Companies say they want control over their AI agents — but does that word mean what they think it means once the agent is actually running?

This explores where human control ranks among the design priorities business users name for workplace AI agents, and what 'control' means once those agents are deployed. The most direct evidence comes from a multi-method study that found eight UX principles for workplace agents. Business users put four of them at the top: human control, reliability, context-awareness, and safety What UX principles do workplace users want in AI agents?. The corpus doesn't order these four against each other. They work as a group of minimum requirements for adoption. Without them, users don't accept the agent and don't collaborate with it. So control isn't traded off against the others. It sits alongside reliability and safety as part of what people need before they will use the tool at all.

The more interesting finding is that what users say they want and what they actually get can be quite different. A study of junior and senior engineers found that company rules (tool mandates, allow-lists, data policies) decide how much control a person keeps over agentic AI before personal preference comes into it Does personal preference shape how engineers use AI tools?. The same pattern shows up at the organizational level. BCG's survey of nearly 12,000 employees found that clear strategic direction predicted value better than access to tools Does strategic direction matter more than AI tool access?. In practice, control is often set by the organization's design choices, not by the individual user.

Holding formal control is also different from feeling in control. Leonardi shows that when people credit an AI with independent judgment, their sense of control and responsibility shrinks, even though they keep full formal authority Does attributing agency to AI actually reduce human sense of control?. The good news in that work is that the loss is reversible: people take control back in later phases of the cycle. Other work goes further and argues that agentic systems can reverse delegation by handing tasks to humans Can AI systems delegate tasks back to humans?. That raises the question of who is controlling whom. It also helps explain why 'context-awareness' sits next to control on users' list. AI context (prompts, history, retrieved data) keeps shifting in ways users can't track the way they would track a normal interface How does AI context differ from conventional software context?, so steering the agent depends on understanding what it is working from.

Control also doesn't mean users want to do everything themselves. A survey of 1,500 workers across 844 tasks found that equal human-AI partnership is the most-wanted level in 45% of occupations, and that 41% of startup investment targets areas that don't match what workers want What collaboration level do workers actually want with AI?. Read this way, 'control' is mostly about having a clear say in how work is divided, not about holding a veto. That fits with framing human-AI decision-making as an organization design problem How should humans and AI systems divide decision-making tasks? and with Human Factors proposals for clear task allocation and continuous feedback, though those proposals haven't been tested yet Can Human Factors design fixes restore GenAI productivity losses?.

One note on what the corpus can't tell you: none of these sources ranks control above or below reliability or safety. The sharper takeaway is that the control users ask for is often already decided by company policy, and it can slip away psychologically even when nobody takes it away formally.