Summary

Document extraction depends on the file, the service and the processing route. Check completeness and critical fields before allowing the output to trigger an action.

What you'll learn

  • Match file and page limits to the actual processing mode.
  • Check handwriting language, embedded images and small text.
  • Test the required values, not just whether an upload succeeds.

Whether AI can read your documents depends on the file, the processing service and the task. A clean exported PDF, a faint scan and an invoice photographed at an angle need different checks. Start with the published format, page and text-height limits, then test the fields your business actually needs. An accepted upload is not proof that every amount, reference or handwritten note was captured correctly.

The floor nobody mentions: text has a minimum height

Textract’s October 2026 documentation gives a minimum detected text height of 15 pixels, approximately 8-point type at 150 DPI. Azure Read gives 12 pixels for a 1024×768 image and also describes approximately 8-point text at 150 DPI. These are service-specific input requirements, not a universal boundary for every AI model.[1][2]

Look at the smallest information your workflow needs: payment references, conditions, handwritten corrections and totals. Large body text can be readable while a critical reference remains unclear. Check the actual image rather than assuming that its file extension or scanner setting guarantees sufficient detail.

As a practical trial, rescan the original at 300 DPI instead of 150 DPI and compare the smallest required fields. More genuine image detail may help; enlarging an existing blurry image does not recreate missing information. Keep the clearer original and test the extraction result before changing the whole scanning process. AWS recommends a high-quality image, ideally at least 150 DPI, and advises against unnecessary conversion or downsampling.[8]

Exported documents may contain a text layer; scans and photographs need image reading. Check both the actual content and the processing route.
Check what is inside the file, then choose and test the extraction path.

What each provider actually accepts

Published limits differ by service and processing route. In October 2026, the compared paths include single-page processing, longer background jobs and model requests that also have context limits. Match the exact row to the service you intend to use.[1][2][4][5][7]

Service / routeFile or request sizePage limitImportant condition
Amazon Textract, PDF/TIFF[1]10 MB immediate /500 MB background1 immediate /3,000 backgroundNo vertical text or XFA PDFs
Azure Document Intelligence Read[2][3]4 MB free /500 MB paidFirst 2 pages free /2,000 paidImage dimensions 50×50 to 10,000×10,000 pixels
Claude PDF requests[4]32 MB request, platform-dependent600; 100 below a 1M-token context windowTotal request and context can bind earlier
OpenAI Responses file input[5]Each file under 50 MB; 50 MB combinedNo page figure quoted hereVision PDF processing includes text and page images
Mistral OCR[6]Confirm deployed service limitsNo page figure quoted hereTable formatting needs OCR 2512+; block extraction OCR 4+
Google Enterprise Document OCR[7]40 MB online /1 GB batch15 online /500 batch; 30 online in imageless modeOther Document AI processors have different limits

Textract’s immediate PDF/TIFF path accepts one page. Use the background route for multipage processing. Make the implementer show that your actual intake takes the correct route before judging a trial successful.[1]

Azure’s free tier processes only the first two pages. Check that the output covers every page you expected. A successful response is not a completeness check for the rest of a longer contract.[2]

Provider names do not identify the whole configuration. Record the service, model or processor, processing mode and account tier with your test. Keep file size and page count separate: a document can fit one limit while exceeding another. A dash or omitted page figure in a comparison must never be read as unlimited capacity.

Your PDF quietly becomes images, and that is where cost hides

Anthropic’s October 2026 guide gives an approximate 1,000-versus-7,000-token example for a three-page PDF on the older Amazon Bedrock Converse path, covering Opus 4.6 and earlier. The difference is between text extraction and visual PDF processing with citations enabled, not a universal multiplier for every PDF or service.[4]

The example illustrates why a text-only budget can miss image processing. Dividing the published approximate counts gives seven times as many tokens in that example. It does not predict your complete invoice: the document, model, processing mode and other request content matter.

OpenAI’s Responses documentation separately says that vision-capable PDF inputs include both extracted text and page images. Its non-PDF document path extracts text only. Ask which path the implementer is testing rather than assuming an uploaded file is processed in the same way everywhere.[5]

Dense PDFs can reach Claude’s context limit before its page limit. Split work along useful document boundaries and verify that the system still has the context needed for the question. If a total depends on a table continued on another page, the extraction needs that relationship preserved.[4]

Provider example for an older Claude Bedrock Converse path: about 1,000 text-only tokens versus 7,000 visual-processing tokens for three pages. Not a universal bill multiplier.
The approximate sevenfold token difference belongs to this provider example and processing path.[4]
A published input limit tells you what the service accepts. It does not prove that every required field was read correctly.

Language and handwriting are narrower than print

Textract’s October 2026 limits list printed-text detection in English, French, German, Italian, Portuguese and Spanish, but handwriting recognition only in English. A German handwritten note is outside that documented handwriting support, even when its printed parts are supported.[1]

Textract also excludes vertical text alignment. This is different from rotating a page within its plane: a sideways photograph of horizontal writing and characters arranged vertically are different inputs.[1]

Azure Read returns handwriting-style information with a confidence score. That score concerns the style classification, not proof that an amount or name was transcribed correctly. Review the actual extracted fields and the service’s supported handwriting languages before automating an action from them.[2]

Mistral describes multilingual OCR across more than 40 languages. It also makes structured extraction version-dependent: table formatting needs OCR 2512 or later and block extraction needs OCR 4 or later. Treat language coverage and layout features as separate questions; a listed language is not a measured accuracy promise for your handwriting.[6]

Textract supports printed detection in six listed languages, including German, while documented handwriting recognition is English-only.
For Textract, printed German and handwritten German fall into different support categories.[1]

Four traps that are not about quality at all

Check file restrictions and processing behaviour before interpreting a failed extraction as a model-quality problem. A valid response can still omit content that the selected route does not support.

