Amazon Textract Guide: Extracting Data From Documents on AWS

Amazon Textract extracts text, tables and forms from documents. Learn its APIs, pricing, use cases and production best practices.

What is Amazon Textract?

Every organization processes documents: invoices, receipts, contracts, tax forms, medical records, loan applications and identity documents. Pulling structured data out of them used to mean manual data entry or brittle OCR tools that needed reconfiguring whenever a form layout changed. Amazon Textract, however, replaces both with ML-powered document processing.

Amazon Textract is a fully managed machine learning service from Amazon Web Services that extracts text, handwriting, layout elements and structured data from scanned documents. Traditional OCR only reads characters off a page, while Amazon Textract also understands document structure: it finds tables, forms with key-value pairs, signatures and the relationships between parts of a document.

Give Amazon Textract an invoice, for example, and it reads the text and also works out that “Invoice Number” is a label and “INV-2024-001” is its value. Similarly, it recognizes that a table’s header row holds “Item”, “Quantity” and “Price”, and that the rows below are data belonging to those columns. That understanding of context is what separates intelligent document processing from basic character recognition, and it makes Textract the base for automated document workflows on AWS.

As a result, the efficiency gains can be large. Teams that process documents by hand, keying invoice data into ERP systems, copying medical forms into electronic records or reviewing loan applications page by page, spend a lot of time per document and introduce errors along the way. Textract can process a page in seconds, with a consistency that manual entry struggles to match when thousands of documents arrive each day.

Textract Abilities at a Glance

6 languages
Supported for Text Detection
4
Input Formats: PDF, PNG, JPEG, TIFF
5
Specialized APIs: Text, Analyze, Expense, ID, Lending

Amazon Textract accepts PDF, PNG, JPEG and TIFF files. It also handles single-page and multi-page documents, reads printed and handwritten text and returns every extracted element with a confidence score and bounding box coordinates.

Amazon Textract processing cycle: classify the document, send it to the right API, check confidence scores, validate against business rules, review low-confidence results
A production cycle for Amazon Textract document processing.

Key Takeaway Amazon Textract goes well beyond basic OCR. It understands document structure, including tables, forms, key-value pairs, signatures and layout, and returns structured, machine-readable data from most document types. If your organization processes documents by hand, Textract is a quick route to automation on AWS.


How Amazon Textract Works

Amazon Textract runs as a serverless API. You send a document, stored in S3 or as raw bytes, choose the type of analysis and get structured JSON back, with every detected element’s text, confidence score and position on the page.

Textract’s pre-trained models work on most documents without templates or configuration. In addition, AWS keeps improving the models, so accuracy can improve over time without any work on your side.

Textract API Overview

Amazon Textract provides several specialized APIs, each built for a different document task:

  • DetectDocumentText: The simplest API, which extracts all text as words and lines. It is plain OCR, but deep learning gives it better accuracy on hard inputs such as handwriting, low-quality scans and noisy backgrounds.
  • AnalyzeDocument: The core API. It extracts text plus structure: tables (rows, columns and cells), forms (key-value pairs and checkboxes), signatures and layout elements such as paragraphs, titles, headers, footers and lists. This is where Textract’s grasp of structure sets it apart from basic OCR.
  • Queries: A feature of AnalyzeDocument that lets you ask natural language questions about a document, such as “What is the patient name?” or “What is the due date?”, and get precise answers. In addition, it is pre-trained on paystubs, bank statements, W-2s, loan applications, mortgage notes and insurance cards.
  • AnalyzeExpense: Built for invoices and receipts. It identifies vendor names, even from logos without labels, along with line items, quantities, prices and totals, and returns normalized field names so different invoice formats can be processed the same way.
  • AnalyzeID: Built for identity documents. It extracts data from U.S. passports and driver’s licenses without templates, which supports automated identity verification, account opening and KYC workflows.
  • Analyze Lending: A managed workflow for mortgage loan packages. It classifies pages into document types, such as W-2s, paystubs, bank statements and tax returns, sends each to the right extraction API and returns consolidated results for the whole package.

