":[{"id":"calc","type":"calc","title":"Calculate your AI return","note":"Pick a study, then enter your own numbers. Everything is calculated in your browser.","studies":[{"id":"nber-support","short":"Customer support (NBER 2023)","title":"Generative AI at Work","authors":"Brynjolfsson, Li, Raymond","year":2023,"design":"Staggered roll-out at one firm","setting":"Customer support agents","task":"Resolving customer chats","effect":"+14% issues resolved per hour (+34% for novice and low-skilled agents)","saving":0.123,"calc":"1 - 1\/1.14","n":"5,179 agents","url":"https:\/\/www.nber.org\/papers\/w31161"},{"id":"copilot-rct","short":"GitHub Copilot RCT (2023)","title":"The Impact of AI on Developer Productivity: Evidence from GitHub Copilot","authors":"Peng, Kalliamvakou, Cihon, Demirer","year":2023,"design":"Randomised controlled trial","setting":"Freelance software developers","task":"Write an HTTP server in JavaScript","effect":"Task completed 55.8% faster","saving":0.558,"calc":"as reported","n":"95 developers","url":"https:\/\/arxiv.org\/abs\/2302.06590"},{"id":"bcg-frontier","short":"BCG consultants (HBS 2023)","title":"Navigating the Jagged Technological Frontier","authors":"Dell'Acqua et al.","year":2023,"design":"Randomised experiment","setting":"BCG consultants","task":"Realistic consulting tasks inside the AI frontier","effect":"12.2% more tasks, 25.1% faster, over 40% higher quality","saving":0.251,"calc":"as reported (speed)","n":"758 consultants","url":"https:\/\/www.hbs.edu\/faculty\/Pages\/item.aspx?num=64700"},{"id":"noy-zhang","short":"Writing tasks (Science 2023)","title":"Experimental evidence on the productivity effects of generative artificial intelligence","authors":"Noy, Zhang","year":2023,"design":"Randomised experiment","setting":"College-educated professionals","task":"Writing tasks (press releases, reports, emails)","effect":"Time taken fell by 40%, output quality rose by 18%","saving":0.4,"calc":"as reported","n":"453 participants","url":"https:\/\/www.science.org\/doi\/10.1126\/science.adh2586"},{"id":"metr-oss","short":"Experienced developers (METR 2025)","title":"Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity","authors":"Becker, Rush, Barnes, Rein (METR)","year":2025,"design":"Randomised controlled trial","setting":"Experienced open-source developers on their own repositories","task":"Real issues in mature codebases","effect":"Tasks took 19% longer with AI tools","saving":-0.19,"calc":"as reported (slowdown)","n":"16 developers, 246 tasks","url":"https:\/\/arxiv.org\/abs\/2507.09089"},{"id":"cui-fieldexp","short":"Developers, 3 field trials (2024)","title":"The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers","authors":"Cui, Demirer, Jaffe, Musolff, Peng, Salz","year":2024,"design":"Three randomised field experiments","setting":"Developers at Microsoft, Accenture and a Fortune 100 firm","task":"Everyday software work","effect":"+26.08% completed tasks","saving":0.207,"calc":"1 - 1\/1.2608","n":"4,867 developers","url":"https:\/\/papers.ssrn.com\/sol3\/papers.cfm?abstract_id=4945566"},{"id":"stlfed","short":"US workers, all work (St. Louis Fed 2025)","title":"The Impact of Generative AI on Work Productivity","authors":"Bick, Blandin, Deming (St. Louis Fed)","year":2025,"design":"Nationally representative survey","setting":"US workers who use generative AI","task":"All work","effect":"Users saved 5.4% of work hours","saving":0.054,"calc":"as reported (share of all hours)","n":"Real-Time Population Survey","url":"https:\/\/www.stlouisfed.org\/on-the-economy\/2025\/feb\/impact-generative-ai-work-productivity"}],"default":"nber-support"},{"id":"c1","type":"charts","title":"What the studies found","charts":["saving"],"captions":[]},{"id":"tbl","type":"table","table":"studies","title":"The studies behind the numbers","note":"Effect: what the study reports. Time saved: the figure the calculator uses."