The outcome bias induces economic agents to be rewarded or punished based on random outcomes irrelevant to assessing performance. In a theory-guided experiment, I study how the outcome bias shapes agents’ choices, which is crucial to understanding welfare implications. I propose a simple model of outcome bias. Its key prediction is that the outcome bias can trick principals into incentivizing agents to choose their least favorite action when it is most likely to yield a higher payoff than the alternatives. In the experiment, agents, acting on behalf of principals, choose between two lotteries. One lottery is clearly optimal (FOSD), but it yields a lower payoff in most states. Principals perfectly observe the agent’s choice and the realized state of the world and decide whether to award a bonus to the agent. Despite having a revealed preference for the dominant lottery, principals are more likely to award the bonus for the dominant lottery. Stated beliefs show that agents anticipate these perverse incentives. In what appears to be a failure of strategic reasoning, most agents choose the dominant lottery, contrary to their beliefs. Strategically sophisticated agents are more likely to choose a dominated lottery when they believe to have an incentive to do so. The results suggest that the outcome bias might be most relevant in settings with sophisticated agents such as politicians, CEOs, or fund managers. Structural estimation suggests that principals fall almost exclusively into one of two types: (near) fully outcome-biased or (near) fully unbiased.