Why do people lose trust in an AI agent right after it sends an email, even when the work itself was fine?

This explores what makes people pull back from an AI agent they had already handed work to. The corpus suggests a counterintuitive answer: how much is at stake is not the main trigger. In a controlled study of students using a general-purpose agent, trust dropped sharply on tasks that were **irreversible and visible to other people**, such as sending an email. This happened even when users rated the agent's output as adequate. High-stakes tasks that could still be corrected produced no such drop What makes people distrust AI agents they delegate to?. So people aren't mainly reacting to bad work. They're reacting to losing the chance to catch a mistake before the world sees it.

That reframes trust as being about control rather than competence. One argument in the collection is that people need the *felt* ability to step in, not constant hands-on oversight. An agent feels safe when the option to intervene is visible, even if the user rarely uses it Does perceived control matter more than actual control for AI trust?. Anthropic's usage data on Claude Code points the same way from a different angle. Experienced users auto-approve far more sessions than newcomers, yet they interrupt about twice as often Do experienced AI users trust agents more or just oversee differently?. Their trust doesn't simply rise over time. They switch from approving each step in advance to watching the work and stepping in when something looks wrong. Withdrawal, in other words, may be less a switch flipping off and more a shift in *where* oversight happens.

The weak point tends to be supervision itself. An analysis of more than 73,000 Reddit posts about one agent found that users' values were mostly met when they described what the agent delivered. In every user group, though, those values broke down during supervision: setting the agent up, watching it run on its own, and reviewing the results Where do user values break down in agent supervision?. At larger scale, Capgemini found enterprise trust in fully autonomous agents fell from 43% to 27% in a single year as deployments moved past pilots. Organizations still planned to use agents, but on teams supervised by humans Why is enterprise trust in autonomous AI agents declining?. Early enthusiasm meets real use, and people respond by asking for more oversight rather than abandoning the tools.

There's also a quieter side to this. Trust doesn't only collapse; sometimes it's misplaced from the start. People often accept fluent AI output without checking it because checking is costly. One note calls this 'cognitive surrender' When do users stop checking whether AI output is actually backed?. Designed personality and human-like warmth can push trust further than the agent's actual competence justifies How far can personality push users toward misplaced AI trust?. One reading of sudden withdrawal, then, is a correction: the gap between assumed and real reliability finally becomes visible, often at an irreversible moment.

The design lesson that falls out of this is to expand an agent's autonomy in stages gated on trust. Under that approach, an agent moves from analysis-only, to low-risk actions, to full autonomy only as users show they understand how it thinks, not on a fixed schedule Should agentic AI autonomy expand based on earned trust?. Taken together, the corpus suggests trust breaks down less when an agent errs and more when users realize they can't see or undo what it's about to do.