Synchronous vs Asynchronous Textract Processing

Amazon Textract offers two processing modes. Synchronous APIs (DetectDocumentText, AnalyzeDocument) process single-page documents in real time and return results immediately, which suits interactive apps where a user uploads a document and expects instant feedback. Asynchronous APIs (StartDocumentTextDetection, StartDocumentAnalysis), by contrast, handle multi-page documents as jobs and notify you through SNS when results are ready, which suits batch pipelines and large document workflows.

For production, the most common design combines S3 event notifications with Lambda. When a document lands in a bucket, Lambda starts the Textract analysis, processes the results and stores the extracted data in DynamoDB, RDS or Amazon OpenSearch for other applications. This event-driven pattern scales on its own and needs no infrastructure management.

For high-volume processing, AWS also publishes Intelligent Document Processing (IDP) CDK constructs: templates that deploy a full pipeline with S3 ingestion, SQS queuing to manage throttling, Lambda orchestration, Textract analysis and result storage. They build in exponential backoff with jitter, error handling, dead-letter queues for failed documents and CloudWatch dashboards, which saves weeks of pipeline work.

Textract also worked with Amazon Augmented AI (A2I) for human review: when a confidence score fell below your threshold, A2I sent the document to a reviewer, and the corrected data flowed back into the pipeline. A2I moved to maintenance mode and closed to new customers on July 30, 2026, so new projects should build their own review step, for example a queue of low-confidence results in a simple internal review tool. Human review still matters wherever extraction errors carry business or compliance risk.

Custom Queries for Textract

You can also customize the pre-trained Queries feature with your own documents. In the AWS Console, you upload a small set of sample documents, annotate the fields you need and train an adapter. This helps with industry-specific documents where pre-trained models may miss specialized fields, such as GST numbers on Indian invoices or policy numbers on your own insurance documents.

Custom Queries keep your data private during training. Your training documents and annotations stay in your AWS account, and the resulting adapter is private to your organization. It works alongside Textract’s pre-trained features, so you can use standard and custom Queries in the same API call.


Core Amazon Textract Features

Beyond the APIs above, several features make Amazon Textract suited to enterprise document processing. Together they handle most document types, from simple one-page forms to long financial reports with nested tables and a mix of handwritten and printed content:

Table Extraction
Keeps table structure during extraction, so rows, columns, cells and headers stay related. As a result, extracted tables can go straight into databases, spreadsheets or data pipelines with their structure intact.
Form Key-Value Pair Extraction
Finds label-value pairs in forms, such as “First Name: Jane”, and selection elements such as checkboxes. It tells keys from values without templates, so you can capture structured data from any form layout.
Handwriting Recognition
Detects and extracts handwritten text alongside printed text. It handles mixed documents where some fields are typed and others handwritten, which is common in healthcare intake forms, applications and field reports.
Signature Detection
Detects whether and where a document is signed, which helps when validating contracts, checks, consent forms and loan agreements. It returns bounding box coordinates and confidence scores.
Layout Analysis
Extracts layout elements: paragraphs, titles, section headers, footers, page numbers and lists. It keeps reading order and hierarchy, so the content works well for downstream NLP.
Natural Language Queries
Ask questions about a document in plain English, such as “What is the total amount due?”, and get precise answers. It is pre-trained on common business documents, and Custom Queries cover domain-specific fields.
Talk to an ExpertStill keying documents in by hand? Plan your Textract pipeline with us.

