When does AI make work slower or worse?
On tasks the AI handles poorly, controlled studies find results 14 to 25 percent worse: consultants using GPT-4 gave fewer correct answers, and experienced developers took 19 percent longer. No study measures how often this happens across all AI-assisted work.
- Medium confidence
- 3 primary sources
- 2025 to 2026
- Last checked
Every study on one scale
Chart values as text
- Scale: how much worse (points fewer correct, or % more time). Shaded range: 14–25.
- Dell'Acqua et al., BCG (2026): 13.9–24.5 points fewer correct
- METR (2025): 19% more time
- UK DBT, Excel task (2025): about 22% more time
Why medium confidence: Three independent creators with different designs each found a penalty in this band on some tasks. Each measures a different outcome, one study has only 3 people per group, and none answers how often it happens across all work.
Sources
| Study | Creator | Year | Sample | Method | Figure | Link |
|---|---|---|---|---|---|---|
| Navigating the Jagged Technological Frontier (Organization Science) | Fabrizio Dell'Acqua, Ethan Mollick, Karim R. Lakhani and colleagues, with Boston Consulting Group | 2026 | 758 BCG consultants; the outside-the-frontier task was run with a subset | Preregistered randomized experiment. One business case was chosen so GPT-4 would give a plausible but wrong answer. | 84.5% correct without AI, 70.6% and 60% with AI: 13.9 and 24.5 points fewer correct answers, while spending less time | doi.org ↗ |
| Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity | Joel Becker, Nate Rush, Beth Barnes and David Rein (METR) | 2025 | 16 experienced developers, 246 tasks in mature open-source projects | Randomized at the task level, February to June 2025, with screen recordings | 19% more time per task with AI; developers accepted under 44% of AI suggestions and spent more time prompting, waiting and reviewing “we find that allowing AI actually increases completion time by 19%” | arxiv.org ↗ |
| The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade | UK Department for Business and Trade | 2025 | 11 civil servants observed, 3 per group on the Excel task | Timed and blind-graded fictional tasks with matched colleagues as controls, late 2024. The authors warn the samples are very small. | Excel analysis took 25:01 with Copilot against 20:33 without, with lower accuracy (1.5 against 2.7 out of 5) “M365 Copilot users completed Excel data analysis more slowly and to a worse quality” | assets.publishing.service.gov.uk ↗ |
How much of the time AI saves goes into fixing its output?
Nearly 40 percent, by AI users' own account in one vendor survey: Workday's November 2025 survey of 3,200 AI users at large firms. Only 14% said they consistently get clearly positive results. In a 2024 Upwork survey, 77% of employees who use AI said it had added to their workload in some way, and 39% spend more time reviewing AI output.
| Study | Creator | Year | Sample | Method | Figure | Link |
|---|---|---|---|---|---|---|
| Beyond Productivity: Measuring the Real Value of AI | Workday, survey by Hanover Research | 2026 | 3,200 full-time employees who use AI, at organizations with $100 million or more in revenue | Online survey, November 2025, self-reported. Read in Workday's press release; the full report is gated. Workday sells AI products. | Nearly 40% of AI time savings lost to rework; 14% consistently get clear net-positive results “Nearly 40% of AI time savings are lost to rework” | newsroom.workday.com ↗ |
| From Burnout to Balance: AI-Enhanced Work Models | Upwork Research Institute with Workplace Intelligence | 2024 | 2,500 respondents in the US, UK, Australia and Canada, including 625 full-time employees | Online survey, April to May 2024, self-reported. Read in Upwork's press release. Upwork sells freelance talent. | 77% of employees who use AI say it added to their workload; 39% spend more time reviewing AI content “77% of employees using AI say these tools have added to their workload” | investors.upwork.com ↗ |
Does AI make software delivery less stable?
Google's DORA surveys link wider AI adoption to less stable software delivery, in both 2024 and 2025. In 2025 the link to throughput turned positive while the stability link held. These are statistical associations from surveys of technology professionals and do not show cause.
| Study | Creator | Year | Sample | Method | Figure | Link |
|---|---|---|---|---|---|---|
| 2025 DORA Report: State of AI-assisted Software Development | Google Cloud DORA | 2025 | Nearly 5,000 technology professionals | Annual survey with statistical modeling. Read in Google Cloud's announcement posts for 2024 and 2025. | AI adoption linked to lower software delivery stability; 30% report little or no trust in AI-generated code “AI adoption does continue to have a negative relationship with software delivery stability” | cloud.google.com ↗ |
Numbers you may have seen
These figures circulate widely. Here is what the original source says.
- Upwork: 77% of workers say AI has decreased their productivity. The survey says 77% of employees who use AI said it added to their workload, in a group of 625 full-time employees. It does not measure productivity.
- Workday: workers lose 40% of their time fixing AI. The survey says nearly 40% of the time AI saves is lost to rework. As a share of the whole work week it is far smaller.
- AI makes consultants 19% worse (Dell'Acqua et al.). The drop applies to one task chosen to sit outside GPT-4's abilities. On 18 other tasks the same consultants did better with AI.
- DORA: every 25% increase in AI adoption cuts delivery throughput by 1.5% and stability by 7.2%. These are survey associations and do not show cause. In 2025 the throughput link turned positive; only the stability link held.
More from What AI Saves
- F-083–5.4%How much time does AI save knowledge workers?
- F-0925–40%How much faster do people finish writing tasks with AI, and is the work better?
- F-1014–15%How much more do customer support agents get done with AI?
- F-1121–26%How much faster do developers code with AI assistants?
- F-1227–46%Does AI help less-experienced workers more than experienced ones?
The whole chapter · Chapter poster · All of The AI at Work Fact Book