A process-model-based methodology for code review supported by artificial intelligence

Abstract

Code review remains a central quality assurance practice in software projects, and a new participant has recently entered it: an assistant based on a large language model that generates review comments, summaries, and draft fixes. Existing descriptions treat such an assistant as an external tool and leave open where it acts and with what authority. This paper develops a methodology for code review supported by artificial intelligence, built on an explicit process model in which the assistant is a first-class participant without gating authority. The methodology comprises six principles, six integration points bound to process-model elements, routing rules driven by predicted review characteristics, a feedback accumulation loop on assistant output, and an adoption procedure. The empirical grounding uses twelve open-source repositories with about sixty-five thousand merged pull requests. Twelve such assistants take part in review; roughly one merged pull request in six carries at least one of their reviews; among all collected pull requests, the share of those with such a review grew from a fraction of a percent to over a third in two years and differs across repositories by three orders of magnitude. A descriptive within-repository comparison shows that pull requests reviewed by an assistant are systematically larger and go through more iterations, which persists within quartiles of change size. Measurability and transparent configuration are grounded empirically; prediction-driven control rests on earlier results; the preconditions of the remaining principles are verified descriptively: assistants rarely issue verdicts, act on the model trigger, and the platform records human reactions.