There are two theories of what AI agents are for. One says they replace workers. The other says they amplify them. The distinction is not academic, because it determines where a firm points its money, and a growing body of evidence suggests a lot of that money is currently pointed the wrong way.
What people want handed over, and what they do not
Start with demand, since it is measurable. Shao and colleagues (2025) audited what U.S. workers actually want from AI, mapping 844 tasks across 104 occupations from a survey of 1,500 professionals against expert assessments of technical capability. The pattern is consistent. Workers want AI to take over low-value, repetitive tasks, the drudgery that fills a day without defining the job. On high-value tasks they want something different, which the authors call shared agency: the machine executes, the human keeps judgment and responsibility.
That is a preference for augmentation over substitution, and it is not vague sentiment. It sorts cleanly by task. The revealing finding is the mismatch. When Shao and colleagues (2025) line up worker desire against where capability and, by implication, investment are heading, a meaningful share of agent development is flowing into zones where workers do not want automation, while some of the tasks people would happily offload go unaddressed. Capital is arriving, and it is arriving in the wrong place.
The evidence favors amplification
The productivity studies point the same direction, because their gains come from augmentation, not from removing the human. Brynjolfsson, Li and Raymond (2025), in a field study of 5,172 customer-service agents, found productivity up about 15 percent on average, with the largest gains, up to roughly 34 percent, among novice and lower-skilled workers, and close to nothing for the most experienced. The mechanism was diffusion: the AI spread the tacit know-how of the best agents to the rest. That is amplification of a workforce, not a replacement of it, and notably the study also found improved customer sentiment and higher employee retention, outcomes that a pure substitution strategy would not produce and might reverse.
Dell'Acqua and colleagues (2025) found the same amplifying shape among 758 BCG consultants: AI compressed the performance distribution, lifting the lowest performers most. But the same experiment carries the warning for anyone leaning toward substitution. On tasks outside the model's capability, AI users produced worse results, deferring to confident wrong answers. Remove the human judgment and you do not just lose the augmentation upside, you inherit the failure mode. The evidence rewards keeping a capable person in the loop and punishes taking them out.
Why capital keeps aiming wrong
If augmentation is both what workers prefer and where the returns are cleaner, why does investment drift toward substitution? Partly because substitution has a legible business case, a headcount line that shrinks, while augmentation shows up as diffuse gains in quality, speed, and retention that are harder to put in a slide. Microsoft's 2025 survey captures the mood driving decisions: 82 percent of leaders plan to use agents to extend workforce capacity within 12 to 18 months, and 46 percent already automate whole workflows. "Extend capacity" can mean amplify people or replace them, and the ambiguity lets a substitution instinct ride in under an augmentation banner.
The World Economic Forum (2025) recommends a corrective: deploy agents with governance proportional to task risk, matching autonomy to the nature of the work rather than automating uniformly because it is possible. That is, in effect, a rule for putting the augmentation-versus-substitution decision where it belongs, at the level of the individual task, judged on its merits, rather than as a blanket direction of travel.
The practical read
The firms that get the most from agents will not be the ones that automated the most jobs. They will be the ones that automated the right tasks and amplified their people on the rest, because that is where the workers want to go and where the field evidence says the value is. Practically, that means two lists. First, the low-value tasks people are glad to lose, which Shao and colleagues (2025) show are under-served today, so start there and win goodwill. Second, the high-value tasks where a human keeps the final call and the agent makes them faster and better. Point capital at both, and be suspicious of any agent project whose entire business case is a smaller headcount. The evidence says that bet is aimed at the wrong tasks.
Sources
- Shao, Zope, Jiang, Pei, Brynjolfsson, Yang et al. (2025). Future of Work with AI Agents. Maps worker desire against AI capability across 844 tasks, showing preference for augmentation and a misalignment of investment. https://arxiv.org/abs/2506.06576
- Brynjolfsson, Li and Raymond (2025). Generative AI at Work. Finds AI amplified a workforce by diffusing top performers' practices, with gains concentrated among novices plus better sentiment and retention. https://academic.oup.com/qje/article/140/2/889/7990658
- Dell'Acqua et al. (2025). Navigating the Jagged Technological Frontier. Shows AI amplifies lower performers but degrades results when it replaces human judgment on out-of-range tasks. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Microsoft WorkLab (2025). Work Trend Index 2025. Reports 82 percent of leaders plan to use agents to extend capacity and 46 percent already automate whole workflows. https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born
- World Economic Forum (2025). AI Agents in Action. Recommends risk-proportional deployment, a rule for deciding automation task by task. https://www.weforum.org/publications/ai-agents-in-action-foundations-for-evaluation-and-governance/