Start with the right question
Extracting fields from a bundle looks mechanical, but a model may read a subtotal as a balance or combine two people with similar names. Design the extraction table so an incorrect value can be found and corrected without rebuilding the entire case summary.
Work through the checks
Define each field
State whether an amount means billed, paid, outstanding, or estimated. Specify the date type and currency. Ambiguous field names encourage internally consistent but practically wrong results.
Require provenance
Attach a filename and page or row reference to each value. Ask for an explicit missing marker when no source exists. A plausible number without provenance should remain outside verified calculations.
Inspect edge cases
Review handwritten notes, low-quality scans, tables spanning pages, negative amounts, and repeated statements. These are useful places to look for extraction errors even when ordinary paragraphs appear accurate.
Reconcile duplicates
Match invoice numbers, service periods, and payment references before summing. Keep the duplicate record but mark its relationship to the original so later reviewers understand why it was excluded from a total.
A record you can use
- Field definition
- Extracted value and unit
- Source page
- Verified value
- Duplicate or missing flag
Keep this record locally. Use document references rather than publishing account numbers or other people's personal details. An empty field means the information is still missing; it should not be filled with an AI guess.
Worked example
A fictional bundle contains a $400 invoice and a later statement showing the same $400 plus a $100 payment. An extractor totals $800 in charges. The reviewer links the documents, records one charge and one payment, and leaves $300 as the arithmetic balance rather than double-counting the invoice.
This example is fictional and illustrates the method, not a real case result.
Limits and next steps
Extraction accuracy does not determine whether a cost is legally recoverable. Keep disputed items and unverified values separate. Do not upload documents you are not authorized to process.
Review the extraction table before using it in a chronology, ledger, or AI prompt. Preserve the raw output and corrections so a later reviewer can reconstruct the calculation.