Checking an Agent Tools extraction result against the original document
An asynchronous extraction API solves one problem and creates another: the output arrives before anyone has decided whether it is fit to use. Skavio Agent Tools makes processed results available through authenticated downloads after a job completes, so verification has to be designed into the pipeline rather than added after an export has already reached accounting. This guide sets out a practical review sequence: confirm the document frame, reconcile the fields a finance team will act on, check the arithmetic, and record a named approval before the result leaves the system.
The result is a draft until the job is complete
Agent Tools is asynchronous. A submission acknowledgement tells you that the API accepted a request; it does not establish that a document has been read or that a result is ready. In a verification-first workflow, keep four events distinct: submission, completion, authenticated download, and human approval. The server quote issued at submission is authoritative for the job. A successful job settles once, while a permanent failure releases the reservation. A data-quality problem found during review is a review event, not automatically a billing event. Use idempotent submission identities so that a client retry after a timeout does not create another job for the same document. Fetch the result with a key scoped to the same company that submitted it, and avoid passing download credentials through chat, logs or browser code. Record the job identifier alongside the source document before anyone opens the output.
- Treat submission acknowledgement and queued states as process information, not as a completed result.
- Use an idempotency key when retrying a submission so a network timeout does not become a second job.
- Download with a company-scoped key and keep that credential out of shared channels.
- Log the job ID, source file, submission time and reviewer before field inspection begins.
Check the document frame before the field values
Start with the coarser facts. Is this the right file? Does the page count match the source? Are the pages complete and in the expected order? A plausible invoice number on the wrong page is still the wrong result. If the documented Agent Tools response exposes page references, warnings or missing-field indicators, verify them against the source. If the published schema does not expose a particular field, do not invent it or assume another Skavio product's response shape applies. Map the output back to the original PDF or image manually where necessary. Look for accidental duplicate submissions and for mixed batches where a credit note sits beside a positive invoice. The documented shared-processing retention window is seven days, so the review step must be scheduled with completion rather than left for a later week. For larger batches, plan around the documented upload, queue, active-job and storage limits before promising an internal turnaround.
- Confirm file identity, page count, page order and completeness against the source.
- Treat warnings, nulls and ambiguous matches as exceptions, not as values to fill from context.
- Check for duplicate documents and for credit notes mixed into ordinary invoice batches.
- Schedule human review within the documented shared-processing retention window.
Reconcile the fields a finance team will act on
Field review is not a spot check of the most visible number. Compare supplier identity, document number, any purchase-order or contract reference, issue and due dates, currency label, tax label and the reported total against the source. If a value is blank, ambiguous or contradicted by the page, route it to a named reviewer rather than letting a guessed value fill the gap. Currency labels need particular care because a label is not a conversion; if the document does not state a rate, do not add one. Tax labels should become review exceptions, not tax advice. Do not infer per-field evidence or confidence metadata unless the published Agent Tools response schema defines it. The review record should say what the reviewer compared and where. This is slower than accepting a table at face value, but it keeps the handoff honest.
- Verify supplier identity, document number and reference numbers against the source page.
- Check issue dates, due dates, currency labels and tax labels without inventing missing context.
- Send blank or contradictory fields to a named person instead of silently completing them.
- Use only the response fields that the published Agent Tools documentation actually defines.
Make arithmetic and approval explicit
After field review, reconcile line-item arithmetic with the reported header total. Check signs on credit notes, quantities and units, and whether tax is stated per line or once at header level. A mismatch should become a visible exception, not a silent adjustment. Then record the approval: reviewer, time, job ID, fields changed, evidence checked and decision. Keep the approved result and the review log together. Keep the final write into accounting outside the extraction step and behind a human gate; an extraction API can prepare structured data, but it should not be treated as an approved accounting entry or an automated posting. If a document is unreadable, reject it or request a better scan rather than filling missing text from context. The difference between a controlled pipeline and an unreviewed data source is not the absence of automation. It is the presence of a named human decision before the result is used.
- Sum the line items and compare the result with the reported header total.
- Review signs, credit notes, quantities and units as explicit exceptions.
- Record reviewer identity, timestamp, changed fields and approval decision.
- Do not treat a downloaded result as an approved accounting entry or an automated posting.