}],"tables":[{"id":"studies","columns":[{"k":"study","label":"Study","type":"text"},{"k":"authors","label":"Authors","type":"text"},{"k":"year","label":"Year","type":"text"},{"k":"design","label":"Design","type":"text"},{"k":"who","label":"Who","type":"text"},{"k":"sample","label":"Sample","type":"text"},{"k":"effect","label":"Effect reported","type":"text"},{"k":"saving","label":"Time saved used","type":"text"},{"k":"source_url","label":"Source","type":"src"}],"rows":[{"study":"Generative AI at Work","authors":"Brynjolfsson, Li, Raymond","year":"2023","design":"Staggered roll-out at one firm","who":"Customer support agents","sample":"5,179 agents","effect":"+14% issues resolved per hour (+34% for novice and low-skilled agents)","saving":"12.3%","source_url":"https:\/\/www.nber.org\/papers\/w31161"},{"study":"The Impact of AI on Developer Productivity: Evidence from GitHub Copilot","authors":"Peng, Kalliamvakou, Cihon, Demirer","year":"2023","design":"Randomised controlled trial","who":"Freelance software developers","sample":"95 developers","effect":"Task completed 55.8% faster","saving":"55.8%","source_url":"https:\/\/arxiv.org\/abs\/2302.06590"},{"study":"Navigating the Jagged Technological Frontier","authors":"Dell'Acqua et al.","year":"2023","design":"Randomised experiment","who":"BCG consultants","sample":"758 consultants","effect":"12.2% more tasks, 25.1% faster, over 40% higher quality","saving":"25.1%","source_url":"https:\/\/www.hbs.edu\/faculty\/Pages\/item.aspx?num=64700"},{"study":"Experimental evidence on the productivity effects of generative artificial intelligence","authors":"Noy, Zhang","year":"2023","design":"Randomised experiment","who":"College-educated professionals","sample":"453 participants","effect":"Time taken fell by 40%, output quality rose by 18%","saving":"40.0%","source_url":"https:\/\/www.science.org\/doi\/10.1126\/science.adh2586"},{"study":"Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity","authors":"Becker, Rush, Barnes, Rein (METR)","year":"2025","design":"Randomised controlled trial","who":"Experienced open-source developers on their own repositories","sample":"16 developers, 246 tasks","effect":"Tasks took 19% longer with AI tools","saving":"-19.0%","source_url":"https:\/\/arxiv.org\/abs\/2507.09089"},{"study":"The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers","authors":"Cui, Demirer, Jaffe, Musolff, Peng, Salz","year":"2024","design":"Three randomised field experiments","who":"Developers at Microsoft, Accenture and a Fortune 100 firm","sample":"4,867 developers","effect":"+26.08% completed tasks","saving":"20.7%","source_url":"https:\/\/papers.ssrn.com\/sol3\/papers.cfm?abstract_id=4945566"},{"study":"The Impact of Generative AI on Work Productivity","authors":"Bick, Blandin, Deming (St. Louis Fed)","year":"2025","design":"Nationally representative survey","who":"US workers who use generative AI","sample":"Real-Time Population Survey","effect":"Users saved 5.4% of work hours","saving":"5.4%","source_url":"https:\/\/www.stlouisfed.org\/on-the-economy\/2025\/feb\/impact-generative-ai-work-productivity"}],"limit":20,"rowhead":"study"}],"preview":{"big":"-19% to 56%","big_label":"time saved across published AI studies","list":[["GitHub Copilot RCT (2023)","55.8%"],["Writing tasks (Science 2023)","40.0%"],["BCG consultants (HBS 2023)","25.1%"],["Developers, 3 field trials (2024)","20.7%"]]},"stored_at":"2026-11-16 16:29:05"}