We document a lower bound for thecontrol premium: agents' willingness to pay to control their own payoff. Participants choose between an asset that will pay only if they later answer a particular quiz question correctly and one that pays only if their partner answers a different question correctly.  However, they first estimate the likelihood that each asset will pay off.  Participants are 20% more likely to choose to control their payoff than a group of payoff-maximizers with accurate beliefs.  While some of this deviation is explained by overconfidence, 34% of it can only be explained by the control premium.  The average participant expresses a control premium equivalent to 8% to 15% of the expected asset-earnings.  Our results show that even agents with accurate beliefs may incur costs to avoid delegating and suggest that to correctly infer beliefs from choices, one should account for the control premium.