Invoices arrive from three separate channels, get OCR'd and structured by AI, checked for duplicates before anything is booked, routed for human approval in Slack, and archived — with a full audit trail of every rejection, duplicate, and review.

Manual invoice entry is slow and error-prone. Invoices arrive from email, an upload portal, and a shared drive folder with no single intake point. Duplicate invoices get paid twice without a systematic check, and finance has no consistent approval trail for what got paid and why.

Invoices can enter through a webhook upload, a monitored Gmail inbox, or a watched Drive folder — each normalized into the same internal format so the rest of the pipeline doesn't care where the file came from. The file is validated, OCR'd, and the extracted text is passed to an AI extraction step that returns structured fields (vendor, amount, date, line items) against a strict schema. Before anything is saved, the system checks for a duplicate invoice; if one exists, finance is notified and the record stops there instead of double-booking. Valid, non-duplicate invoices get an AI-generated summary and go to Slack for a one-click approval decision — approved invoices are marked and archived, rejected ones notify the submitter, and anything ambiguous routes to human review rather than being forced through.

Three intake channels normalized to one pipeline — no change needed in how the business sends invoices

OCR + structured AI extraction, not a single free-text prompt — line items and totals come back schema-validated

Duplicate detection before booking — the single most common cause of AP overpayment, handled automatically

Human-in-the-loop approval — AI drafts and summarizes, a person still authorizes payment

Original document reconstruction and archival, not just the extracted data

Full audit trail — every duplicate, validation error, and human review is logged