Microsoft has updated Azure Content Understanding so extracted document fields can point back to their supporting text and carry a learned confidence score. The combination is intended to help applications show evidence and prioritize likely mistakes before an invoice, claim or policy record triggers a business action.
Grounding is now integrated more deeply into extraction, which Microsoft says reduced average inference-token use by as much as 28 percent while improving accuracy. Confidence comes from a separate model rather than the language model’s self-assessment. It considers the document, the location and form of the evidence, the field schema, the predicted value and diagnostic signals from extraction.
A large insurer used the features to label fields for human attention. Microsoft reports more than 97 percent page-level classification accuracy, a nearly 50 percent reduction in review burden and 75 percent faster onboarding of new forms. Because the data was sensitive, people still checked every field; risk labels merely guided their attention.
The confidence value is not a universal automation threshold. Teams must calibrate each field and document type against representative, human-verified production examples, then validate the rule on held-out documents. Empty fields may need separate treatment, and some fields may remain unsuitable for automation. The customer results therefore demonstrate one calibrated workflow, not a guarantee that a score such as 0.8 is safe elsewhere.