I’m building a feature that asks an OpenAI model to return data matching a JSON schema. Structured output sounds like it should remove most of the parsing headaches, but I’m unsure how much validation still belongs in my application.

For example, would you still check required fields, allowed values, and business rules after the response matches the schema? I’d be interested in hearing how people handle OpenAI structured outputs in production, especially when a valid response can still contain the wrong information.