Document automation: make exceptions visible
Extraction is the beginning of a decision
Reading fields from a document is different from accepting them as correct. A legible value can still belong to the wrong supplier, use the wrong currency or conflict with another page. Decide which checks and human approvals are required before data reaches a business system.
This is a workflow design guide, not a statement that a particular service is approved for your documents. Confirm permitted data use, access, retention and vendor arrangements before a pilot.
Define the accepted input
List the document types, languages, versions and required fields you intend to handle. Choose representative approved test material, including poor scans, missing pages, duplicates and unfamiliar layouts. Keep originals available to authorized reviewers under an agreed retention policy. A readable file is not necessarily a complete or supported document.
Preserve the link between the original, extracted values and processing version. Treat instructions embedded in a document as untrusted content; they must not change approval rules or authorize tools to act.
Check fields and relationships
Validate required values, formats and references against the destination's rules. Check relationships such as line totals versus a stated total using agreed rounding and currency rules. Flag disagreement with known business records instead of silently replacing them. A check passing does not prove the whole document is authentic or legally valid.
If a tool supplies confidence scores, evaluate them against your own reviewed sample. A high score is not a universal guarantee, and not every field or service supplies a comparable score. Set review thresholds by document type and the consequence of an error; do not copy a vendor example threshold into production without testing.
Give exceptions a reason and an owner
Route incomplete, contradictory, unsupported or uncertain items to a queue with a clear reason. Show only the source material needed by the authorized reviewer. Record the original value, correction, reviewer, decision and time. Distinguish correction of extracted data from approval of a business action.
For example, an invoice might be readable but contain a total that disagrees with its line items. The illustrative response is to hold it for reconciliation, not silently change the amount or release payment. Review access, escalation times and a substitute owner so the queue does not depend on one person's availability.
Resume without repeating the business action
Keep a stable document and operation identity. After correction, revalidate the approved data and check whether the destination write already succeeded. Use tested destination duplicate protection where available. If the result is uncertain and cannot be reconciled safely, hold it for manual investigation.
Test acceptance, rejection, correction, repeated upload and restart. Measure field errors on reviewed samples, review time, oldest unresolved item and duplicate writes. Compare results separately for each supported document group; aggregate accuracy can hide a weak category. Agree when to pause or return to manual handling.
Read next and sources
Continue with recovery without duplicates and ongoing costs.
Reviewed 21 October 2026. Microsoft's Document Intelligence transparency note explains evaluation, limitations and human oversight. It is background for assessing a tool, not evidence that it meets your workflow's acceptance criteria.