Prepare the document taxonomy
Supply file types, naming conventions, expected document sets and field definitions. Agree page boundaries, version precedence and identifiers that connect each extracted value to the source document or page.
Separate extraction from verification
Check required pages, signatures or fields only to the documented administrative standard. Route unreadable text, inconsistent identifiers and missing attachments for decision. Preserve source evidence rather than filling gaps from assumption or unrelated records.
Assess document complexity
Review scans, page counts, layouts, languages and sensitive-data requirements in a pilot. Define whether acceptance and pricing use pages, documents, cases or extracted records, including treatment of duplicate submissions and corrected versions.
Define the document set as well as the extracted fields
A document taxonomy should distinguish type, case association, version and page range. State how to treat a single PDF containing several documents and several files containing one document. File names alone are weak evidence: “final” may be outdated and a five-page scan may omit a required numbered page.
Keep classification, extraction, administrative completeness and substantive review as separate results. An application form can be correctly classified and fully extracted while its supporting evidence is missing. A complete document set can still require a qualified decision on eligibility, legal sufficiency or validity. Administrative statuses must not imply those decisions.
Illustrative example: an application package with revised pages
A hypothetical case AP-22 requires application pages 1–3, an identity attachment and a signed declaration. The approved taxonomy assigns codes APP, ID and DEC. The input contains APP-v1 pages 1–3, APP-v2 pages 1–2, ID-v1 and DEC-v1 without the signature page. The client’s rule says an explicit newer version supersedes the older form as a whole; pages from v1 cannot silently fill the v2 gap.
The operator indexes every file, extracts the case ID from APP-v2 page 1 and marks the package incomplete: APP-v2 page 3 and the declaration signature page are missing. The case owner supplies both through approved intake. The operator updates the index, checks page sequence and links extracted fields to the final source pages. Acceptance is the corrected index, field file and completeness evidence. The case remains “ready for client review,” never “eligible” or “approved.”
| Document | Initial condition | Treatment / accepted evidence |
|---|---|---|
| APP-v1 | Pages 1–3 present, older version | Retain as superseded; exclude from current extraction |
| APP-v2 | Pages 1–2 present, page 3 missing | Hold completeness; link supplied replacement page 3 |
| ID-v1 | Required attachment present | Index document; no eligibility conclusion |
| DEC-v1 | Signature page missing | Request missing page through case owner |
| AP-22 package | Two missing-page exceptions resolved | Ready for client review; source links preserved |
Make each extracted value recoverable from its source
Store document ID, version and page with the extracted value. If a field spans pages or requires a permitted calculation, record that method and its inputs. OCR confidence can help route review, but it is not acceptance evidence by itself; a confidently misread identifier is still wrong.
Use explicit unreadable and conflicting-value codes. When two current documents show different names or dates, preserve both values and ask the authorized owner which source controls. The operator should not select the more plausible value or combine pages from different versions without an approved assembly rule.
Count pages, documents and cases without confusing them
A case can contain several documents and many pages; a replacement can change the page count without creating a new case. Delivery and billing registers should show the agreed unit and how duplicates, blank pages and revised versions are treated. A page-processing total does not indicate how many packages are complete.
At handover, reconcile supplied files, classified documents, current versions, extracted records and open exceptions. Identify the client reviewer for each incomplete case. Agree retention and return arrangements for originals and working copies so the final structured dataset does not become detached from its authorized evidence.