Companies reported early, large gains in high-volume work with fairly clear answers, including customer inquiries, HR questions and contract review. These tasks were easy to measure, so companies could assess results before investing further.
Companies managed quality risks by having staff check AI drafts before they reached customers or patients. Klarna cut customer-service staff while relying on its AI assistant, then brought human agents back because of service-quality concerns.
Shinhan Bank reorganized its banking knowledge before adding a generative AI assistant. Walmart, POSCO and Kurly also cleaned, connected and shared data before choosing a model.
A Canadian tribunal held Air Canada liable for refund terms that its website chatbot had stated wrongly, and rejected the argument that the chatbot was responsible for its own words. McDonald's ended its AI voice-ordering test at drive-throughs in 2024 while order accuracy remained a problem.
Zillow bought homes at prices estimated by its model. When the market changed, forecast errors led to write-downs and the company closed its home-buying unit. MD Anderson spent tens of millions of dollars on an AI project with weak oversight that was never used to treat patients.
Amazon dropped a résumé-screening model that learned from past hiring records to rate applications associated with women lower. In hiring and evaluation, companies faced questions about bias in training data and whether they could explain decisions that affected people’s lives.
JPMorgan let staff choose whether to use its AI platform. Moderna built habits with an internal chatbot before expanding AI across the company and merging HR with technology. Shopify made AI use part of performance reviews, while Musinsa promoted AI skills across its workforce.
On Naver's shopping platform, the sales effect of AI shopping features was larger for small sellers than for large ones. Small Korean manufacturers used government smart-factory programs to cover much of the up-front cost of AI-based quality control.
Coca-Cola’s AI-generated holiday ad drew widespread criticism for lacking warmth. Duolingo’s public AI-first memo also drew immediate criticism. Brands that relied on emotion and trust faced limits from audience reactions.
Some companies used AI in their own operations before selling it to other businesses. Octopus Energy, Samsung SDS and Salesforce sold outside customers tools they had tested in their own customer service or office work.
issues resolved per hour by agents with AI assistance (novice and low-skilled agents: about +30%)
Stanford and MIT economists studied a GPT-based chat assistant introduced to 5,172 support agents at a Fortune 500 business-software company, mostly between late 2020 and early 2021. Issues resolved per hour rose by 15% on average and about 30% for the least experienced and lowest-skilled agents, whose speed and quality both improved. The most experienced agents gained little speed and saw a small decline in quality. These results came from the peer-reviewed 2025 study, which followed a 2023 working paper reporting gains of 14% and 34% among 5,179 agents.
The tool passed experienced agents' knowledge to newer staff, so teams with more inexperienced staff gained more.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,172 support agents and published their peer-reviewed findings in volume 140(2) of The Quarterly Journal of Economics. Issues resolved per hour rose 15% on average and about 30% for the least experienced and lowest-skilled agents, while quality declined slightly for the most skilled. Newer agents were less likely to leave, and customers were more polite.
The researchers first released their working paper in April 2023 and revised it in November 2023. They reported a 14% average productivity gain among 5,179 agents and a 34% gain for novice and low-skilled staff, with little effect on experienced, highly skilled agents. Customer sentiment and staff retention also improved.
Stanford HAI reported a 14% productivity gain in the call-center study, concentrated among less experienced agents. Agents with two months of tenure who used the assistant performed like agents with six months who did not. Experienced agents saw little benefit.
faster completion of tasks within AI's capabilities (correct answers fell 19 points on a task outside them)
Researchers from Harvard Business School, other universities and BCG randomly assigned 758 BCG consultants to no AI, GPT-4, or GPT-4 with a short prompting introduction in a preregistered experiment. On 18 consulting tasks within the model’s capabilities, AI users completed 12.2% more tasks, worked 25.1% faster and received quality scores about 32% higher on average. On a business-analysis task outside those capabilities, AI users were 19 percentage points less likely to answer correctly, at about 60–71% versus 84.5% without AI. The 2023 working paper reported a quality gain of more than 40%, while the peer-reviewed 2026 study reported about 32%.
