peekgenticBook a call

Analysis · AI in business

OCR vs. IDP for packing slips: a 20-document test

Before you buy anything to read packing slips, run a test with 20 of your own documents. OCR turns a picture into text. IDP also knows which number is the PO and which is the quantity, and says how sure it is. How to pick the 20, how to score them, what the big three cloud services charge per page, and what a passing result looks like on a receiving dock.

By Miguel GutierrezCo-founder, Peekgentic · Dallas
· 8 min read
Preview card: OCR vs. IDP for packing slips, a 20-document test.
Test with your own paper: clean slips, new layouts, handwriting, phone photos and the ugly ones.Peekgentic

In brief

OCR (optical character recognition) turns a scan or photo into text. It doesn't know which number is the PO and which is the quantity. IDP (intelligent document processing) reads the same page, pulls out named fields and tables, and gives each one a confidence score so a person can check the shaky ones. For receiving paperwork you almost always need IDP, because packing slips change layout from supplier to supplier and the received quantity is often written by hand. Don't take a vendor's word for it: run 20 of your own documents through it, score every field, and count the confident wrong answers. That number decides it.

  • OCR gives you text. IDP gives you named fields, tables and a confidence score for each one.
  • At list price, plain OCR is about $1.50 per 1,000 pages; field and table extraction runs about $10 to $70.
  • Microsoft, Google and Amazon list no prebuilt model for packing slips or bills of lading. Expect a general or custom model.
  • Test with 20 of your own documents, scored against an answer key. Count the confidently wrong answers.
  • The review step is yours to design. Amazon closed its review service to new customers and Google deprecated its version.

Why receiving paperwork is a hard test

A packing slip looks simple. It isn't. Every supplier uses its own layout. Some put your PO number at the top, some bury it in a reference field, some leave it off. The quantity shipped is printed, but the quantity your dock actually counted is often written in pen next to it. Add a bill of lading, a coffee stain, a phone photo taken at an angle, and a stamp across the middle.

That's why the question "OCR or IDP?" matters here more than almost anywhere else in the office. And it's why the only honest way to answer it is with your own paper, not a demo file.

OCR vs. IDP in plain English

OCR turns a picture of a page into text. Amazon's own pricing page describes its basic text API as one that uses OCR technology to extract text and handwriting. You get the words and where they sit on the page. You don't get "this is the PO number."

IDP goes a step further. Google describes its Document AI as taking unstructured data from documents and turning it into structured data. Microsoft's Layout model extracts text, tables, selection marks and other structure, and its other models return named fields. Amazon Textract's analysis features pull out forms and tables, not just lines of text.

The piece that matters most for a receiving dock is the confidence score. Microsoft says its service returns an estimated confidence for words, key-value pairs and selection marks. A score of 0.95, Microsoft explains, means the prediction is likely right 19 times out of 20. Microsoft says you can use that score to decide whether to accept a value automatically or flag it for a person.

If you want the longer background, we explain the category in what intelligent document processing is, in plain English. This page is about testing it.

OCR vs. IDP on a receiving dock

OCRIDP
What you get backLines of text and their positionNamed fields (PO, item, qty) and tables
New supplier layoutSomeone writes a new rule or templateOften handled; test it
Line itemsText in rows; you rebuild the tableTable rows with columns
How sure it isPer word, at bestPer field, so you can route doubtful ones to a person
Price per 1,000 pages (list, first tier)About $1.50About $10 to $70, depending on features

What the big three actually charge

Price is usually the first objection to IDP, so here are list prices from the vendors themselves, checked Nov. 14, 2026. Amazon's figures come from the worked examples on its Textract pricing page (US West, Oregon). Microsoft's come from its official Azure retail prices feed (East US), because the pricing page itself shows placeholders until you pick a region. That page does confirm the first 500 pages a month are free.

List price per 1,000 pages, first pricing tier

Plain OCR is cheap. Fields and tables cost more. Prices checked Nov. 14, 2026.

