When evaluated in retrospect, decision-makers are often judged as if they could have foreseen a random outcome, even when the outcomes carry no relevant information. Although such outcome-biased retrospection is typically viewed as detrimental, it remains unclear how it impacts the incentives of decision-makers when choosing between different actions, the extent to which decision-makers can anticipate the outcome bias, how it influences their choices, and ultimately, welfare. I address these questions through a theory-guided experiment set in a delegated risk-taking context. In this scenario, an agent, acting on behalf of their principal, must choose between a first-order-stochastic-dominant and a dominated lottery. The principal observes the choice and outcome and decides whether to grant a bonus to the agent. I propose a simple model of outcome bias in which bonus decisions depend on an ex-post counterfactual comparison between the obtained outcome and the forgone one. This mechanism can create a new type of agency problem. When the dominant lottery is likely to yield a lower outcome than the alternative, the outcome bias can effectively incentivize the principal to encourage the agent to choose the dominated lottery. I provide experimental evidence confirming this prediction. Stated beliefs suggest that, on average, agents anticipate the outcome-dependent nature of bonus payments, with a majority believing they have strong incentives to choose the dominated lottery. However, despite their stated beliefs, most agents predominantly choose the dominant lottery. Experimental subjects with high scores in cognitive reflection and those who have studied economics exhibit a greater propensity to choose the dominated lottery when they believe they have an incentive to do so. These exploratory results imply that outcome bias may indeed have a detrimental impact on actions and welfare, particularly in settings where decision-makers excel in strategic reasoning. Estimates of finite mixture models suggest that heterogeneity in the principals’ outcome bias can be parsimoniously characterized by two types: fully outcome biased and nearly unbiased, with roughly 62% falling into the fully outcome biased category. High outcome bias correlates with low cognitive reflection and fast response times, supporting its interpretation as a cognitive heuristic. Agents tend to overestimate the degree of outcome bias among principals, with 50% believing that 100% of the principals are fully outcome biased.