Every vendor pitch for enterprise AI shows the same curve. Work gets faster, quality climbs, one person now does the job of three. The best field evidence agrees with that curve, and adds a warning most decks leave out. On the tasks AI is good at, it is very good. On the tasks sitting right next to them, the ones that look identical but are not, it can make your strongest people measurably worse.
Two findings from one experiment
The cleanest test comes from inside a working firm. Dell'Acqua and colleagues (2025) ran a pre-registered field experiment with 758 Boston Consulting Group consultants, roughly seven percent of the firm's individual consultants, randomly assigned to work with no AI, with GPT-4, or with GPT-4 plus prompt training. On a battery of tasks that sat inside the model's capability, the consultants using GPT-4 completed 12.2 percent more tasks, worked 25.1 percent faster, and produced output rated about 40 percent higher in quality than the group working without it. Those are not marginal gains. That is the number that sells the software.
Then the same paper turned the result over. On a task deliberately built to fall outside the model's competence, one that required reconciling misleading data with careful judgment, the consultants using AI were more likely to reach the wrong answer than those working without it. They were not lazy. They were persuaded. The model returned answers that were fluent, confident, and incorrect, and experienced professionals deferred to them. The authors called the underlying shape the "jagged frontier": capability that is high in one place and absent a short distance away, with no visible marker between the two. The danger is precisely that the boundary is invisible from the user's chair.
Why fluency is the trap
The customer-service evidence points the same way from a different angle. Brynjolfsson, Li and Raymond (2025), studying 5,172 support agents given a generative-AI assistant, found average productivity up about 15 percent, with the gains heavily concentrated among newer and lower-skilled staff. The mechanism they identify is that the tool spreads the tacit playbook of the best agents to everyone else. That is a real benefit, and it carries a hidden cost. A junior worker who has absorbed the model's suggestions has less practice at telling a good suggestion from a plausible one. Competence at the task and competence at auditing the tool are not the same skill, and AI can grow the first while starving the second.
This is why "human in the loop" is a slogan, not a control, until you specify which loop. A person nominally reviewing agent output who has no independent way to judge it is a rubber stamp with a pulse. Chan and colleagues (2024) argue that as more work is delegated to agents that pursue goals with limited supervision, organizations need actual visibility into what those agents did and why, through activity logs and monitoring, not just a checkbox that a human "approved." Approval without the means to disagree is not oversight.
Managing the boundary
The practical response is not to slow adoption. The productivity numbers are too large to ignore, and Microsoft's 2025 survey of 31,000 workers reports that 82 percent of leaders expect to use AI agents to extend workforce capacity within the next 12 to 18 months. The response is to treat the jagged frontier as a management object, something you map, mark, and revisit, rather than a property of the model you hope your staff will intuit.
Three moves follow directly from the evidence. First, sort the actual work into tasks where the model is reliable and tasks where it is confidently unreliable, and route them differently. Dell'Acqua and colleagues (2025) found that even a short prompting intervention did not close the gap on outside-frontier tasks, so training people to prompt better is not a substitute for keeping certain judgments human. Second, preserve independent verification for the high-stakes zone. Shao and colleagues (2025), auditing what U.S. workers actually want from AI, find that people are happy to hand off low-value tasks but insist on "shared agency" for high-value ones, where the machine executes and the human keeps judgment and responsibility. That preference is not nostalgia. It is a correct read of where deference gets expensive.
Third, measure the error mode you are most exposed to, not just the speed you gained. The seductive quality of the jagged frontier is that the failures look like successes. A wrong answer arrives on time, well formatted, and internally coherent. The firms that get hurt will be the ones that celebrated the 40 percent quality lift on the easy tasks and never noticed the quiet degradation on the hard ones. The frontier is jagged whether or not you draw it. The only choice is whether your organization knows where the edge is before it steps off.
Sources
- Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon and Lakhani (2025). Navigating the Jagged Technological Frontier. Field RCT with 758 BCG consultants showing large gains inside AI's capability and worse performance outside it, the core evidence for the jagged-frontier argument. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Brynjolfsson, Li and Raymond (2025). Generative AI at Work. Field study of 5,172 support agents where AI spread top performers' practices to novices, supporting the point that fluency can crowd out independent judgment. https://academic.oup.com/qje/article/140/2/889/7990658
- Chan et al. (2024). Visibility into AI Agents. Argues for activity logs and monitoring so oversight of delegated agents is real rather than a rubber stamp. https://doi.org/10.1145/3630106.3658948
- Shao, Zope, Jiang, Pei, Brynjolfsson, Yang et al. (2025). Future of Work with AI Agents. Finds workers accept automation of low-value tasks but want shared agency on high-value ones, supporting the case for keeping judgment human at the frontier. https://arxiv.org/abs/2506.06576
- Microsoft WorkLab (2025). Work Trend Index 2025. Survey of 31,000 workers reporting 82 percent of leaders plan to use AI agents to extend capacity within 12 to 18 months. https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born