The public debate about AI and work runs on a binary that the evidence does not support: either the machines take the jobs or they do not. The better-measured reality is more granular and more useful. AI acts on tasks, not jobs, and it acts on them unevenly.
The productivity gains are real and large
The experimental record is unusually strong for such a young technology. In a field experiment at a large customer-service operation, Brynjolfsson, Li and Raymond (2025) studied 5,172 agents given a generative AI assistant and found average productivity, measured as issues resolved per hour, rose about 15 percent, with the gains concentrated among novice and lower-skilled workers. A pre-registered field experiment with 758 Boston Consulting Group consultants found that on tasks within the tool's ability, those using GPT-4 completed 12.2 percent more tasks, 25.1 percent faster, and at higher quality (Dell'Acqua and colleagues, 2025). For software specifically, a controlled experiment found developers with GitHub Copilot finished a programming task about 55.8 percent faster than a control group (Peng and colleagues, 2023). These are not speculative projections. They are measured effects in real work.
But the frontier is jagged, and the losses are concentrated
The same BCG study named the crucial complication: a jagged technological frontier. On tasks just outside the model's capability, consultants using it did worse, because the tool was confidently wrong and they trusted it. AI does not raise all work uniformly. It lifts some tasks sharply and quietly degrades others, and the boundary is not obvious from the outside.
The distributional picture at the level of jobs is early but pointed. Brynjolfsson, Chandar and Chen (2025) found that early-career workers aged 22 to 25 in the occupations most exposed to AI saw a relative employment decline of about 16 percent, while experienced workers in the same occupations and workers in less-exposed fields held steady or grew. The labor-market adjustment is showing up first in entry-level hiring, which has uncomfortable implications for how the next generation is supposed to acquire the experience that later protects it.
The organizing variable is human agency
If AI operates task by task, the design question becomes which tasks to hand over and which to keep. Shao and colleagues (2025) built a large audit, the WORKBank, crossing what workers actually want automated across 844 tasks and 104 occupations with what AI can technically do, and introduced a Human Agency Scale to describe the preferred level of human involvement for each task. That is the right unit of analysis. Some tasks people are glad to offload; others they want to keep, not out of nostalgia but because judgment, accountability, or relationship lives there. Automating the second kind because it is technically possible is how organizations destroy value while chasing efficiency.
The macro numbers point to churn, not disappearance
At the aggregate level, the World Economic Forum (2025) projects that 22 percent of jobs will be disrupted by 2030, with 170 million roles created and 92 million displaced, a net gain of about 78 million, and estimates that nearly 40 percent of required skills will change. The headline is churn, not extinction: a large reshuffling of tasks and skills rather than a simple subtraction of workers. That is a management challenge, redeploy and reskill, not an apocalypse to brace for.
The practical read
Stop asking whether AI will replace your people and start mapping which of their tasks it should touch. The evidence shows big productivity gains that favor novices on tasks within the frontier (Brynjolfsson, Li and Raymond, 2025; Dell'Acqua et al., 2025), real degradation just beyond it (Dell'Acqua et al., 2025), and early employment pressure concentrated on entry-level roles (Brynjolfsson, Chandar and Chen, 2025). Use a task-level lens and an explicit choice about where humans keep agency (Shao et al., 2025), and treat the macro shift as churn to be managed rather than a wave to be survived (WEF, 2025). The firms that win will not be the ones that automate the most. They will be the ones that automate the right tasks and keep humans on the ones where judgment still pays.
Sources
- Brynjolfsson, Li and Raymond (2025). Generative AI at Work. Across 5,172 customer-service agents, an AI assistant raised productivity about 15 percent, concentrated among novices. https://academic.oup.com/qje/article/140/2/889/7990658
- Dell'Acqua and colleagues (2025). Navigating the Jagged Technological Frontier. BCG consultants using GPT-4 did more, faster, higher-quality work within the frontier but worse just beyond it. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Peng and colleagues (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. Developers with Copilot finished a task about 55.8 percent faster. https://arxiv.org/abs/2302.06590
- Brynjolfsson, Chandar and Chen (2025). Canaries in the Coal Mine? Early-career workers in AI-exposed occupations saw a roughly 16 percent relative employment decline. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/
- Shao and colleagues (2025). Future of Work with AI Agents. Introduces a Human Agency Scale mapping which tasks workers want automated versus kept. https://arxiv.org/abs/2506.06576
- World Economic Forum (2025). Future of Jobs Report 2025. Projects 22 percent of jobs disrupted by 2030 with a net gain of about 78 million roles and 40 percent of skills changing. https://www.weforum.org/publications/the-future-of-jobs-report-2025/