Most production AI features are not a single prompt. They retrieve data, classify inputs, call models, parse structured outputs, validate results, write records, notify users, and handle failures. That is a workflow.

An AI workflow organizes model calls and ordinary software steps into a controlled execution graph. The system, not the model, decides the sequence. LLMs appear at specific nodes where they are useful: classification, extraction, summarization, synthesis, ranking, or validation.

This chapter covers deterministic pipelines, DAGs, routing, map-reduce, retries, checkpointing, and a full document-processing workflow. The theme is control: use the model where it helps, and keep orchestration in code when the path is knowable.