What Is Amazon Bedrock Data Automation?

Last updated July 2026.

Amazon Bedrock Data Automation is an AWS service that takes unstructured content, documents, images, audio or video, and returns structured data from it in a single API call. For documents it does two things: standard output gives you text, tables and a summary in a fixed shape for $0.010 per page, and custom output returns a schema you defined yourself in a blueprint for $0.040 per page. That works out to $10.00 and $40.00 per 1,000 pages. It is not a cheaper replacement for Amazon Textract. For plain text extraction it costs about 6.7 times more.

Most explanations of Bedrock Data Automation get stuck describing what AWS says it is, which is a way to "generate insights from multimodal content". That is marketing language and it tells you nothing about whether to use it. What follows is what the service actually does, what it actually bills you, and the specific jobs where it is the right call.

What Bedrock Data Automation actually does

Think of it as the layer that sits between a raw file and a database row. You hand it a scanned purchase order. Instead of getting back a wall of text with coordinates, which is what a traditional OCR engine gives you, you get back a JSON object with the fields you asked for: the PO number, the vendor, the line items, the total.

The mechanism is a thing AWS calls a blueprint. A blueprint is a description, in plain language, of the fields you want. You write something like "invoice_number: the unique identifier the supplier assigned to this invoice" and the model finds it. You are not drawing boxes on a template or writing coordinate rules. That distinction is the whole point of the service and it is what you are paying the premium for.

It covers four modalities through one API:

• Documents. PDF, TIFF, JPEG, PNG and DOCX. Printed and handwritten text, at any in-plane rotation.
• Images. JPEG and PNG, up to 5 MB.
• Audio. Transcription and speaker analysis, up to four hours per file.
• Video. Scene detection, transcript and description, up to four hours.

Textract, by comparison, does documents and nothing else.

Standard output vs custom output: what is the difference?

This is the choice that decides your bill, and the naming does not help you.

Standard output is what BDA returns by default. Text, a summary, tables, figure captions, in a shape AWS decided. You do not control the structure. It costs $0.010 per page.

Custom output means you supply a blueprint and BDA returns exactly the schema you named. It costs $0.040 per page.

The important thing to internalize is that the four-times price difference is not buying you better OCR. The text recognition is the same. You are paying for the schema, for the service doing the mapping work that you would otherwise write yourself. That reframing makes the pricing decision much easier: if you do not need a specific schema, do not pay for custom output, and honestly, if all you need is text, do not use this service at all.

How much does Amazon Bedrock Data Automation cost?

Here is every published meter, normalized onto the per-1,000-pages unit the rest of the industry quotes.

Two things about this table are worth more than the numbers in it.

First, there is no volume discount at any scale. Textract drops to $0.60 per 1,000 pages above a million. Azure sells commitment tiers. Google discounts above five million pages. Bedrock Data Automation charges the same rate on your ten millionth page as on your first. That inverts the usual assumption that the hyperscaler gets cheaper as you grow.

Second, the surcharge above 30 blueprint fields is real and nobody talks about it. A blueprint can hold up to 100 leaf fields. Take it to the maximum and you pay $0.040 plus 70 extra fields at $0.0005, which is $0.075 per page, or $75.00 per 1,000 pages. AWS is the only one of the three big cloud extraction services that charges you more for asking for more fields. Azure and Google do not care how wide your schema is.

The full meter-by-meter breakdown, including the worked field math and how the rates compare to Azure and Google, is on our Amazon Bedrock Data Automation pricing reference.

Is Bedrock Data Automation better than Amazon Textract?

It is better at a narrow, specific thing, and worse at almost everything else. That sounds harsh but it follows straight from the rate card.

All rates per 1,000 pages. So BDA wins exactly one column: the full structured form where you would otherwise have chained Forms, Tables and Queries together for $70.00 and then written code to stitch three different output shapes into one record. That is a real and common problem, and $40.00 in one call is genuinely a better deal than $70.00 in three.

Everywhere else, Textract is cheaper, often dramatically. We go through this decision in detail, including the architecture that uses both services together, on Bedrock Data Automation vs Textract.

Does Bedrock Data Automation replace Textract?

No. Both services are live, both have active billing meters, and AWS has announced no end of support for Textract. This matters because the assumption that a newer generative service must supersede the older API is doing real damage to people's cloud bills. If you are currently calling Detect Document Text at $1.50 per 1,000 pages and you "upgrade" to standard output at $10.00, you have multiplied that line of your bill by nearly seven and bought yourself nothing.

The pattern that works is not migration. It is routing. Run Textract cheaply across everything to get text and figure out what each document is, then call Bedrock Data Automation only on the subset that genuinely needs a structured record.

What are the limits of Bedrock Data Automation?

The page limit is the one people get wrong, because there are three different numbers and the internet repeats the wrong one.

• Synchronous API: 10 pages, 50 MB. This is the real constraint if you need an answer in the request.
• Asynchronous, from the console: 20 pages. This is the figure everyone quotes as "the limit". It is not.
• Asynchronous, splitter enabled: 3,000 pages, 500 MB. This is actually the highest single-document ceiling of any major service. Azure Document Intelligence stops at 2,000 and Google's online endpoint stops at 15.

Two other limits deserve more attention than they get. Document support covers six languages only: English, German, Spanish, French, Italian and Portuguese. There is no Chinese, Japanese or Korean, and AWS states plainly that vertical text alignment does not work. For US paperwork that is a non-issue. For anyone with CJK documents it is a hard gate, and no pricing guide anywhere mentions it.

And DOCX files lose their page numbers. Word documents are converted to PDF before processing, and AWS notes that page number mapping does not survive that conversion. If you need to trace an extracted value back to a source page in a Word file, that link is gone.

Who should actually use it?

Use Bedrock Data Automation when your documents vary in layout and you can describe what you want as a list of named fields. A blueprint is a description of intent, not a template, so when a supplier redesigns their invoice, your extraction keeps working. That resilience is worth real money to anyone who has maintained a template-based extraction pipeline and watched it break every quarter.

Use it when the alternative is three Textract calls. Use it when you have audio or video alongside documents and you want one API rather than four.

Do not use it to read text off a page. Do not use it for invoices and receipts, where Analyze Expense is a purpose-built model at a quarter of the price. Do not use it at high volume without modelling the bill first, because the missing volume discount means the gap to every competitor widens as you scale rather than closing.

And be clear about what any of these APIs actually hand you, which is a JSON response and nothing else. There is no classifier deciding which documents need which treatment, no confidence threshold routing the doubtful ones to a person, no validation checking that the line items sum to the total, no export into your accounting system. Extraction is one step. If the output is headed for a finance workflow, you will also want a way to turn that extracted ledger data into board-ready financial statements, and that is a different tool again. On most projects the pipeline around the API costs more to build than the API costs to run, which is the single most useful thing to know before you start comparing per-page rates.

The short version

Bedrock Data Automation is a schema service wearing an OCR service's clothes. Pay it for structure. Never pay it for text. If you are choosing between it and Textract, the honest answer for most teams is that you want both, with Textract underneath doing the cheap work and BDA called only where a structured record actually earns its $40.00 per 1,000 pages.

Related reading: Bedrock Data Automation pricing, AWS Textract pricing, OCR pricing per 1,000 pages across every vendor, and the OCR API pricing comparison pillar.