Password restrictions. Textract excludes password-protected PDFs; Azure requires removing the lock; Claude requires standard unencrypted PDFs. Arrange authorised access to an unlocked working copy without weakening how sensitive originals are stored or shared.[1][2][4]

Embedded images in Office files. Azure Read does not extract embedded or linked images in the compared Office path. An invoice pasted as a picture into Word needs a different input route if that picture contains the required information. The surrounding typed text being returned does not prove the image was read.[2]

Billing units are format-specific. Azure counts up to 3,000 characters as a page unit for Word and HTML, while an Excel worksheet or PowerPoint slide is a unit. Keep that distinction when comparing page-based estimates with actual files.[2]

Rotation is not vertical writing. Textract supports in-plane rotation, including a 45-degree example, while excluding vertical text alignment. A supported rotation does not guarantee a readable image; blur, cropping and missing detail still need inspection.[1]

Before changing models, compare a clearer original scan with the current file. Keep the change only if the fields your workflow needs improve. A scanner setting is a testable intervention, not a guarantee of correct extraction.

Classify your own pile before you spend anything

Start with a representative set that includes the difficult files, then record what the workflow must extract from each. Do not accept the trial from a successful upload or a readable first page alone.

1. Identify the contents. Is the file exported, scanned or photographed? An exported PDF may contain real text, images or both. A scan may already have an OCR layer. Inspect the file and the intended service path rather than assuming its origin settles the answer.

2. Check the smallest required text. Compare its pixel height and clarity with the chosen service’s requirements. If a reference cannot be read, try a better original or a genuine higher-resolution rescan, then compare that field again.

3. Check the four traps. Record locks, Office images, handwriting language and text orientation. Have the implementer show how each exception is detected and routed instead of silently omitted.

4. Verify completeness and action accuracy. Count expected pages and returned pages, match the processing route and inspect the fields that trigger a payment or update. Use a person to approve uncertain results. Record the model, service, settings and test outcome so the check can be repeated after a change.

When the documents will not work

If the required fields remain unreadable, change the input, the service path or the scope of automation before allowing the output to trigger business actions.

Improve the source. Try a clearer original or a genuine higher-resolution scan. Check the required fields again; do not assume a resolution change fixes handwriting, missing pages or a damaged original.

Move the intake point. Ask whether the supplier can provide a native export or structured file. That can remove an unnecessary reading step, but confirm that the new file contains the same required information and reaches the right workflow.

Automate the demonstrated part. Keep readable document classes on the tested route and send exceptions to people. A partial workflow is useful when the system reliably identifies which files and fields it cannot handle. Measure that routing too: a file that slips through with a wrong payment reference is different from one held for review.

What the extraction limits do not cover

Published input limits are not accuracy measurements for your document collection. Passing a size, page or language check is the start of a test, not the acceptance result.

Keep the comparison tied to its named service and processing mode. Google’s processor-specific limits, for example, should not be reduced to one maximum for every Document AI task.[7]

For acceptance, compare extracted values with the original, include difficult files and record omissions as well as incorrect text. Judge the effect on the business action, not merely whether the generated summary reads well. Repeat relevant cases after model or processing changes.

Frequently asked questions

Can AI read a scanned document? Often, but test the actual file and service. Textract documents 15 pixels as its minimum text height, approximately 8-point at 150 DPI; Azure gives its own 12-pixel requirement. Meeting an input requirement does not prove that every required field was extracted correctly.[1][2]

Can ChatGPT read a PDF? OpenAI’s API documentation describes PDF input for vision-capable models as extracted text plus page images. That is an API processing path, not a complete statement of ChatGPT’s product upload limits. Test the actual product or connection and inspect the result.[5]

Why does AI extract the wrong data from my invoices? Check image detail, supported layout, completeness and the selected processing path. For example, Azure Read’s Office route does not extract embedded images, so a picture inside Word may be missed. A successful response is not a check of the invoice’s totals or payment reference.[2]

Can AI read handwriting? Support depends on language and service. Textract documents handwriting only in English, while its printed-text list includes German. Azure’s handwriting-style confidence does not prove transcription accuracy. Test the actual note and validate fields before acting on them.[1][2]

How many pages can AI process at once? Textract’s immediate PDF/TIFF path accepts 1 page and its background path up to 3,000. Azure Read processes the first 2 pages on the free tier and up to 2,000 on paid. Claude documents 600 pages, reduced to 100 when context is below 1M tokens, with other request/context limits still applying.[1][2][4]

Notes

  1. Set Quotas in Amazon Textract. October 2026 documentation Accessed October 15, 2026.
  2. Read model OCR data extraction. October 2026 documentation Accessed October 15, 2026.
  3. Service quotas and limits. October 2026 documentation Accessed October 15, 2026.
  4. PDF support. October 2026 documentation Accessed October 15, 2026.
  5. File inputs. October 2026 documentation Accessed October 15, 2026.
  6. Document AI OCR processor. October 2026 documentation Accessed October 15, 2026.
  7. Document AI limits. October 2026 documentation Accessed October 15, 2026.
  8. Best Practices. October 2026 documentation Accessed October 15, 2026.