PO intake agentEmailed PO PDFs → AI reads them → ERP checks → rep review → one CSV per order

Before

  • Purchase orders arrive as PDFs from dozens of customers, and no two look alike: system printouts, faxes, crooked scans, cover sheets, handwritten forms.
  • Customer service retypes every PO into the ERP by hand, translating customer and old part numbers, units and prices from memory.
  • Price mismatches, new ship-to addresses and missing PO numbers are caught by eye, if they're caught.

After

  • An AI vision model reads each page: it spots POs in the inbox, turns rotated pages upright, skips cover and terms pages, and fills a fixed set of fields with a confidence for each.
  • Plain code checks everything against your ERP data: customers, ship-tos, item numbers (customer and legacy too), units and exact prices.
  • Anything unsure is flagged with a plain reason for a rep. Clean orders come out as one CSV per order in your ERP's import format.
What's real here. Every result on this page is a recorded run of a real AI model (…) on the sample PDFs shown, with its real latency and cost. AI reads the documents; plain code does the checks; a person approves. Handwritten orders always go to a rep. The samples are fictional and few, so the accuracy below is a smoke test, not a promise: a real project measures it on 50–100 of your own POs first.

1 · Order inbox (the model sorted each email: PO, not a PO, or unsure)

FromSubjectAttachmentModel saysWhy
Open a PO:

Measured accuracy on the samples

FieldCorrectAccuracy
SampleFieldsPagesERP matchStatus

How it works

  1. Triage. A small model call reads each email's sender, subject, body and attachment names and labels it PO, not a PO, or unsure. Unsure goes to a person.
  2. Prepare pages. The PDF becomes page images. The model labels each page (PO, cover, terms) and says how it's rotated; code turns it upright and drops cover and terms pages.
  3. Read. A vision model fills a fixed JSON schema (PO number, customer, bill-to, ship-to, dates, ship via, carrier account, notes, line items) with a confidence for every field. No per-customer templates.
  4. Check. Plain code matches the customer (fuzzy name + address), the ship-to on file, items via your item number, customer part crosswalk or legacy part list, converts units, and compares prices exactly to the price list.
  5. Flag. Low confidence, missing PO numbers, new ship-tos, unknown items, unsold units and price mismatches each get a plain reason. Handwritten orders always go to a rep.
  6. Write. One CSV per order, in a fixed column format, dropped where the ERP import already looks. Flagged orders wait for a rep.

Production architecture

Architecture: inbox triage, document prep, AI reading, plain-code ERP checks, rep review and CSV to the ERP import folder