We developed statistical equations to predict abnormalities on eight commonly ordered diagnostic tests and we gave the predictions to 112 physicians practicing in an academic internal medicine practice. Half of each physician's patients were randomized to intervention status. All diagnostic tests were ordered by microcomputer, and when a physician ordered one of the eight study tests for an intervention patient, the computer displayed the probability (0% to 100%) that the test would be positive for the main abnormality being tested for. The physician could then cancel the test if desired. During a six-month controlled trial, when there were more than 15,000 scheduled patient visits, patient charges for the eight study tests were 8.8% less for the intervention patients. The largest reductions (greater than 10%) were for serum electrolyte level tests and complete blood cell counts, the two most commonly ordered tests. Physicians ordered fewer low-probability tests for intervention patients than for controls, suggesting that with timely predictive information, physicians can target tests to higher-risk patients.