Most automation plans start from a question about the machine: what can it do? A more useful question, and one almost nobody asks first, is about the person: how much of this task do they actually want to hand over? The two questions have different answers, and the gap between them is where a lot of AI budget is about to be wasted.
A scale for a word we use too loosely
"Automation" hides a spectrum. Letting an agent draft a routine email is not the same as letting it decide which customers to drop, yet both get filed under the same verb. Shao and colleagues (2025) at Stanford give the spectrum a vocabulary. Auditing what U.S. workers want from AI across 844 tasks in 104 occupations, drawn from 1,500 professionals, they introduce a Human Agency Scale running from H1 to H5, describing how much human involvement a worker prefers on a given task, from near-total automation at one end to the human staying firmly in charge at the other.
The point of the scale is that the right level is not a company-wide setting. It varies task by task, and workers hold consistent views about which is which. The same study finds a clear pattern: people want AI to take over low-value, repetitive work, and they want "shared agency" on high-value work, where the machine executes but the human keeps judgment and responsibility (Shao and colleagues, 2025). This is not resistance to technology. It is a considered distinction between the parts of a job that are a burden and the parts that are the job.
The misalignment nobody priced in
Here is where it turns into a management problem rather than a survey result. When Shao and colleagues (2025) cross-reference what workers desire against where AI capability is actually heading, the two do not line up. A meaningful share of agent investment is flowing toward tasks that workers do not want automated, while some of the tasks people would gladly offload remain untouched. That misalignment is a leading indicator of adoption failure. Tools that automate the wrong thing get quietly abandoned, or worse, resented, no matter how capable they are.
This should sober anyone reading the adoption headlines. Microsoft's 2025 survey of 31,000 workers reports that 82 percent of leaders plan to use AI agents to extend workforce capacity within 12 to 18 months, and that 46 percent already use agents to fully automate some workflows. Capacity and speed are the metrics driving those decisions. Desire is not on the dashboard. A firm can be simultaneously on schedule against its automation roadmap and pointed at the tasks its people most want to keep.
Designing for agency, not just capability
The fix is not complicated, though it inverts the usual order of operations. Before deciding what to automate, ask the people who do the work where on the H1-to-H5 scale each task belongs for them, and treat serious divergence from the technical plan as information rather than obstruction. The World Economic Forum (2025) makes a compatible recommendation from the governance side: agents should be introduced with a defined role, permissions, and supervision, proportional to the risk of the task, rather than deployed uniformly because they can be. Both frames arrive at the same design principle. The unit of decision is the task, not the department.
There is a self-management logic here that predates the technology. Organizations that pushed decisions down to the people closest to the work did not do it out of sentiment; they did it because those people held information the center could not see. Task-level preference for AI involvement is exactly that kind of local knowledge. A support agent knows which parts of a call are drudgery and which require reading a frightened customer. A senior analyst knows which judgments they would trust to a model and which they would stake their name on. Ignoring that knowledge to hit a uniform automation target is the same mistake, in new clothes.
The practical read
Shared agency is not a compromise between full automation and none. On high-value work, Shao and colleagues (2025) find it is what workers actively prefer, machine speed under human accountability. The firms that will get the most from agents are not the ones that automate the most tasks. They are the ones that automate the right ones, which means measuring desire alongside capability and building the deployment around both. Ask your people how much control they want to keep. Then believe the answer, because the alternative is buying tools that sit idle while the work they were meant to do stays exactly as heavy as before.
Sources
- Shao, Zope, Jiang, Pei, Brynjolfsson, Yang et al. (2025). Future of Work with AI Agents. Introduces the Human Agency Scale (H1-H5) and documents misalignment between where workers want AI and where capability is heading. https://arxiv.org/abs/2506.06576
- Microsoft WorkLab (2025). Work Trend Index 2025. Survey of 31,000 workers showing 82 percent of leaders plan agent adoption and 46 percent already fully automate some workflows, driven by capacity rather than worker preference. 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 deploying agents with role, permissions, and supervision proportional to task risk rather than uniformly. https://www.weforum.org/publications/ai-agents-in-action-foundations-for-evaluation-and-governance/