People tell surveys they trust AI agents, then quietly keep their hands on the wheel when it's time to actually click buy.

This explores why people say they trust AI agents in surveys but hand over less real control than those numbers suggest. One caveat first: the corpus has no study that puts stated trust and observed delegation side by side for the same people. What it does have, spread across several notes, is a fairly clear picture of why a 'yes, I'd trust it' answer doesn't turn into handing over the keys.

The first reason is that survey trust is abstract and delegation is specific. Accenture found 74% of consumers would trust an AI agent with purchases, but 56% would still pick the brand themselves, and many loyal shoppers wouldn't let the agent switch Do consumers actually want to delegate shopping decisions to AI?. So the 'yes' was really a conditional yes: the agent can do the errand, but I keep the decisions. The more concrete the question gets, the lower the number. Public First found 45% think AI will change the world, but only 29% would give an agent access to their bank account How willing are people to let AI control their finances?. A general attitude toward 'AI' gets measured differently from permission to touch one particular thing that matters.

The second reason is the most surprising. What makes people pull back isn't how high the stakes are. It's whether a mistake can be undone and whether other people will see it. In a controlled study, students lost trust sharply and demanded approval steps for actions like sending an email, even when they rated the agent's output as fine. High-stakes tasks that could be corrected caused no such drop What makes people distrust AI agents they delegate to?. A survey question rarely raises irreversibility or public exposure, but real delegation raises them right away. That may be most of the gap: people answer surveys as if every mistake can be fixed, and they delegate as if it can't.

Third, using an agent tends to wear trust down rather than build it. Capgemini found enterprise trust in fully autonomous agents fell from 43% to 27% in one year, as companies moved past pilots into real deployment Why is enterprise trust in autonomous AI agents declining?. KPMG's 47-country survey shows the same decline, and ties it partly to people's own bad experiences Why does AI trust keep falling even though it matters most?. The way back is feedback on outcomes. People who avoided a partner once it was revealed to be AI changed their minds only after repeatedly seeing its results Does revealing AI identity help or hurt user trust?. Shneiderman's framework explains why capability alone won't close the gap: reliability, safety and trust each come from different sources, and real trust depends on independent oversight, not just better performance Can we separate reliability, safety, and trust into distinct mechanisms?.

Here's the twist you might not expect: the gap also runs the other way. Among developers, 80% use AI tools while only 29% trust their accuracy Why do developers keep using AI tools they don't trust?. They delegate more than they trust, because they can check the output before it ships. Read together, these findings suggest that 'trust' isn't really the variable that decides delegation. Verifiability and reversibility are. People delegate freely when they can catch and undo errors, whatever they tell a survey, and hold back when they can't. A design lesson follows: an agent gets more real delegation by making its actions reviewable and reversible than by sounding more trustworthy. Sounding trustworthy is already a known trap, since trust tends to attach to how authoritative a chatbot sounds rather than to whether it's right Does chatbot language style actually shape how much we trust it?.