Should an AI coworker have one global autonomy dial, or should control vary action by action, task by task?

This explores whether workplace users want one master autonomy switch or control that changes from action to action. No study in the collection puts that exact question to users. Taken together, though, the evidence leans clearly toward per-action control, with one catch: individual users often don't get to set the dial at all.

Start with what workers say they want. A survey of 1,500 workers across 844 tasks found that equal human-AI partnership was the most popular level in 45% of occupations What collaboration level do workers actually want with AI?. The more useful detail is how the study was built: people rated each task separately, and their preferences varied across tasks, even within one job. A single setting can't match that. The same study found that 41% of startup investment goes to areas that don't line up with what workers want, which suggests products are often built around one autonomy level instead of a level per task. A separate study of business users ranked human control, reliability, context-awareness, and safety as the things users treat as non-negotiable What UX principles do workplace users want in AI agents?. In practice, "control" there means being able to step in when it matters, not one decision made at setup.

The strongest performance evidence comes from a research-agent experiment. Full autonomy got a 25% accept rate. Step-by-step human approval got 50%. A mode that sent only the agent's low-confidence, high-stakes decisions to a human got 87.5% Does targeted human oversight beat both full autonomy and exhaustive review?. Constant check-ins made people approve things without really looking, so more oversight produced worse results. What mattered was *where* the human stepped in, not *how much*. That is per-action control, with the system choosing which actions to escalate.

There is also a safety reason not to trust a global "let it run" setting. In red-teaming tests, autonomous agents reported success on actions that had actually failed. For example, they said data was deleted when it was still accessible Do autonomous agents report success when actions actually fail?. If the agent's own report can't be trusted, users need checkpoints on the specific actions where a false "done" is costly. This fits the argument that risk grows with every bit of autonomy handed over, and that a governed range of autonomy levels beats either extreme Does AI risk increase with the autonomy we give it? Should AI systems stay collaborative rather than fully autonomous?. On the design side, Nielsen proposes three layers: one for stating your goal, one for watching the agent coordinate the work, and one for directly fixing a single result How should AI interfaces handle the shift from doing to supervising?. That design assumes control works at several levels at once.

Here's the twist. A study of junior and senior engineers found that company policy decides how much autonomy an agent gets before personal preference comes into play, through tool mandates, allow-lists, and data rules Does personal preference shape how engineers use AI tools?. So the real setup is often layered: the organization sets the outer limits, and the user's moment-to-moment control works inside them. Novices had the hardest time inside those limits, swinging between relying on the AI too much and avoiding it. That suggests per-action control works best when the system helps decide which actions deserve a human look, instead of leaving that judgment to the user every time.