Which Extracted Fields Deserve a Human Check in Skavio Flow
Skavio Flow turns invoices, orders, scans and text into reviewed XLSX, CSV and JSON. The operative word is reviewed. A value in a clean spreadsheet is still a draft until someone confirms it against the source. The practical question is not whether an extraction looks tidy; it is what goes wrong if the field is wrong.
Treat review as a risk decision, not a blanket ritual
Applying the same attention to every cell wastes effort and dulls attention where it matters. A stronger policy starts with consequence. If a supplier name, total, tax amount, due date or invoice reference is wrong, payment, reconciliation or follow-up can go astray. Those fields deserve mandatory review. A slightly awkward product description usually carries a lower cost; it can be sampled or checked by exception. That is a policy decision inside your company, not something the extraction tool should silently make for you.
- Mandatory review: fields that determine payment, alter totals, identify a counterparty or set a deadline.
- Sampled review: descriptive fields whose errors are visible and lower-consequence, checked against a documented sample.
- Escalation: any field tied to a review warning, incomplete source evidence or an ambiguous tax or currency label.
The five anchors: money, identity, reference, date and tax
For supplier invoices and purchase orders, start with a fixed anchor list. Check header totals against line-item sums; confirm supplier identity; verify invoice, order and contract references; distinguish issue dates from due dates; and inspect tax rates, tax amounts and currency labels. These checks do not require a person to re-key every row. They require a decision on the values that change what the business does next. A mismatch should become a visible exception, not a quiet edit.
- Money: header total, line totals, quantities and units; investigate unexplained differences before export.
- Identity and references: supplier name, invoice or order identifiers and any payment reference your process relies on; do not fill missing references by inference.
- Date and tax: separate issue date from due date, and flag unclear tax labels or mixed currencies for the responsible finance owner.
Use source evidence and warnings to resolve exceptions
Skavio Flow outputs include source evidence and review warnings alongside XLSX, CSV and JSON. Those outputs are review inputs, not decoration. Embedded PDF text is read locally, scans use Worker Fabric OCR, and structured extraction uses the configured OpenAI model. That pipeline affects where a reviewer should look, but it is not evidence that every value is correct. A clean-looking row still needs the five anchors above. A warning needs a disposition: correct the value, reject the document or involve a second reviewer. In a verified acceptance on 1 October 2026, two synthetic PDF invoices were extracted and exported; a correction to a single extracted value regenerated the results without a second job. That is useful workflow evidence, not a substitute for approval.
- Review warnings before signing off an export, and keep a separate company record of what was checked and why an exception was accepted or rejected.
- Compare disputed values with the source page or scan rather than correcting from memory.
- Regenerate exports after a correction, and keep the reviewed output paired with its evidence.
Make the policy repeatable without pretending it is automatic
Custom extraction fields and saved company workflows help standardize field names and checkpoints across batches. Use them to make the review policy consistent: which fields are mandatory, which are sampled, who resolves exceptions and what must be true before XLSX, CSV or JSON leaves the workspace. They should not be described as automatic approval controls. The handoff remains a human decision. For lower-risk descriptive fields, sample a defined number per batch and review every field connected to an exception. For the five anchors, check every document. That is a reasonable starting point for a small finance or operations team; adjust it to your own error tolerance and downstream process.
- Standardize field names for money, supplier identity, references, dates and tax so reviewers see the same checkpoints in every batch.
- Keep mandatory review and exception handling separate from descriptive-field sampling; do not let a sample stand in for an anchor check.
- Export only after the reviewed output and its source evidence are paired for the finance or operations handoff.