A field is absent from the source: how do you make AI mark it missing instead of guessing?

Preserve the source and identify invented values

This commonly affects invoice organization, product specifications and personnel tables: an unstated date becomes a plausible date in the output. Preserve the source and unedited response and inspect missing fields individually. Judging only table completeness mistakes a model's guesses for retrieved data.

Distinguish absent information from extraction loss

Check the original first; if an image contains the information but extracted text does not, repair parsing or OCR. Next, verify field definitions: order time and payment time are not interchangeable merely because both are dates. Distinguish blanks, zero and not applicable because they can have different business meanings.

Define an explicit representation for unknown values

Require evidence-backed values, a consistent null or missing marker for absent information, and source locations. In structured output, permit the appropriate nullable type in the schema and ensure the application handles unknowns. Do not replace all blanks with a default date in postprocessing; list information needing confirmation separately instead of silently filling official fields.

Use missing-field cases to check for continued guessing

Compare complete inputs with intentionally missing-field examples. Existing values should be extracted accurately while absent fields remain unknown. Include explicit-zero cases so zero is not mistaken for a blank. Retain a field-to-source mapping rather than relying on another model saying the output looks correct.

Missing-value rules must fit the business process rather than disguising guesses as defaults. This page provides extraction and validation methods, handles no real customer records and guarantees no model will stop completing missing values. Where sources conflict, retain the conflict or seek evidence instead of choosing a value and calling it verified.