This is what separates an answer you read from an answer a machine can use. Without it, plugging a model into an automation means hoping it will respect the requested shape. With it, the shape is guaranteed.
Definition
Structured output is a mechanism constraining a model to produce its answer in a format defined in advance, with named fields and expected types.
The difference from a plain instruction is fundamental. Writing "answer only in JSON with the fields category and urgency" is a request: the model follows it most of the time. Declaring that format through the provider's dedicated mechanism is a constraint: the answer cannot fall outside it.
"Most of the time" is fine for personal use. Over a thousand runs a month, a 98 percent compliance rate means twenty failures, often silent, leaving your Workflow in an inconsistent state.
What it is actually for
| Need | What the structure guarantees |
|---|---|
| Classifying requests | The category belongs to the allowed list |
| Extracting from an invoice | The amount is a number, the date a date |
| Qualifying a lead | The score stays within the expected range |
| Summarising in points | The number of points is respected |
| Chaining two steps | The next step finds the expected fields |
The last line changes the nature of your automations. While the output is free text, every following step has to guess what it received. With a guaranteed structure, a Node can read a field without wondering whether it exists.
Best practices
Describe every field. The model relies on those descriptions to decide what goes in. "urgency" with no explanation gives arbitrary results; "urgency: high if the customer mentions a deadline, otherwise normal" gives a reproducible classification.
Close your lists. A field accepting only three values beats a text field you will filter afterwards. That is half the point of the mechanism.
Provide an "I do not know" field. Without it, a model forced to fill a field will fill it, even with no information. That is the most direct route to a Hallucination presented in a clean, and therefore credible, form.
Keep the structure short. A format with forty nested fields degrades the quality of each one. Two calls with two simple structures beat one call with a complex structure.
A well-formed answer is not a correct answer. The structure guarantees the amount is a number, not that it is the right amount. Checking the substance remains entirely on you, and an impeccable form tends to lull vigilance.
Frequently asked questions
Is it available everywhere?
At every major provider through their API, under different names. In chat interfaces, no: it is a mechanism for programmatic calls, hence for automation.
Do you need to code to use it?
Not really any more. Automation platforms let you describe the expected fields in a form and handle the rest. Understanding the notion is still useful to know what to ask for.
Does it cost more?
The format itself counts towards the Token sent, so marginally yes. The saving comes from elsewhere: fewer retries, fewer failures to fix by hand.
When should you use it rather than a plain instruction?
As soon as the answer is read by a machine rather than by you. Our n8n course shows that connection at the point a model enters a scenario, with the field descriptions that make all the difference.