A study of 289 workers found that humans draw a sharp line between tasks they will hand to an AI agent and tasks they insist on keeping. Researchers used a statistical method called generalized estimating equations to measure how much people wanted to stay involved across dozens of everyday and work tasks. The findings show that desire for involvement is not uniform. It shifts by task type, stakes, and personal domain.
People were most willing to hand over information-gathering, scheduling, and routine data work. They pulled back sharply on tasks involving judgment, relationships, or personal values. The study found that financial and health decisions, creative work with personal meaning, and anything touching other people's opinions sat firmly in the keep-it column. This is not a story about trust in AI in general. It is a story about where trust ends and the human feels the work is theirs.
For a team building an agent product, the practical read is this: the tasks humans most want to delegate are also the tasks where mistakes matter least. The tasks humans most want to keep are exactly where an agent error would be costly or embarrassing. That gap is the design problem. An agent that cannot signal its own limits, or that takes on work the human wanted to own, will lose trust faster than one that does less.