AI workflow

In an AI workflow, the path is decided in advance by your code. A typical one might extract fields from an email, look up the matching record, apply business rules, and draft a response, with the model handling only the steps that need language understanding. Prompt chains, routing by classification, and fan-out-then-merge pipelines are all workflows.

Workflows are the default shape for most AI product features because they are easy to reason about. Each step can be tested on its own, costs are predictable, failures are traceable to a specific stage, and you can put deterministic checks and human approvals exactly where they belong. Automation platforms like Zapier, Make, and n8n build the same pattern visually.

The trade-off is flexibility. A workflow handles the cases you designed for; when the right next step depends heavily on what an earlier step found, the decision tree can grow unwieldy. A common middle ground is a workflow with one bounded model-driven stage that may call a few tools.

Workflows that run longer than a request can wait belong in a background job, with each step's result persisted so a crash or deploy doesn't restart the whole thing.