The same AI tool improved some consulting tasks and reduced accuracy on another. Skilled professionals had difficulty recognizing which tasks exceeded its capabilities.
Fabrizio Dell’Acqua and colleagues published their peer-reviewed study online in March 2026. On tasks within the model’s capabilities, AI users completed 12.2% more tasks, worked 25.1% faster and improved quality by about 32% on average. Correct answers fell by 19 percentage points on the task outside those capabilities. Less skilled participants gained the most.
The September 2023 Harvard Business School study assigned 758 knowledge workers to three groups. AI users worked faster and completed more tasks within the model’s capabilities, but were less likely to solve a task outside them. The research was later published in Organization Science.
BCG reported that its consultants scored about 40% higher with GPT-4 on creative product tasks. On business problem-solving tasks, scores were 23% lower because many accepted incorrect output. Ideas from GPT-4 users were 41% less diverse.
shorter time to finish one programming task with an AI coding assistant
Researchers from Microsoft Research, GitHub and MIT Sloan recruited 95 freelance developers through Upwork in May–June 2022 and randomly split them into a Copilot group (45) and a control group (50); 35 in each group completed the task. Asked to write an HTTP server in JavaScript, the Copilot group took 71 minutes on average against 161 minutes, a 55.8% reduction (95% confidence interval 21–89%). Less experienced developers benefited more. Copilot is made by GitHub, a Microsoft subsidiary, and three of the four authors worked at Microsoft or GitHub.
Developers saved substantial time on one well-defined coding task. Later studies of large existing codebases found much smaller gains or slowdowns in that different setting.
Sida Peng, Eirini Kalliamvakou, Peter Cihon and Mert Demirer recruited 95 developers through Upwork and assigned 45 to Copilot and 50 to a control group. In each group, 35 completed the HTTP-server task, with the Copilot group taking 55.8% less time. Its success rate was 7 percentage points higher, a difference that was not statistically significant.
GitHub reported that 78% of the Copilot group and 70% of the control group completed the task. The Copilot group finished about 55% faster, taking 1 hour 11 minutes versus 2 hours 41 minutes. In a separate survey of more than 2,000 Copilot users, most said the tool helped them work faster and stay focused.
share of manufacturing SMEs naming initial cost as the biggest barrier to AI adoption
The Ministry of SMEs and Startups cited a Korea Federation of SMEs survey when it announced its AI-based Smart Manufacturing Innovation 3.0 strategy on 24 October 2025. Manufacturing SMEs named initial cost (44.2%) as their biggest barrier, followed by staff shortages (20.5%) and lack of strategy (14.9%), while 72.3% wanted direct financial support. The government put AI adoption at about 1% and set targets of 10% adoption and 12,000 AI-centered smart factories by 2030. The survey date and sample size were not reported.
Manufacturers cited initial cost more often than staff shortages or lack of strategy. Government measures included subsidies and subscription-based smart factories to reduce that initial burden.
Money Today reported that manufacturing SMEs cited initial cost (44.2%), staff shortages (20.5%) and lack of strategy (14.9%) as barriers to AI, while 72.3% wanted direct financial aid. Government measures included subscription-based smart factories and a one-stop support platform. The government aimed to raise AI adoption from about 1% to 10%.
Herald Business reported that a Korea Federation of SMEs survey found initial cost was manufacturers’ biggest barrier to AI and about seven in ten wanted direct funding. Government targets included 12,000 AI smart factories, 10% AI adoption, 500 specialist firms and 1,000 shared datasets.
Financial News reported government targets of 12,000 AI-centered smart factories and an increase in AI adoption from 1% to 10% by 2030. It also described support for individual companies under the smart-factory AI program.
longer task completion time when AI tools were allowed (developers had expected 24% faster)
From February to June 2025, nonprofit METR randomly allowed or prohibited AI on 246 real tasks completed by 16 experienced developers in large open-source projects they had maintained for years. When AI was allowed, developers mostly used Cursor Pro with Claude 3.5/3.7 Sonnet and took 19% longer, with a 95% confidence interval of +2% to +39%. They had expected to work 24% faster and afterward still believed they had been 20% faster. A follow-up begun in August 2025 with 57 developers produced estimates whose confidence intervals included zero, and METR said in February 2026 that selection effects made the data unreliable.