Amazon Textract Pricing Model

Fundamentally, Amazon Textract uses pay-per-page pricing with no minimum commitments. Rather than listing specific dollar amounts that change over time, here is how the cost structure works:

Understanding Textract Cost Dimensions

  • Pages processed: Charged per page, with a separate rate for each API. DetectDocumentText (plain OCR) is the cheapest, AnalyzeDocument costs more because of its structural analysis, and AnalyzeExpense and AnalyzeID have their own per-page rates.
  • Feature combinations: In AnalyzeDocument, you can turn on several features per call, such as tables, forms, queries, signatures and layout. Each one adds to the per-page cost, so enable only what each document type needs.
  • Volume tiers: Per-page prices fall as monthly volume rises, so high-volume workflows benefit from the tiered pricing.
  • Custom Queries: Documents processed with a custom adapter are charged at their own per-page rate.
Cost Optimization Strategy

Match the API to the task. If you only need raw text, use DetectDocumentText and skip AnalyzeDocument’s structural analysis. For invoices, use AnalyzeExpense, which returns normalized fields, instead of generic AnalyzeDocument. And classify documents first, so each goes to the cheapest API that can handle it. For current prices by API and tier, see the official Textract pricing page.


Amazon Textract Security and Compliance

Textract processes sensitive business documents, such as financial records, identity documents, medical forms and legal contracts, so security matters.

Textract Data Protection

Data processed by Amazon Textract is encrypted in transit with TLS and at rest with AWS KMS. Textract supports interface VPC endpoints through AWS PrivateLink, so document traffic can stay off the public internet, and IAM policies control which users and applications can call the APIs. AWS may store and use content processed by Textract to improve the service, unless you opt out through an AWS Organizations AI services opt-out policy.

For data residency, Textract processes documents in the AWS Region where you call the API. Combined with S3 bucket policies, KMS keys and IAM controls, that keeps sensitive document data inside the security boundary you define.

Amazon Textract is HIPAA eligible, so healthcare organizations can use it for medical records, insurance claims and patient intake forms that contain protected health information. It is also covered by AWS compliance programs such as SOC, PCI DSS and ISO 27001, which helps financial services firms processing loan documents, invoices and tax forms meet their compliance obligations.


What’s New in Amazon Textract

AWS keeps updating Amazon Textract. The Layout feature, for example, extracts paragraphs, titles, headers, footers and lists while keeping the reading order and hierarchy of complex documents. That helps downstream NLP and content management, where structure and reading order matter as much as the raw text.

Custom Queries adapters let organizations tune extraction with a small set of annotated samples, so domain-specific processing does not need ML expertise. Together with the pre-trained Queries, which cover paystubs, bank statements, W-2s, loan application forms, mortgage notes, claims documents and insurance cards, they let one API handle standard and proprietary formats.

Textract has also improved handwriting recognition and support for pages that mix printed and handwritten text. In healthcare, insurance and government, where handwritten notes on printed forms are common, this reduces manual review and correction.


Real-World Amazon Textract Use Cases

Amazon Textract serves organizations in every industry that handles paper or digital documents, from banks processing loan applications to healthcare organizations digitizing decades of patient records. These are the use cases we implement most often for enterprise clients:

Invoice and Receipt Processing
Extract vendor names, line items, quantities, prices, totals and dates from invoices and receipts with AnalyzeExpense. Field names are normalized across vendor formats for consistent processing in accounts payable systems.
Mortgage and Loan Processing
Use Analyze Lending to classify and extract data from whole loan packages, including W-2s, paystubs, bank statements and tax returns, which shortens mortgage processing a lot.
Healthcare Document Automation
Extract patient data from intake forms, insurance claims, pre-authorization forms and medical records. HIPAA eligibility supports compliant processing of protected health information.
Identity Verification
Extract data from passports and driver’s licenses with AnalyzeID for automated KYC checks, account opening and onboarding, which cuts manual identity verification time.
Contract Analysis
Extract key terms, dates, parties and clauses from contracts and legal documents. Add Amazon Comprehend for entity extraction and analysis of contract language.
Searchable Document Archives
Turn scanned paper archives into searchable digital libraries. Extract text from historical documents and index it in Amazon OpenSearch for full-text search across the whole repository.

