Why do people keep wanting a hand on the wheel, even when an AI agent's work is good enough?

This explores why people at work keep wanting a hand on the wheel when AI agents act for them. The corpus has no single survey that ranks control against other preferences. What it does have is a set of studies that point to the same answer: people aren't mainly worried about bad output. They're worried about what happens when an agent acts in ways they can't undo or can't see, and about losing their sense of authorship over their own work.

The sharpest clue comes from a study where students handed tasks to a general-purpose agent. Trust dropped steeply on tasks that were irreversible and visible to other people, like sending an email. That happened even when people rated the agent's work as adequate. Tasks with high stakes that could still be corrected caused no such drop (What makes people distrust AI agents they delegate to?). So the demand for control isn't really about how risky a task is. It's about whether a mistake can be taken back before a colleague, client, or boss sees it. Work is full of exactly these tasks: messages sent, records changed, commitments made under your name.

A second answer is psychological rather than practical. Leonardi's 'agency loops' framework finds that once people start crediting an AI with its own judgment, they feel less in control and less responsible, even though they still hold formal authority (Does attributing agency to AI actually reduce human sense of control?). The insistence on control can be read as a 'reassertion' move: people pulling authorship back. Gunasekaran adds a useful twist. People don't actually want constant oversight. They want to feel that they *could* step in (Does perceived control matter more than actual control for AI trust?). That suggests the heavy weighting is partly about something users rarely exercise: knowing an override exists.

Designers are already building toward this. One pattern is the 'autonomy slider', which lets users set an agent to suggest, draft, or execute on a per-task basis (Should agent autonomy be a continuous slider or fixed levels?). Cocoa's system goes further. Users assign each step of a plan to themselves or to the agent and can revise the plan mid-task, and field users especially valued being explicit about who does what (Can users steer AI agents better by revising plans mid-execution?). The direction can even flip: some researchers argue that agentic systems will increasingly hand tasks *back* to humans, which turns 'who controls whom' into a two-way negotiation (Can AI systems delegate tasks back to humans?).

There's also a workplace-specific reason that's easy to miss. Individual employees often have less control than they think, because the company decides first. A study of junior and senior engineers found that tool mandates, allow-lists, and data policies set how much agency people keep over AI before personal preference comes into play (Does personal preference shape how engineers use AI tools?). Above that sit system prompts, a hidden layer of rules set by developers that end users can't see or change (Who actually controls what AI systems say?). Seen this way, the strong preference for control may be a reaction to working inside stacked layers of authority that users don't own. They hold on tightly to the one layer they can touch.