All three hyperscalers charge exactly $1.50 per 1,000 pages for basic OCR. They give three completely different answers to what happens to the document afterwards. Here is every retention, residency and model-training policy in one table, quoted from the vendors' own documentation.
Upload a document to extract
Drop files here or click to upload
Up to 50 files
Uploading...
Try it on a document you would actually send. Use a redacted sample if the real one is sensitive.
Google Document AI does not persist online requests to disk at all and states that it never uses customer data to train its models. Azure AI Document Intelligence stores the data and results temporarily in the same region and deletes them 24 hours after the analyze request, with an API to purge sooner. AWS may store customer content for service improvement, model training included, possibly in a region other than the one you called, unless you set an AWS Organizations opt-out policy that is switched off by default. Three vendors, one price, three different defaults.
This is the part of a document-AI evaluation that usually gets settled last and should be settled first. Teams compare accuracy on a sample set, compare the rate card, pick a winner, and only then send the contract to legal, who ask where the documents go. On a payroll file, a patient intake form or a set of signed loan documents, that question can end the project after the integration is already built. The rate card is public and easy. The retention behavior is spread across a developer guide, a responsible-AI page, an organization-level policy reference and a compliance list, and no two vendors organize it the same way.
So this page collects it in one place. Every cell below was read on the vendor's own documentation on July 22, 2026, and the verbatim sentences are quoted in the sections underneath rather than paraphrased. Where a vendor's product documentation does not answer a question, this page says that it does not, which is a more useful answer than a confident guess.
Read the last three rows together. The basic OCR rate is identical across all three, so data handling, not price, is the thing that actually differentiates them at the entry tier.
| Question | AWS Textract | Azure AI Document Intelligence | Google Document AI |
|---|---|---|---|
| Basic OCR rate per 1,000 pages | $1.50 | $1.50 | $1.50 |
| Documents kept after the response | Possibly, for service improvement | 24 hours | No (online), 1 day failsafe (batch) |
| Processed only in your region | Not guaranteed for improvement copies | Yes | Residency controls available |
| Used to train the vendor's models | Yes unless you opt out | Product doc is silent | No, stated explicitly |
| Opt-out on by default | No | Not applicable | Not applicable |
| Where the control lives | AWS Organizations policy | Per-request delete API | Built into the service |
| Explicit no-training statement on the product doc | No | No | Yes |
| Listed for HIPAA workloads | Yes, HIPAA eligible services list | Yes, under the Microsoft BAA | Yes, HIPAA covered |
One row deserves a caveat rather than a checkmark. "Product doc is silent" for Azure is not an accusation that Microsoft trains on your documents. It means the Document Intelligence data-privacy page and FAQ describe storage, region and deletion in detail and do not make a statement either way about model training, so the commitment you would rely on lives in your Azure agreement and the Microsoft Trust Center rather than in the product documentation. Google is the only one of the three that puts the sentence on the product's own security page, which is why it gets the only unqualified "no" in that row.
AWS is unusually direct about this, which is to its credit. The Organizations user guide states, word for word: "AWS AI services may use and store customer content for service improvement, such as fixing operational issues, evaluating service performance, debugging, or model training. For this purpose, we might store such content in an AWS Region outside of the AWS Region where you are using the service. You can opt out of use of your content for service improvement by using the AWS Organizations opt-out policy."
Read that twice, because two separate things are in it. The first is model training, named explicitly. The second is quieter and matters more for regulated work: content stored for that purpose may leave the region you deliberately chose. A team that picked us-east-1 for a residency reason has not, by picking it, confined every copy of the document to us-east-1.
The control is genuine and it works. The catch is where it lives. It is an AWS Organizations policy, which means it is set at the organization or organizational-unit level, not as a setting inside the Textract console. A company running one standalone AWS account, which describes a great many teams shipping their first document pipeline, has to enrol in Organizations before it can express the preference at all. Nothing in the Textract getting-started path prompts you to do it. So the honest description of the default is not that AWS hides anything. It is that the safe configuration requires an action in a different service that most builders never learn they needed.
There is a reward for setting it. AWS states that "When you opt out of content use by an AWS AI service, that service deletes all of the associated historical content that was shared with AWS before you set the option", limited to copies "not required to provide service functions". The opt-out is retroactive to the copies kept for improvement, not just forward-looking.
One frequently misquoted point is worth correcting here, because it appears in a lot of third-party write-ups. Textract's own data-protection page does contain a strong promise that customer content is "not logged or retained even for debugging purposes", but that sentence sits in the Custom Queries section and describes the images and annotations used to train an adapter, which AWS says are deleted once training completes. It is a promise about adapter training data. It is not a general statement about what happens to every document you pass to an ordinary analyze call, and quoting it as one is the most common error on this topic.
