Structured Outputs: Getting AI to Fill Forms and Sheets Reliably

What structured outputs are, in plain English

Normally an AI model replies in free text. That's fine for a person, but an automation tool such as Zapier, Make or your own code needs something predictable: a field called invoice_number, a field called total, a date in one format. If the AI sometimes writes "Total: $1,070" and sometimes "the amount due is 1070 SGD", the next step breaks.

A JSON schema is a short description of the answer you want: which fields, what type each one is (text, number, date, yes/no, a fixed list of choices) and which are required. With structured outputs switched on, the model must reply in exactly that shape.

Structured outputs vs JSON mode

Older models offered a looser setting called JSON mode. OpenAI's structured outputs guide compares them: both produce valid JSON, but only structured outputs make the reply follow your schema. JSON mode can still leave out a field or invent a value outside your list.

Anthropic's structured outputs documentation describes the same idea for Claude: JSON outputs constrain the response to your schema, and strict tool use applies the same guarantee to the inputs Claude sends to your functions. It lists current Opus, Sonnet and Haiku models as supported.

Where this helps a small business

Structured outputs are what make AI usable inside ordinary business automations, rather than only in a chat window:

• Supplier invoices to Xero or QuickBooks: read a PDF, return supplier, number, date, line items and total, then create a draft bill. Our guide to AI invoice and receipt extraction covers the full flow.
• Enquiry emails to a CRM: name, company, phone, service wanted and urgency, sent straight into a lead record.
• Classifying messages: tag each support email with one category from a fixed list, so it lands with the right person.
• Spreadsheet rows from documents: delivery orders, quotes or forms turned into one clean row each.

In every case the AI step sits between a trigger and an action, often connected through webhooks or APIs. Our guide to APIs and webhooks explains how those pieces connect.

The limits: right shape, not always right answer

This is the part vendors are careful to state and buyers often miss. OpenAI's guide says outright that structured outputs "can still contain mistakes", and that if the input has nothing to do with the schema, the model may fill the fields with invented values. A correctly shaped invoice with the wrong total is still wrong.

Both vendors also list cases where the reply may not match the schema at all:

• Refusals: if the model declines a request for safety reasons, the refusal takes priority over the schema.
• Token limit reached: if the answer is cut off, the JSON can be incomplete. Anthropic says to retry with a higher max_tokens value.
• Unsupported schema features: both vendors support only a subset of JSON Schema. Anthropic lists numeric limits like minimum and maximum as unsupported, and OpenAI caps schemas at 5,000 properties and 10 levels of nesting.
• Capitalisation of choices: Anthropic notes enum values may come back with different capitalisation, so compare them case-insensitively.

How to validate AI output before it reaches your systems

A dependable automation adds checks around the AI step. These are usually a few lines in the automation tool or in code.

1. 1

   Define a tight schema

   Required fields only where they always exist, fixed lists for categories and currencies, a date format. Add a field like "confidence" or "not_an_invoice" so the model can say the input doesn't fit.

2. 2

   Check the stop reason

   Treat refusals and cut-off responses as failures and send them to a person, not to the next step.

3. 3

   Check the numbers add up

   Line items should sum to the subtotal; subtotal plus GST should equal the total. Reject if not.

4. 4

   Check against what you know

   Is the supplier in your contact list? Is the invoice number already in Xero? Is the date in the past few months?

5. 5

   Route exceptions to a person

   Anything that fails a check goes to a review queue or an email, with the original document attached.

6. 6

   Sample the successes

   Once a month, compare a handful of processed documents with the originals to catch quiet errors.

Costs and practical details

Structured outputs don't have a separate price; you pay normal token prices. Anthropic notes it adds a short system prompt explaining the format, so input tokens rise slightly, and the first request with a new schema is slower while it is compiled, with compiled schemas cached for 24 hours from last use. For per-token prices, see our LLM API pricing comparison.

Anthropic also notes that while prompts and responses are covered by its zero data retention arrangements, the schema itself is cached separately, so don't put personal data in field names or fixed lists. That is good practice with any vendor.

Do it yourself or get help?

Zapier and Make both have AI steps, and many let you define output fields, which gives you some of this without code. The validation checks above are what most do-it-yourself setups skip, and they are what stops a bad extraction from becoming a bad payment.

Our Workflow Automation (S$390 per workflow) builds one workflow of up to five steps on Make, Zapier or Google Apps Script in your account, with error alerts to your email and documentation. For larger jobs, such as an extraction tool with a review screen and checks against your accounting data, our Custom Tool or Dashboard starts from S$1,990 with a free written scope.