This study compares the documentation of ambulatory care visits and diagnoses in Medicaid paid claims and in medical records. Data were obtained from Maryland Medicaid's 1988 paid claims files for 2407 individuals who were continuously enrolled for the fiscal year, had at least one billed visit for one of six indicator conditions, and had received the majority of their care from one provider. The patients sampled were also stratified on the basis of the case-mix adjusted cost of their usual source of care. The medical records for these individuals as maintained by their usual source of care were abstracted by trained nurse reviewers to compare claims and record information. Linked claim and medical record data for sampled patients were used to calculate: (i) the percent of billed visits documented in the record, (ii) the percent of medical record visits where both the date and the diagnosis agreed with the claims data, and (iii) the ratio of medical record visits to visits from billed claims. Included in the analysis were independent variables specifying place of residence, type and costliness of usual care source, level of patient utilization, and indicator condition on which patient was sampled. Ninety percent of the visits chronicled in the paid claims were documented in the medical record with 82% agreeing on both date and diagnosis. Compared to the medical records kept by private physicians and community health centers, a significantly lower percent of hospital medical records agreed with the claims data. Total volume of visits was 2.6% higher in the medical records than in the claims. Claims data substantially understated visits in the medical record by 25% for low cost providers and by 41% for patients with low use rates (based on claims information). Conversely, medical records substantially understated billed visits by 19% for rural patients and by 10% for persons with high visit rates. Although Medicaid claims are relatively accurate and useful for examining average ambulatory use patterns, they are subject to significant biases when comparing subgroups of providers classified by case-mix adjusted cost and patients classified by utilization rates. Medicaid programs are using claims data for profiling and performance assessment need to understand the limitations of administrative data.