Google is the clearest illustration. For a synchronous request, Google states the document data "is processed in memory, encrypted in flight, and not persisted to disk". That is about as strong as a cloud retention answer gets. Switch the same workload to the batch endpoint, which is what most teams do the moment volume grows or files get long, and the behavior changes: the document is stored, "encrypted with an ephemeral key, meaning that no human has access to it", and "typically deleted immediately after the processing, with a failsafe Time to live (TTL) of one day".
Neither of those is a problem. The problem is a team that verified the synchronous answer during the pilot, wrote it into the security review, and then moved to batch for production without anyone revisiting the paperwork. The migration from sync to batch is driven by page limits and throughput, so it feels like an engineering decision. It quietly changes the data-handling answer that legal signed off on.
Azure has its own version of this. The standard cloud endpoint holds results for 24 hours; a connected container changes the picture substantially, since Microsoft states that "Connected containers don't send customer data, such as the image or text that's being analyzed, to Microsoft", while still requiring an outbound connection for billing. If your requirement is that documents never leave your network, that distinction between the meter calling out and the documents staying in is the entire decision, and it is covered in more depth on our page about on premise OCR software.
The practical rule: ask for the retention answer per endpoint, and re-ask it whenever the integration changes shape. Page limits, file size ceilings and throughput quotas differ per endpoint too, which we track on the OCR API limits comparison.
A growing number of teams skip the OCR products entirely and send page images to a general language model, because the per-page arithmetic can be attractive. That decision moves you onto a different set of data policies, written for a different product category and with different defaults. The general shape across the major language model APIs is that some retention exists by default for abuse monitoring and safety review, and that zero retention is an arrangement you request rather than a switch you flip, typically tied to an enterprise agreement.
Two things follow. First, scope: a zero data retention arrangement usually covers the inference endpoint, while request metadata, error traces and platform logs are governed separately. If the reason you need zero retention is a specific regulatory clause, the clause almost certainly cares about those too. Second, the batch endpoints that make language model OCR affordable are frequently governed by their own terms, so the same sync-versus-batch trap applies here with an extra layer.
Because this changes fast and varies by contract, we do not publish a per-vendor retention table for the language model providers on this page. The rates are a different matter and those we do track in detail: see LLM OCR pricing for the token-to-page conversion across every major model, and OCR pricing per 1,000 pages for the cross-vendor rate reference. If you are running a language model path in production, get the retention answer in your contract rather than from any documentation page, including this one.
Any vendor can answer these in a sentence each. How quickly they do it tells you most of what you need to know.
A number in hours, not "we do not retain your data longer than necessary". Azure says 24 hours and names the API that shortens it. That is what a usable answer looks like.
Ask separately about processing and about any copy kept for improvement, debugging or support. These can live in different places, and only one of them usually appears in the marketing material.
Not "we respect your privacy". The question is whether your documents can enter a training set, and whether preventing that is the default or a setting somebody has to find.
A control that exists but ships off protects nobody who did not read the documentation. This is the difference between the AWS position and the Google position, and it is the whole point of this page.
On Google these two endpoints have genuinely different retention behavior. Assume nothing carries across, and get both answers.
Documentation changes without notice. If retention is material to your deal, it belongs in the agreement or the DPA, where it is enforceable.
It would be poor form to publish this table and be vague about ourselves. DocuOCR is a cloud product. Documents you upload are processed and stored in US infrastructure so you can retrieve extractions, re-run them and export them, which is the point of having a workflow rather than a bare API, and you can delete them from the dashboard. We do not use customer documents to train models. If your requirement is that documents never leave your network at all, we are not the right fit and we say so plainly on on premise OCR software, which walks through what the on-premise options actually are and what they cost you in model freshness.
The honest summary of this whole page is that data handling has become a real differentiator between products that are otherwise priced identically, and it is the one dimension buyers routinely evaluate last. If you are choosing between the three hyperscalers on price alone, you are choosing between three identical numbers and ignoring the row that will actually matter to your security review. Start with the rate card if it helps, at OCR API pricing comparison, then come back to this table before anyone writes an integration.
Upload a document, look at the fields and the confidence scores, and see whether the output is worth the security review in the first place.
OCR API limits comparison
Page, size and rate ceilings per vendor
On premise OCR software
When documents cannot leave your network
OCR pricing per 1,000 pages
Every vendor, one rate table
High volume OCR API
What the rate really is above 1M pages
KYC document extraction
Identity documents and the retention question
OCR API pricing comparison
The buyer view across the market
From the same family of tools