Amazon Textract vs Azure AI Document Intelligence

If you are comparing document processing services across cloud providers, here is how Amazon Textract compares with Microsoft’s Azure AI Document Intelligence (formerly Form Recognizer):

Ability Amazon Textract Azure AI Document Intelligence
Core OCR Yes: DetectDocumentText with deep learning Yes: Read API with advanced OCR
Table Extraction ✓ Preserves full table structure Yes: Table extraction with cell merging
Form Key-Value Pairs ✓ Automatic without templates Yes: Pre-built and custom models
Natural Language Queries ✓ Ask questions in plain English ◐ Field extraction with labels
Invoice Processing Yes: AnalyzeExpense with normalization ✓ Pre-built invoice model
ID Document Processing Yes: AnalyzeID (U.S. documents) ✓ Broader international ID support
Lending / Mortgage ✓ Analyze Lending managed workflow ◐ Prebuilt mortgage document models
Custom Model Training Yes: Custom Queries adapters Yes: Custom models with labeling
Handwriting Recognition Yes: Mixed print and handwriting Yes: Mixed print and handwriting
Language Support ◐ 6 languages (EN, ES, DE, FR, IT, PT) ✓ Many more languages for print OCR

Choosing the Right Textract Alternative

Both services are mature, so your cloud ecosystem usually decides. On AWS, Textract’s native links to S3, Lambda and SQS make it the natural choice for automated document pipelines. On Azure, Document Intelligence integrates natively with Azure Blob Storage, Logic Apps and Azure AI services.

Azure offers prebuilt models for common U.S. mortgage documents, while Textract’s Analyze Lending adds a managed workflow that classifies and extracts a whole loan package in one go, which matters for lenders automating loan origination. Textract’s natural language Queries also make ad hoc extraction easier than label-based field extraction.

Azure has clear advantages in two areas. First, languages: Azure’s OCR covers many more languages than Textract’s six, which matters for global organizations with multilingual documents. Second, identity documents: Azure recognizes IDs from many countries, while Textract’s AnalyzeID covers U.S. passports and driver’s licenses.

AWS teams that need more languages can use a hybrid approach: Textract for documents in its supported languages, where its structural understanding is strongest, and Amazon Bedrock with multimodal models for other languages.


Getting Started with Amazon Textract

Amazon Textract needs no setup: the pre-built APIs have no models to deploy or train. You call the API with a document and get structured results right away. The free tier gives enough pages to test your use case before you commit to production spending.

Your First Textract API Call

Here is a minimal Python example that extracts text from a document stored in S3:

import boto3

# Initialize the Textract client
client = boto3.client('textract', region_name='us-east-1')

# Extract text from a document in S3
response = client.detect_document_text(
    Document={
        'S3Object': {
            'Bucket': 'my-documents',
            'Name': 'invoices/invoice-001.pdf'
        }
    }
)

# Print extracted lines of text
for block in response['Blocks']:
    if block['BlockType'] == 'LINE':
        print(f"{block['Text']} ({block['Confidence']:.1f}%)")

For structured extraction of tables, forms and queries, replace detect_document_text with analyze_document and specify the feature types you need. For invoices, use analyze_expense, and for identity documents, analyze_id. Each API returns structured JSON with confidence scores and bounding boxes for every element. For more details and patterns, see the Amazon Textract documentation.


Amazon Textract Best Practices and Pitfalls

Advantages
Goes beyond OCR: understands tables, forms, key-value pairs and layout
Natural language Queries let you extract specific fields without templates
Purpose-built APIs for invoices (AnalyzeExpense), IDs, and lending workflows
Fully serverless with automatic scaling and no infrastructure to manage
Custom Queries adapters trainable on a small set of samples
HIPAA eligible, SOC compliant, PCI DSS compliant
Limitations
Text detection in 6 languages, so less suited to multilingual work
Complex table structures can produce inconsistent extraction results
A2I human review closed to new customers since July 30, 2026
Cloud-only, with no offline option for air-gapped environments
AnalyzeID currently limited to U.S. government-issued documents

Recommendations for Textract Deployment

  • Classify documents before processing: Send each document type to the most suitable API: invoices to AnalyzeExpense, IDs to AnalyzeID and general forms to AnalyzeDocument. That improves accuracy and lowers cost, because each API is tuned for its document type.
  • Set confidence thresholds for your use case: Textract returns a confidence score for every element. For high-stakes work such as financial processing and compliance, send low-confidence extractions, for example below 95%, to human review instead of processing them automatically. For lower-stakes work such as search indexing, a lower threshold, say 80%, may be enough.
  • Add retry logic with exponential backoff: Textract enforces per-account rate limits in transactions per second. Use exponential backoff with jitter to handle throttling, especially in large batch runs. The AWS IDP CDK constructs include these patterns.

