CI/CD for AI applications is not just "run tests, build a container, deploy." The deployable behavior of an AI system may change when code changes, prompts change, retrieval data changes, model configuration changes, safety policy changes, or a hosted model provider updates a model behind an alias.

Good pipelines make those changes visible before they reach users. They also give the team a fast way to stop, roll back, or limit exposure when the evaluation signal is uncertain.

This chapter applies the same engineering standard you would use for any production service: repeatable builds, clear gates, traceable artifacts, progressive rollout, and rollback paths.