Structured output (JSON mode)

You need it the moment an AI answer feeds an automation rather than a person. A human can read a paragraph and pick out the supplier name. A workflow step expecting a field called "supplier" breaks when it gets prose.

How it works

You give the model a schema: the field names, their types and, for some fields, a fixed list of allowed values (an enum). The model fills in those fields. Many APIs have a setting that enforces the format. OpenAI's documentation separates two levels: JSON mode, which guarantees valid JSON but does not check it against your schema, and Structured Outputs, which makes the response match the JSON Schema you supply.

A correctly shaped answer can still hold wrong values. Validate the output in your code before it changes anything, such as updating a CRM record or sending an email.

Example

An operations manager extracts details from supplier invoices into a sheet. Her schema has five fields: supplier, invoice_date, total, currency and category, where category is an enum of "software", "contractors", "ads" or "unknown". When an invoice is ambiguous, the model returns "unknown" and the workflow sends that row to her for a quick check instead of filing it under the wrong budget line.

Common mistake

Building a schema with no way to say "I don't know". If every field must hold a real value, the model guesses, and the guess looks as confident as a correct answer (a form of hallucination). Add an "unknown" option to enums and keep schemas small. A long schema asking for data the source document does not contain mostly produces invented values.