A structured output is a model response that follows a defined shape, such as a JSON object with required fields and allowed values. Instead of returning a paragraph that software must interpret, the AI might return {"priority":"high","owner":"legal"} according to a schema the application already understands.
This is useful whenever an AI response feeds another system. A support workflow can require a category, confidence level, and escalation reason. A research pipeline can require a claim, supporting excerpt, and source URL. An agent can choose its next step from a small set of permitted actions rather than inventing a new label each time.
Structured output is related to tool calling: tool arguments are usually structured so code can validate and execute them. But a structured response does not have to call a tool; it may simply produce clean data for a form, database, report, or later agent.
Why it matters at your desk. A marketer can turn a folder of campaign briefs into consistent records. A researcher can extract the same fields from every paper. An engineer using Cursor can ask for output that a test or script checks automatically, while teams building Workspace Agents in ChatGPT can use defined fields to hand work safely from one step to the next.
Schema compliance is not truth. A model can place a fabricated fact inside perfectly valid JSON, omit nuance because the schema has nowhere to put it, or choose the wrong allowed category. Validate the content as well as the shape, preserve source evidence, and add an explicit unknown or review path when the model cannot support an answer.