Genetic mutations are being thoroughly mapped in human cancers, yet a fundamental question in cancer biology is whether such mutations are functionally required for cancer initiation, maintenance of established cancer, or both. Here, we study this question in the context of human acute myeloid leukemia (AML), where DNMT3A R882 missense mutations often arise early, in pre-leukemic clonal hematopoiesis, and corrupt the DNA methylation landscape to initiate leukemia. We developed CRISPR-based methods to directly correct DNMT3A R882 mutations in leukemic cells obtained from patients. Surprisingly, DNMT3A R882 mutations were largely dispensable for disease maintenance. Replacing DNMT3A R882 mutants with wild-type DNMT3A did not impair the ability of AML cells to engraft in vivo, and minimally altered DNA methylation. Taken together, DNMT3A R882 mutations are initially necessary for AML initiation, but are largely dispensable for disease maintenance. The notion that initiating oncogenes differ from those that maintain cancer has important implications for cancer evolution and therapy.