  1. Azure Read (OCR)$1.50
  2. Textract text detection (OCR)$1.50
  3. Azure prebuilt / Layout$10
  4. Textract tables$15
  5. Azure custom extraction$30
  6. Textract forms$50
  7. Textract tables + forms + queries$70

Sources: AWS Textract pricing (US West, Oregon, first 1M pages); Azure retail prices API (East US, first 1M pages for Read). List prices; contracts and regions differ.

Put that in dock terms. If you receive 1,500 pages of paperwork a month, the gap between plain OCR and the most expensive option above is roughly $2 versus $105 a month at list price. Real projects cost more than the per-page fee, because someone has to connect the result to your ERP and build the review screen. But the per-page fee is rarely the reason to pick the cheaper tool.

None of them has a "packing slip" button

This surprises people. Microsoft's list of prebuilt models covers invoices, receipts, IDs, tax forms, bank statements, contracts and a few others. We found no prebuilt model for packing slips, bills of lading or delivery notes. Google's processor list has invoice, expense and form parsers, a custom extractor, and no packing-slip processor. Amazon's specialized APIs cover expenses, IDs and lending documents.

So for receiving paperwork you'll use a general tool (layout, form or table extraction) or train a custom model. Microsoft says a custom model needs only five examples of the same document type to get started. Five examples per supplier layout adds up fast if you have 60 suppliers. That's exactly what the 20-document test should reveal.

Handwriting: check the fine print

On a receiving dock, the most important number is often the one written in pen: what you actually counted. Read the language limits before you assume it works.

Numbers and checkmarks are easier than words. But a 7 that looks like a 1 is still a wrong count. Put your real handwritten slips in the test.

The number that decides it isn't accuracy. It's how many wrong answers came back marked as sure.

The 20-document test

This is the test we'd run before any receiving project, ours or anyone else's. It takes an afternoon to gather the paper and a day to score it.

1. Pick the 20 on purpose

A fair 20-document mix

CountWhat to includeWhy
5Clean, typed packing slips from your top suppliersThe easy case. If it fails here, stop.
5Slips from five different, smaller suppliersTests new layouts, not just the ones it has seen
3Slips with received quantities or notes written in penTests handwriting where it matters
3Phone photos taken on the dock, not scansThat's how the paper really arrives at the office
2Multi-page bills of lading or slips with long line listsTests tables that run across pages
2The ugly ones: stamped, folded, faxed, partly tornShows what it does when it can't read something

2. Write down the right answers first

Before you run anything, have someone on your team type the correct values for each document into a spreadsheet: supplier, your PO number, packing slip number, each line's item number, quantity shipped, quantity received, and lot or serial number if you track them. This is your answer key. It's boring, and it's the only way the test means anything.

3. Run all 20 through each option

Same 20 documents, every option you're considering: plain OCR, a cloud IDP service, a vendor's product, a custom build. Don't let anyone hand-pick or clean up the files first.

4. Score every field three ways

  • Right. Matches the answer key.
  • Flagged. Wrong or missing, but marked low confidence, so a person would have checked it.
  • Confidently wrong. Wrong, and marked as sure. This is the dangerous one.

Score the fields that touch inventory and money separately: PO number, item number and quantities. A wrong ship-to address is annoying. A wrong received quantity becomes a wrong stock count and a wrong payment.

5. Time the person, too

Time how long your receiving clerk takes to key the 20 today. Then time how long it takes to check the flagged fields on the screen each option gives you. That difference is the real saving.

What a passing result looks like

There's no industry standard here, so this is our rule of thumb, not a published benchmark:

  • Zero confidently wrong answers on PO number, item number and quantity across the 20. One is a warning. Two means the tool, or the threshold, isn't ready.
  • Most fields right on the clean slips, and the misses on the ugly ones show up as flagged, not as quiet errors.
  • The person checking spends clearly less time than typing from scratch.

Vendors give similar advice on where to set the bar. Amazon's own guidance says business processes involving financial decisions might require confidence thresholds of 90% or higher. For its custom models, Microsoft says it's best to target an accuracy score of 80% or higher. Your test tells you whether those thresholds catch your errors.

Why "a person checks it" has to be designed, not assumed

Almost every pitch says a person stays in the loop. Ask how. The cloud vendors used to sell a ready-made review step, and that's changing: Amazon says its Augmented AI review service is no longer open to new customers, and Google's Human-in-the-Loop feature is on its deprecation list. So the review screen is something you or your builder provides.

