Skavio Flow: check the evidence before accepting invoice and order fields
Skavio Flow turns invoices, orders, scans and text into structured results with source evidence and review warnings. For finance and procurement teams, the important step comes between extraction and export: checking whether each consequential value is supported by the original document. A tidy spreadsheet is an output format—not proof that its contents are correct.
Read the value and its source together
Flow supports custom extraction fields and saved company workflows, with XLSX document and line-item worksheets, CSV and JSON outputs. Source evidence gives reviewers material to inspect alongside the extracted results. Start with the fields that determine how a document will be handled: supplier identity, invoice or order number, dates, currency, totals and line items. Compare the available evidence with the original document. Check not only whether the characters match, but whether they belong to the right field. An order reference is not necessarily an invoice number; a delivery date is not necessarily a payment due date. Microsoft’s Document Intelligence transparency note explains why these checks matter: a single-character error in an invoice amount can make the invoice incorrect even when document-wide text accuracy appears strong. That is external context for reviewing consequential fields, not a measurement of Flow’s accuracy.
Use warnings to direct attention—not to certify the rest
Inspect Flow’s review warnings before accepting the results. Return to the source and establish what can actually be confirmed. If the document is unclear or incomplete, a plausible value is not a substitute for evidence. Use warnings to guide attention, but do not treat the absence of a warning as proof of correctness. Do not assume that every erroneous field will be flagged or that evidence will be available for every value. This distinction matters especially for scanned paperwork. Microsoft documents that resolution, contrast, rotation and text characteristics can affect OCR accuracy. Where a character or amount is difficult to read, inspect a clearer original or request clarification rather than guessing. NIST’s Generative AI Profile also identifies over-reliance on automated output as a risk: confident presentation does not establish correctness.
An illustrative review: dates, references and totals
Consider a hypothetical supplier invoice with the date “04/05/2026,” an order reference near the invoice number and a faint decimal separator in a line-item amount. These are illustrative reviewer scenarios, not verified Flow warning categories. The review task is to resolve meaning before handing the data onward. Use the following checks for an invoice or purchase order, adjusting the fields to suit the document:
- Dates: confirm which date is being extracted and whether the source establishes day/month or month/day order. If it does not, seek clarification.
- References: distinguish invoice numbers, purchase-order numbers and supplier references. Do not fill a missing identifier with a nearby but different reference.
- Currency and totals: confirm the currency label and compare the extracted total with the amount stated in the document.
- Line items: check descriptions, quantities, units, unit prices and amounts against the source—not just the overall total.
- Arithmetic: manually compare line-item sums with the stated totals, accounting for tax, discounts and charges as the document presents them. This is a reviewer check, not a claim of automatic reconciliation.
Correct supported values, then export the reviewed result
Resolve discrepancies using the source before preparing the final handoff. In an internal acceptance test on October 1, 2026, two synthetic PDF invoices were extracted and exported, and a scalar correction regenerated results without a second job. That verifies the tested scalar-correction path—not every field-editing, line-item or bulk-correction scenario. After review, choose the output that suits the receiving process: XLSX for document and line-item worksheets, CSV for tabular exchange, or JSON for structured data. Keep the original document available for follow-up questions, and communicate unresolved issues rather than presenting uncertain values as confirmed. The boundary is straightforward: extraction proposes structured fields; human review determines what is ready to use. Export does not imply automatic posting to an accounting system or ERP.