The extracted table adds up. Are its fields still correct?

Check identifiers, units and each source row before trusting the total. A correct sum can coexist with incorrect extracted values.

Keep the source document beside the extracted result. Ask for the page and source text for each important field, and mark unreadable fields for review rather than asking AI to fill them in.

A total cannot check every field

Our fictional document has ID 0071, currency KRW, row amounts 3,600 and 1,500, and total 5,100. Changing the rows to 3,500 and 1,600 still gives 5,100. Changing the ID to 71 or currency to USD also leaves that sum unchanged. These are constructed errors, not observed OCR outputs.

Check the source at field level

Compare the identifier as text, the currency, row labels and amounts, then the total. For a quantity-by-price table, check multiplication as well. In this example, 3 × 1,200 = 3,600 and 2 × 750 = 1,500.

Use confidence as one review signal

Microsoft’s Document Intelligence guide describes confidence at different levels and notes that not all fields return a score. A correct cell can sit in a row with other errors. Its guide includes human review; do not treat a single high score as verification of the whole document.

Keep uncertain entries visible

Our suggested extraction request includes value, unit, source page and a “needs review” status. Do not invent an unreadable digit to make a table look complete. Use material you are allowed to process and remove unnecessary private details before an external upload.

Published October 14, 2026. We checked the invented arithmetic and counterexamples. We did not run OCR, Azure Document Intelligence or a chatbot on a document.

Sources: Microsoft Learn · Extraction confidence.