Experienced developers working in familiar codebases felt faster with AI even as measured completion time increased. Results from early-2025 tools may differ from those of later tools.
Joel Becker, Nate Rush, Elizabeth Barnes and David Rein found a 19% increase in completion time in a trial with 16 developers and 246 tasks. Developers had expected a 24% speed-up and perceived a 20% speed-up afterward, while machine-learning and economics experts had forecast gains of 38–39%. The researchers examined 20 possible explanations and judged the slowdown unlikely to be mainly an artifact of the study design.
METR studied projects with an average of more than 22,000 stars and more than a million lines of code, using tasks that took roughly two hours. It noted the small developer sample, focus on a single domain and possible learning effects. METR later described these results as out of date and released follow-up data.
METR began a follow-up in August 2025 with 57 developers and more than 800 tasks. It estimated speed-ups of 18% for returning developers and 4% for new participants, with both confidence intervals including zero. As developers increasingly refused to work without AI, METR judged the data unreliable and announced a new study design.
measured effect on earnings and recorded hours two years after ChatGPT's launch (effects above 2% ruled out)
Anders Humlum of the University of Chicago and Emilie Vestergaard of the University of Copenhagen worked with Statistics Denmark on representative surveys of 11 occupations highly exposed to AI chatbots. Their late-2024 survey covered 25,000 workers at 7,000 workplaces and was linked to official earnings, hours and occupation records through December 2024. Where employers encouraged use and provided enterprise chatbots and training, 93% had used chatbots, 28% used them daily and 19% reported saving more than an hour a day. Comparing adopters with non-adopters before and after chatbots arrived, the researchers found no effect on earnings or recorded hours, although work changed as staff generated content, reviewed AI output and integrated AI.
Workers' reported time savings did not change pay or hours within two years. Most users said they spent the saved time on other tasks, changing how work was organised.
The researchers first released their working paper in May 2025 and revised it in March 2026. Using linked survey and administrative data, they compared changes for adopters and non-adopters and ruled out average effects on earnings and hours larger than 2%. About 85% of users reported spending saved time on other tasks, and some moved into better-paid occupations.
The Becker Friedman Institute reported no effect on earnings or recorded hours two years after ChatGPT’s launch. The finding held for intensive users, early adopters and workplaces that invested heavily. Chatbot adoption was linked to changes in tasks and occupations without net changes in pay or hours.
weekly completed tasks among developers who used the AI coding assistant (pooled estimate, standard error about 10 points)
Economists from Princeton, MIT, Wharton and Microsoft analyzed three randomized experiments at Microsoft, Accenture and an unnamed Fortune 100 electronics manufacturer from September 2022 to May 2023. A random subset of 4,867 developers received GitHub Copilot access during normal work, and users completed 26.08% more weekly tasks (standard error 10.3%), made 13.55% more commits and ran 38.38% more builds across the experiments. Recent hires and junior developers used the tool more and gained more, while senior developers had no statistically significant gain. Two of the six authors worked at Microsoft, which owns GitHub.
Company developers gained less than in the single-task lab trial, with more uncertainty in results. Less experienced developers gained most.
Kevin Zheyuan Cui, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng and Tobias Salz reported the experiments in February 2025. The trials lasted two to eight months before all groups received access. Across the three trials, completed tasks rose 26.08%, commits 13.55% and builds 38.38%. The researchers said each trial alone had substantial statistical uncertainty and that initial uptake was low.
The researchers registered the three-company experiments with the American Economic Association after they had run. The experiments took place from September 2022 to May 2023 and assigned access randomly by team or individual at each company. Primary measures included pull requests, commits and builds.