This paper investigates the optimal acquisition of information in a model of job assignment within a firm. We consider a firm with two types of jobs, skilled and unskilled. The firm draws workers randomly from the general population, and a worker is either talented or untalented. Initially, a worker's productivity in the firm is unknown to the worker and the firm. Workers are equally productive in the unskilled job, but talented workers are more productive in the skilled job than in the unskilled job, and untalented workers are more productive in the unskilled job than in the skilled job. Before assigning a worker to a job, the firm can test whether the employee is talented, and the firm is able to choose the accuracy of this test. We assume that the cost of a test is increasing and convex in test accuracy. We show that (1) the accuracy of the firm's test increases with the cost of a mismatched worker; (2) increased optimism about the worker's ability need not lead to less rigorous testing; (3) the probability that a worker is assigned to the skilled job need not increase as the gain from assigning a talented worker to a skilled job increases, or the loss from assigning an untalented worker to a skilled job decreases, or the fraction of the population that is skilled increases; and (4) a longer testing period, allowing as many as two tests of workers, leads the firm to use a less expensive, and less accurate, test initially than when there is only one opportunity to gather information.