And it has to be a real check. In a peer-reviewed 2011 study of 195 people, researchers found that checking data by eye left about 30 times as many errors as entering it twice and comparing. That study used typed research data, not packing slips, so treat it as a warning, not a dock statistic. The point holds: a screen that shows 40 fields and asks "look OK?" doesn't work. A good review screen shows the image next to only the doubtful fields and makes the person type or confirm those.

We cover that design in human-in-the-loop automation.

Where plain OCR is enough

To be fair to the cheaper option: OCR alone is fine when you only need the paper to be searchable. If the goal is "find the packing slip for PO 4471 when a supplier disputes a short shipment," OCR plus a good file name does the job for about $1.50 per 1,000 pages. It's also fine when every document comes from one supplier with one layout that never changes, and a simple rule can find the fields.

Where it breaks is the moment you want the numbers to flow into receiving in your ERP without someone retyping them. That's when you need fields, tables and confidence scores.

After the test

  1. If nothing passes, don't force it. Fix the paper first: ask your top five suppliers to put your PO number on every slip, or send an advance ship notice.
  2. If something passes, start with your top suppliers' slips only and keep the person checking every received quantity for the first month.
  3. Keep the 20 documents and the answer key. Rerun them every time the vendor changes its model or you add a supplier layout.

For the full receiving workflow, see automating packing lists and bills of lading, and measure minutes per receipt before and after, as in how to measure the ROI of office automation.

Common questions

What is the difference between OCR and IDP?

OCR turns a scan or photo into text. IDP also pulls out named fields like the PO number and quantity, reads tables of line items, and gives each value a confidence score so a person can check the doubtful ones.

Can OCR read packing slips?

It can read the text, but it doesn't know which number is the PO and which is the quantity. With many supplier layouts, you need IDP or a custom model to get fields and line items you can send to your ERP.

How much does IDP cost per page?

At list price, checked Nov. 14, 2026, plain OCR is about $1.50 per 1,000 pages on both Amazon Textract and Azure. Azure's prebuilt and layout models are $10 per 1,000 pages, custom extraction $30, and Textract's forms feature $50.

Is there a prebuilt model for packing slips or bills of lading?

We found none in Microsoft's, Google's or Amazon's lists. You use a general layout, form or table model, or train a custom model. Microsoft says a custom model needs five examples of a document type to start.

Can these tools read handwritten received quantities?

Often, but check the limits. Amazon says Textract reads handwriting only in English. Microsoft and Google list handwriting support language by language. Put your real handwritten slips in the test.

How do I test OCR or IDP before buying?

Pick 20 of your own documents on purpose, type the correct values into an answer key, run all 20 through each option, and score each field as right, flagged or confidently wrong. Aim for zero confidently wrong PO numbers, item numbers and quantities.

What confidence threshold should we use?

There's no single standard. Amazon says financial processes might need 90% or higher. The 20-document test shows whether your threshold actually catches your errors.

Sources

  1. AWS, Amazon Textract pricing, checked Nov. 14, 2026
  2. AWS, What is Amazon Textract?
  3. AWS, Amazon Textract best practices (confidence thresholds)
  4. AWS, Amazon Textract hard limits (handwriting languages)
  5. AWS, Using Amazon Augmented AI with Amazon Textract
  6. Microsoft Learn, Accuracy and confidence scores (Document Intelligence)
  7. Microsoft Learn, Document Intelligence model overview
  8. Microsoft Learn, Read model
  9. Microsoft Learn, OCR language support
  10. Microsoft, Azure Document Intelligence pricing page
  11. Microsoft, Azure retail prices API, Document Intelligence, East US, checked Nov. 14, 2026
  12. Google Cloud, Document AI overview
  13. Google Cloud, Document AI processor list
  14. Google Cloud, Document AI deprecations
  15. Barchard and Pace, Computers in Human Behavior, 2011 (abstract)

More from Peekgentic

Published November 14, 2026 · Written by Miguel Gutierrez, Co-Founder, Peekgentic (Dallas, TX). Every statistic links its source.