F-12 · What AI Saves
Does AI help less-experienced workers more than experienced ones?
Mostly yes, on well-defined tasks: the newest or weakest workers gained about 27 to 46 percent in support, consulting and software studies, against 0 to 17 percent for the strongest. On open-ended business decisions the pattern reversed: in Kenya, strong entrepreneurs gained from an AI adviser while weak ones did nearly 10 percent worse.
27–46% gain for the newest or weakest workers; the strongest gain 0 to 17%
- Medium confidence
- 5 primary sources
- 2023 to 2026
- Last checked
Every study on one scale
Shaded: 27–46% One figure A range in one study
Chart values as text
- Scale: % gain for the newest or weakest group. Shaded range: 27–46%.
- Brynjolfsson et al., support (2025): 30–46% (top skill: about 0)
- Dell'Acqua et al., BCG (2023): 43% (top half 17%)
- Cui et al., software (2026): 27–39% (long tenure 8–13%)
Why medium confidence: Five independent teams in different settings find bigger gains for weaker or newer workers. Each measures a different outcome, the software gaps are not statistically significant, and a randomized trial on open-ended business decisions finds the reverse.
Studies without a figure on this scale are listed in the table only.
Sources
| Study | Creator | Year | Sample | Method | Figure | Link |
|---|---|---|---|---|---|---|
| Generative AI at Work (Quarterly Journal of Economics) | Erik Brynjolfsson (Stanford), Danielle Li and Lindsey Raymond (MIT) | 2025 | 5,172 customer support agents at one Fortune 500 software firm | Staggered rollout, compared before and after, split by earlier skill and by tenure. The tenure figure is from the 2023 NBER version. | Less skilled and less experienced agents +30%; lowest-skilled fifth +36%; agents with under a month of tenure about +46%; no significant change for the most skilled “Less skilled and less experienced workers improve significantly across all productivity measures” | doi.org ↗ |
| Navigating the Jagged Technological Frontier (HBS Working Paper 24-013) | Fabrizio Dell'Acqua, Ethan Mollick, Karim R. Lakhani and colleagues, with Boston Consulting Group | 2023 | 758 BCG consultants | Preregistered randomized experiment. Gains compare AI users with their own earlier scores on a baseline task. | Consultants below the median improved 43%, those above 17% “those below the average performance threshold increasing by 43% and those above increasing by 17%” | hbs.edu ↗ Archived copy ↗ HBS no longer hosts this version of the PDF; the archived copy is the 2023 working paper we read. |
| The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers (Management Science) | Kevin Zheyuan Cui, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng and Tobias Salz | 2026 | 4,867 developers at Microsoft, Accenture and a Fortune 100 electronics maker | Three randomized company experiments, split by tenure and job level. The splits are not statistically significant. | Short-tenure developers +27% to 39% completed tasks; long-tenure +8% to 13% “less experienced developers had higher adoption rates and greater productivity gains.” | doi.org ↗ |
| Experimental evidence on the productivity effects of generative artificial intelligence (Science) | Shakked Noy and Whitney Zhang (MIT) | 2023 | 453 college-educated professionals | Preregistered online randomized experiment on writing tasks | The gap between stronger and weaker writers halved: low scorers improved in grade and speed, high scorers kept their grade and got faster “ChatGPT compresses the productivity distribution by benefiting low-ability workers more.” | science.org ↗ |
| Generative AI in Action: Field Experimental Evidence from Alibaba's Customer Service Operations | Xiao Ni, Yiwei Wang, Tianjun Feng, Lauren Xiaoyuan Lu, Yitong Wang and Congyi Zhou | 2026 | 5,940 new customer service agents, about 2.56 million chats | Four-week randomized trial run by Alibaba, split by earlier customer ratings | Ratings for the lowest-performing fifth rose 0.81 points on a 5-point scale and fell 0.28 for the top fifth “Lower-performing agents benefit the most, indicating that gen AI can narrow performance gaps.” | arxiv.org ↗ |
When does AI help strong performers more than weak ones?
On open-ended decisions, where people must judge which advice fits. In a five-month trial with 640 entrepreneurs in Kenya who got a GPT-4 business mentor on WhatsApp, low performers' revenue fell about 9% while high performers' rose about 18%, a gain the authors treat as tentative. They trace the gap to weaker owners acting on generic advice such as price cuts.
| Study | Creator | Year | Sample | Method | Figure | Link |
|---|---|---|---|---|---|---|
| The Uneven Impact of Generative AI on Entrepreneurial Performance: Evidence from a Field Experiment in Kenya (HBS Working Paper 24-042) | Nicholas G. Otis, Rowan Clarke, Solène Delecourt, David Holtz and Rembrand Koning | 2026 | 640 entrepreneurs in Kenya, five months | Randomized field experiment: a GPT-4 mentor on WhatsApp or a standard business guide | No average effect; low performers about 9% lower revenue, high performers about 18% higher (p = 0.07) “low performers did nearly 10% worse due to the AI assistant” | hbs.edu ↗ Archived copy ↗ |
Numbers you may have seen
These figures circulate widely. Here is what the original source says.
- AI mostly helps top scientists (Toner-Rodgers, MIT, 2024). MIT said in May 2025 that it has no confidence in the paper's data and asked for it to be withdrawn.
- Below-average BCG consultants improved 43% more than the control group. The 43% and 17% compare consultants using AI with their own scores on an earlier task without AI.
- Novice support agents gained 34% with AI. That is the 2023 working-paper figure. The peer-reviewed 2025 version says 30% for less skilled and less experienced agents and 36% for the lowest-skilled fifth.
- AI levels the playing field for every worker. It holds in studies of well-defined tasks. In the Kenyan trial on open-ended business decisions, AI widened the gap.
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-1314–25%When does AI make work slower or worse?
The whole chapter · Chapter poster · All of The AI at Work Fact Book