Document capture turns invoices, forms and other files into data that a business can search or use in a workflow. The output is a prediction about the source. Before it becomes a saved record or triggers an action, check that the required information is complete, correctly interpreted and suitable for that action.
A confidence score can help prioritise review, but it does not establish the business meaning of a field. This guide explains how to assess extraction, preserve evidence and design a practical review process.
Define the required result
List the fields and structures needed by the receiving process. Identify required values, acceptable formats and checks between related fields. For an invoice, correctly reading a number is different from establishing whether it is the total, a line quantity or a reference.
Record how missing or ambiguous information should be handled. Do not invent a value to satisfy a required field. Distinguish an intentionally empty field from a value the extraction process failed to find.
Test representative original files
Include the formats, suppliers, layouts and quality levels the business actually receives. Digital PDFs, scans and photographs can behave differently. Check rotated pages, small print, handwriting, repeated labels and documents containing several records.
Compare extracted output with the source, including the association between labels and values. A visually similar document can contain a meaningful structural difference. Keep difficult examples as reviewed cases rather than assessing only clean demonstration files.
Interpret confidence at the appropriate level
Platforms may provide separate scores for words, fields, document types, tables or other structures. Availability and meaning depend on the service, model and version. Check the documentation for the returned score rather than treating every number as the same measure.
A high score can still accompany an inappropriate business value, and a missing score is not a pass. Use observed results on your own reviewed examples to decide how confidence supports routing. Avoid setting a universal acceptance threshold without understanding the consequences of errors.
Check completeness and relationships
Validate required fields, dates, currencies and expected ranges where those rules genuinely apply. Compare related amounts or identifiers with the receiving system. These checks should expose inconsistencies rather than silently changing extracted values to make the record appear plausible.
Keep extraction confidence separate from workflow validation. A recognised invoice number may still refer to a duplicate or an unknown supplier. The ability to read a document does not grant authority to approve its contents.
Inspect tables as connected structures
Rows, cells and headings need to remain correctly associated. Merged cells, optional columns and multi-page tables can introduce errors even when individual words are readable. Review the complete table and the lines used to construct the business record.
Test omitted values and page boundaries explicitly. An empty predicted cell can mean a genuine absence or a missed value. A high score for one cell does not establish that the whole row has been captured correctly.
Preserve traceable source evidence
Keep an authorised reference to the original file and relevant page. Text spans and bounding regions can help a reviewer locate an extracted value, but they are not a substitute for durable access to the original evidence.
Check that the reference continues to work after processing and respects the source's access restrictions. Define retention and deletion for originals, extraction results and review records together. Do not make sensitive documents publicly accessible to simplify review.
Make review a usable step
Show the proposed value, relevant source area and reason for review. Allow staff to correct or reject the result and record the decision. Route missing required fields and inconsistent relationships as well as low-confidence predictions.
Define which actions require approval and enforce the hold in the receiving workflow. Review capacity must match expected volume, including busy periods. A growing queue should have an owner and an escalation route.
Measure errors and improve the process
- Compare accepted values with independently reviewed source evidence.
- Track corrections by field, format and document type.
- Measure unnecessary review alongside missed errors.
- Retest after model, template or configuration changes.
- Keep new difficult examples in a regression set.
- Verify the final saved record, not just the extraction response.
Our AI evaluation guide helps separate model scores from checked outcomes. The AI data settings guide covers the information path, retention and access questions behind a responsible capture workflow.
