A parks AP review agent needs structured vendor fields such as amounts and invoice numbers from PDF invoices. What should the extraction step provide?

A parks accounts-payable review agent needs field extraction from vendor invoice PDFs that yields structured vendor fields such as amounts and invoice numbers. Semiautonomous AP review depends on those fields for matching, thresholds, and exception queues; raw table cells, layout text, or generic pairs do not pay or reject invoices. Table cell extraction from vendor invoice PDFs pulls grid rows and columns from line-item tables but does not identify the structured vendor fields such as invoice number and totals, so three-way match still has nothing to compare. Layout text extraction from vendor invoice PDFs returns page text with positions but not as structured fields: the agent cannot reliably consume a dollar amount or invoice number as a typed value for matching. Key-value extraction from vendor invoice PDFs surfaces generic label-value pairs found on the page but does not provide the normalized invoice schema — a specific dollar figure, a specific invoice number — in a schema the AP system can key off of. Exam caveat: municipal AP agents over invoice PDFs need field extraction of amounts and identifiers — not raw table cells, layout text, or generic key-values. Operational check: extract invoice number, date, vendor, line amounts, and totals from parks vendor PDFs; validate against a labeled sample; feed fields into the AP review agent with dual-control thresholds; keep table-cell, layout-text, and generic key-value outputs as metadata, not as the source of the numbers the agent acts on. Normalize currency and invoice-number formats during extraction so duplicate-pay detection and purchase-order matching work across vendor template variants used by parks contractors.