Extract receipt and invoice data with checks before posting

Register the original before cleaning or extracting it

Assign a document ID and record the arrival channel, original filename, received time, page count and file fingerprint. Keep the original and any improved scan linked rather than replacing one with the other. Check that every expected page is present and that the total, document number and supplier details are readable.

Restrict access to the people and services involved in the process. Agree retention and permitted processing before sending documents to an external service. A cropped preview can assist review, but retain access to the whole document so a reviewer can notice a credit note label or a qualification outside the selected area.

Choose fields for the document type and keep their evidence

Receipts and invoices answer different questions. Microsoft’s Document Intelligence references describe separate models: invoice extraction includes invoice fields and line items; receipt extraction includes merchant, transaction date and total. Verify your own formats, languages and required fields rather than assuming one model covers every document.

For each material field, keep its raw text, normalised value and source page or location. Leave absent values empty and distinguish them from extraction failures. Keep invoice total and amount due separate: deposits, previous payments and credit balances may make them different. A currency symbol alone may be insufficient to identify the currency.

• Document type, supplier identity and document number.
• Issue date, due date and currency where actually stated.
• Line descriptions, quantities, unit amounts and line totals.
• Subtotal, stated tax, discounts, charges, total and amount due.
• Source locations, missing-field reasons and extraction version.

Use arithmetic and document identity as independent checks

Calculate with decimal currency rules and an agreed rounding tolerance. Compare quantities and unit amounts with line totals where the document supplies them. Reconcile subtotal plus stated charges and tax minus stated discounts with the displayed total, using the document’s actual structure. Do not add tax again when the printed amounts already include it, or repair a discrepancy by inventing a balancing line.

A file fingerprint catches an identical upload, but a new scan of the same invoice needs a second check. Compare supplier ID and invoice number, then inspect date, currency and amount. Receipts without a reliable document number need a different review rule. Flag suspected duplicates; preserve credit notes and revised documents as related records until their status is resolved.

Worked example: a clean-looking extraction with a wrong total

Hypothetical example, not a client project: an invoice shows a USD 300 subtotal, a USD 20 discount, a USD 10 delivery charge and no separate tax amount, with a total of USD 290. Extraction returns USD 390. The arithmetic check expects 300 − 20 + 10 = 290 and sends the draft to review. The reviewer opens the source total and corrects the field; the workflow does not silently choose its calculated result.

The next upload is a phone photograph of that invoice. Its fingerprint differs, but the supplier and invoice number match the accepted record. Hold it as a possible duplicate. If the document is actually a revised invoice, retain both versions and ask the finance owner how the revision should replace or adjust the earlier record.

Make review and recovery visible parts of the workflow

Microsoft’s confidence guidance describes estimated confidence and recommends considering human review; not every field has a score. Treat confidence as one routing signal. A confident extraction can still concern the wrong page, wrong document or an invoice that should not be paid. Test thresholds against your reviewed sample rather than copying an arbitrary percentage.

Give reviewers the source beside the draft, check results and a reason for the exception. Record accepted, corrected, rejected and waiting-for-source states with an owner. During service failure, retain documents as pending extraction. During export failure, resume from the recorded export step and check the destination ID before replaying. Missing sources and unknown posting outcomes require investigation, not an automatic retry loop.

Measure checked records before expanding to posting

Create a baseline for entry plus checking, using a mix of suppliers, scans and exceptions. Required-field accuracy = correctly extracted required values divided by required values checked. Also measure whole-document acceptance without correction, duplicate flags resolved, review minutes per accepted document and the age of unresolved documents. An average field score can conceal a wrong total.

Compare total effort, including rescanning, corrections and failed exports. Pause expansion if material amount or supplier errors escape review, duplicates reach the destination, sources cannot be reopened or the exception queue exceeds review capacity. Keep the manual finance process available. Move to posting only after separate approval of field mappings, permissions, reconciliation and reversal procedures.