Purchase order OCR when every customer's PO looks different
Purchase order OCR turns a PO PDF or scan into data you can use. Template OCR works when every PO looks the same. When every customer sends a different layout, AI reading copes better, as long as it tells you which fields it is unsure of.
Template OCR and AI reading compared
Both turn a document into data. They differ in how they find the data, and that difference decides how much setup and checking you carry.
| Compared on | Template OCR | AI reading |
|---|---|---|
| Setup per layout | A template drawn for each customer's layout: where the PO number sits, where the table starts | None. The document is read as a person would read it |
| A new customer's layout | Fails or misreads until someone builds a template | Reads on the first order |
| Scans and faxes | Skew and noise shift fields out of their boxes | Reads through skew and noise; lookalike characters still need a check |
| Handwritten changes | Usually missed, so the printed value goes through | Often caught, and flagged for a person to confirm |
| Telling you when it is unsure | Character confidence scores, if any, not field-level reasons | Should flag each uncertain field with a plain reason. Ask to see it |
Template OCR still makes sense when a few customers send most of your orders in fixed layouts that never change. For a long tail of buyers, each with their own procurement system, the templates never stop needing work.
What PO OCR should pull out
Whatever does the reading, a usable result has these fields, each one checkable:
- Header: customer name, PO number, order date, requested delivery date, ship-to and bill-to addresses, and currency.
- Each line: the customer's part number, description, quantity, unit of measure, unit price, line total and any date for that line.
- Notes: delivery instructions, freight, discounts and GST, kept out of the lines so the line count matches the items ordered.
Two details separate a careful reader from a careless one. Pack sizes such as "box of 50" must not be read as quantities, and lines quoted from an earlier email in the thread must not be read as part of this order.
Try it on one of your POs
Drop a PO from your inbox or paste an order email. Three reads are free and need no sign-up. Uploads are read in memory and never stored.
PO reader
Drop a PO file here
PDF, scan or photo (PNG, JPG, WEBP), TXT or CSV. Up to 10 MB. Photos over 5 MB are shrunk before sending.
A file is selected. Remove it to paste text instead.
0 / 60,000Read in memory and discarded. Never stored, never used for training. Privacy policy
The accuracy number that matters
OCR tools are often sold on character accuracy: one high percentage for how many characters come out right. On a purchase order that figure says little. A single wrong character in a part number or a quantity makes the whole line wrong, and a line that needs an edit costs nearly as much time as keying it.
The number to ask for is the share of lines right with no edit, measured on your own documents: a month of real orders, compared line by line with what your team keyed. Pair it with a second question: of the lines that were wrong, how many did the tool flag as uncertain? A wrong line that was flagged costs a glance. A wrong line that was not flagged costs a credit note.
That is what a shadow test measures, on your orders rather than a demo set.
Getting a clean read from scans and photos
- Flat page. Lay the PO flat. Curled edges bend lines of text.
- Whole page in frame. Include every edge, so no column is cut off.
- Good light. Even light without shadows across the table. Avoid a flash glare spot on glossy paper.
- PDF over photo. If the customer can email the PDF their system produced, ask for it. It always reads better than a photo of a printout.
- One order per file. Split a batch scan into one file per PO.
More on the full flow in AI order processing, and on the options beyond OCR in order entry automation. The reader also has its own page: PO reader.
Purchase order OCR questions
What is purchase order OCR?
Purchase order OCR is software that turns a PO in PDF, scan or image form into structured data: customer, PO number, dates, addresses and line items. Classic OCR does it with a template per layout. AI reading does it by understanding the document, whatever its layout.
Is OCR accurate enough for purchase orders?
Character accuracy on a clean PDF is high, but that is not the number that matters. What matters is the share of order lines that come out right with no edit, measured on your own documents, and whether the tool tells you which fields it is unsure of.
What is the difference between template OCR and AI reading?
Template OCR is told where each field sits on a known layout, so it needs a template per customer and breaks when the layout changes. AI reading works out what each part of the document means, so a new customer's layout reads on the first order.
Can it read scanned or faxed POs?
Yes. Scans, faxes saved as PDF or image and phone photos all work. They read less cleanly than digital PDFs, mostly on lookalike characters such as O and 0 or I and 1, so those fields come back flagged with the reason.
Which fields should PO OCR extract?
At least: customer name, PO number, order date, requested delivery date, ship-to and bill-to addresses, currency, and for each line the part number, description, quantity, unit of measure, unit price, line total and any line date. Freight and tax rows should be kept out of the lines.
Do I need to set it up for each customer's layout?
With template OCR, yes, and again whenever a customer changes their system. With AI reading, no. Each PO is read on its own terms, and a new layout needs no setup.
See it on your own orders
A free shadow test runs a month of your emailed orders through the agent. Read-only, about a week.