Document intelligence validation
Introduce automatic acceptance one document class at a time
Begin with a class whose fields and validation rules are understood. Expand using reviewed evidence rather than assuming every supplier format behaves the same way.
In this article
Define the first eligible class
Choose documents with a clear business purpose, supported fields and enough reviewed examples to assess correctness. Define eligibility using observable characteristics, not a vague label such as "simple invoices".
Keep unfamiliar documents on the review path. A classifier's confidence that a file belongs to the class should also be evaluated, because a misclassified document may receive inappropriate validation rules.
Record the acceptance rule version, extraction configuration and source fixture set. These artifacts establish what was actually approved for automation.
Compare without posting automatically
Run the proposed acceptance rule alongside the existing review process. Compare its decisions with reviewed outcomes and inspect false acceptance, not only the number of documents it would have passed.
Use this phase to identify missing fields, ambiguous adjustments and layouts that need separate treatment. Do not silently edit labels to agree with the candidate. Resolve disagreements against the source and business requirements.
Keep the mode clearly separated from execution. An assessment flag should not accidentally become permission to post a record.
Enable a limited production path
Turn on automatic acceptance for the agreed class and retain sampled review of accepted records. Monitor disqualifying errors, exception volume and downstream rejection.
Provide a control that stops new automatic acceptance while preserving extraction and manual review. This allows containment without disabling the entire document workflow.
Plan for records already accepted or submitted when the control changes. Some may need reconciliation or targeted review, and the application should identify them by rule version and operation record.
Expand only with new evidence
Add another class when its reviewed results justify the decision. Reuse common infrastructure, but do not assume the same confidence threshold or arithmetic relationships apply unchanged.
When changing the extraction model, rerun the accepted classes as well as the new one. A model update can improve one layout while regressing another. Keep the previous acceptance configuration available during the observation period, with a tested manual fallback if restoration is incompatible.
The rollout is complete for a class when the team understands its observed error rate, review burden and recovery path. Broader coverage should follow that evidence, not a target percentage detached from document quality.
Primary sources
Microsoft Learn: extraction confidenceMicrosoft Learn: invoice extractionReferences checked 11 September 2026.