Architecture and Validation Best Practices

  • Use asynchronous APIs for multi-page documents: Synchronous APIs handle single-page documents only. For multi-page PDFs, use the asynchronous Start and Get pattern with SNS notifications, so your application is not blocked. Queue documents in SQS to smooth traffic and stay within rate limits.
  • Validate extracted data against business rules: ML extraction is accurate but not perfect. Check that dates are valid, totals match line items, required fields are present and number formats are consistent, so errors are caught before they reach ERP, CRM or compliance systems.

Key Takeaway Amazon Textract turns manual document processing into automated, scalable pipelines, extracting structured data from invoices, forms, IDs and loan packages through purpose-built APIs. Successful deployments match the right API to each document type, set sensible confidence thresholds and validate extracted data against business rules. An experienced AWS partner can help you design document processing architectures that raise accuracy while keeping costs down.

Our SolutionAI-Powered Document Processing and Workflow Automation

Frequently Asked Questions About Amazon Textract

Common Questions Answered
What is Amazon Textract used for?
Amazon Textract extracts structured data from documents such as invoices, receipts, forms, contracts, tax documents, identity documents, medical records and loan packages. It goes beyond OCR by understanding structure: tables, key-value pairs, checkboxes, signatures and layout. Common uses include invoice processing in accounts payable, mortgage processing, healthcare document digitization, identity verification for KYC, contract analysis and searchable archives of paper records.
How is Amazon Textract different from OCR?
Traditional OCR reads characters and returns raw text without understanding the document’s structure. Amazon Textract uses deep learning to understand tables, forms, key-value pairs, signatures and layout. For example, Textract knows that “Due Date” is a label and “March 15, 2026” is its value, while OCR returns them as unrelated strings. Textract can also answer natural language questions about documents and normalize invoice data across vendor formats.
What languages does Amazon Textract support?
Amazon Textract detects text in English, Spanish, German, French, Italian and Portuguese. If you need more languages, Azure AI Document Intelligence covers many more for printed text, and Amazon Bedrock multimodal models can read documents in other languages. Amazon Rekognition’s DetectText API also reads text in images, although it lacks Textract’s understanding of document structure.

Technical and Integration Questions

Does Amazon Textract store my documents?
Textract returns its results to you and does not keep a copy you manage, but AWS may store and use content processed by Textract to improve the service unless you opt out through an AWS Organizations AI services opt-out policy. Processing is encrypted in transit and at rest, and VPC endpoints keep document traffic inside your private network.
What is the difference between Amazon Textract and Amazon Rekognition for text?
Amazon Rekognition’s DetectText API is built for text in natural scenes, such as street signs, product labels and graphics in images and videos. Amazon Textract is built for business documents, such as invoices, forms and contracts. So use Rekognition to find text in photos or video frames, and Textract to extract structured data from documents.

References

  1. APIs, Supported Languages and Formats, Queries and Adapters: AWS — Amazon Textract Developer Guide
  2. Per-Page Pricing by API and Tier: AWS — Amazon Textract Pricing
  3. SageMaker A2I Moving to Maintenance (Closed to New Customers July 30, 2026): AWS What’s New — AWS Service Availability Updates, June 2026
Newsletter

Security Insights, Delivered.

Practical, vendor-neutral analysis of threats, tools, and architecture decisions.

Book a Consultation

A practice principal will contact you within 1 business day.

Thank you! A Signisys principal will be in touch within 1 business day.