Structured output makes AI responses easier for software to read, but a tidy object is not automatically a correct business instruction. When commissioning an extraction or assistant workflow, ask where shape validation ends and business validation begins. The boundary matters whenever generated fields could create, change or route an operational record.
Related service: AI automation development- 01Validate structure
- 02Check business meaning
- 03Confirm authority
- 04Apply or review
SECTION 01
Ask for a layered contract
Start with the output fields, allowed values and treatment of missing information. JSON Schema supports object properties and required fields; additional fields are a separate decision rather than something to assume away. Ask the agency to show accepted and rejected examples, including unexpected nested data and values outside the agreed format.
Next define semantic checks: does the customer exist, is the date plausible for this task, and does the referenced document belong to the same account? Some checks need a trusted lookup rather than a model opinion. Authorization belongs in the receiving system. OWASP's excessive-agency guidance is relevant when generated output can invoke real tools.
SECTION 02
Illustrative example: a draft purchase request
Consider an email asking to reorder twelve boxes of packaging. Extraction returns a product code, quantity and delivery address. The shape is valid, but the selected product is sold individually, not by box. The workflow should flag the unit mismatch instead of confidently creating twelve individual items.
The reviewer sees the original sentence and the matched catalog record, chooses the correct pack unit, and approves a specific draft. A later change to quantity requires another check. This is a hypothetical implementation brief, not a claim that any model can infer every purchasing rule. Direct catalog matching may be preferable to AI where the input is already structured.
SECTION 03
Buy useful rejection and repair behavior
Agree whether invalid output is retried, returned for correction or routed to a person. Put a limit on repair attempts and retain the reason for rejection. Automatically asking the same model to fix its own object is not independent verification, especially when the source itself is incomplete.
Acceptance should include missing required fields, wrong types, extra commands, unknown record IDs, unauthorized accounts and valid-looking but contradictory values. Request versioned schemas, business-rule ownership and examples that reproduce failures. The handoff should explain how an added field affects older integrations and how a rejected proposal can be corrected without silently overwriting evidence.
SECTION 04
Frequently asked questions
Does valid JSON mean the answer is correct?
No. It means the response is syntactically readable. Schema rules add structural checks, while factual and business correctness require further validation.
Should every output need manual approval?
Not necessarily. Agree which bounded, reversible cases may proceed after checks and which uncertain or consequential cases require review.
Can schema validation replace permission checks?
No. The destination must independently verify the user's authority and the permitted action, regardless of how well formatted the generated request is.
PRIMARY REFERENCES
Sources and further reading
These references cover the standards, platforms, or published prices discussed in the guide. Worked examples and checklists are our editorial guidance.
- JSON Schema object validationJSON Schema
- Excessive agencyOWASP
EDITORIAL METHOD
About this guide
We use AI to assist with drafting and editing. Catapult AI Work is responsible for the published content. Examples illustrate possible approaches; they are not client case studies unless identified as such.
Budget examples are not Catapult package prices. Check linked provider pages for current fees and plan limits before making a purchase.
Read the editorial policy