Raising Document Extraction Accuracy to 92– 95%: How an AI Software Company Modernised Its Document Platform with Amazon Textract and Bedrock
AI & Software Development

Raising Document Extraction Accuracy to 92– 95%: How an AI Software Company Modernised Its Document Platform with Amazon Textract and Bedrock

Back to case studies
Confidential
AI & Software Development

Company Overview

The customer is an India-based AI and software development company that offers a document intelligence platform used to extract and structure information from business documents at scale. As document volumes and complexity grew, the platform's existing extraction approach began to limit both accuracy and cost efficiency, prompting the Customer to look for a Generative AI upgrade that could handle a wider range of document types while integrating cleanly into its existing AWS environment.

Customer Challenge

1. Low Extraction Accuracy on Complex Documents

• The existing pipeline relied on rule-based OCR and struggled with complex and unstructured documents.

• Field extraction accuracy sat at only 70–75%, falling short of what the platform needed.

 

2. Heavy Manual Review Burden

• Roughly 35% of all documents processed required manual human review.

• Each review took 15 to 20 minutes, creating a significant operational bottleneck.

 

3. Inefficient, Always-On Infrastructure

• The EC2-based processing infrastructure ran continuously regardless of document volume.

• This produced idle compute costs and poor overall cost efficiency.

 

4. Limited Document Coverage

• The system handled only pre-defined templates and could not reliably process unstructured or semi-structured documents.

 

5. Need for Clean AWS Integration

• Any upgrade had to integrate into the Customer's existing AWS environment rather than require a separate platform.

Solution Approach

CloudTry integrated Amazon Textract and Amazon Bedrock (Claude Haiku) into the platform's document intelligence pipeline on a fully serverless AWS architecture, combining high-accuracy extraction with contextual reasoning over the results.

1. Generative AI Document Intelligence (Amazon Textract + Amazon Bedrock – Claude Haiku)

• Amazon Textract performs high-accuracy OCR, form field extraction, and table detection across complex document types.

• Amazon Bedrock (Claude Haiku) then reasons over the raw Textract output via a structured prompt, classifying the document type, validating extracted fields, filling in missing data through contextual reasoning, and returning a clean structured JSON output.

 

2. Event-Driven Serverless Pipeline

• Documents uploaded to the platform are stored in Amazon S3, which triggers an AWS Lambda function via S3 Event Notification.

• Lambda calls Textract, passes the output to Bedrock, and stores the structured results in Amazon DynamoDB with a full audit trail per document.

• The application retrieves results through Amazon API Gateway, backed by Lambda.

 

3. Scalable Serverless AWS Architecture

• Amazon S3 — stores incoming documents and triggers processing via event notifications.

• AWS Lambda — orchestrates the pipeline on a pay-per-invocation basis with no idle cost.

• Amazon Textract — provides OCR, form field extraction, and table detection.

• Amazon Bedrock (Claude Haiku) — classifies, validates, completes, and structures extracted data into JSON.

• Amazon DynamoDB — stores results and a full audit trail per document on on-demand pricing.

• Amazon API Gateway — serves results to the application layer through a managed endpoint.

• Amazon SES — sends processing completion notifications and manual review alerts.

• Amazon CloudWatch — monitors the full pipeline with dashboards tracking error rates, processing times, and Bedrock invocation metrics.

 

4. Structured Prompt Design for Reliable Output

• Prompt engineering, refined during the proof-of-concept phase, drives document classification and field validation.

• The structured approach produced measurably more consistent JSON output than a generic prompt.

 

5. Compliance-Ready Audit Trail

• A full per-document audit trail is maintained in DynamoDB.

• The schema was shaped by the Customer's compliance requirements and agreed before the build.

 

6. Fully Serverless Operations (No EC2)

• The entire pipeline is serverless with zero EC2 infrastructure, scaling automatically with document volume on a pay-per-document model.

 

The solution runs in a live AWS production environment integrated with the Customer's platform, with no EC2 instances to manage.

Outcomes & Impact

Transformation Overview (Before vs After)

Aspect 

Before Generative AI 

After Generative AI

Extraction Accuracy 

70–75% (rule-based OCR) 

92–95% (Textract + Bedrock)

Processing Time 

15–20 minutes (manual

review) 

Under 30 seconds (automated)

Manual Review Rate 

~35% of documents 

Under 8% of documents

Document Coverage 

Pre-defined templates only 

Unstructured and semi-structured, all formats

Compute Model 

Always-on EC2 

Serverless, pay-per-document

Infrastructure Cost 

High and fixed 

An estimated 35% lower, scales with volume

 

 

Results Achieved

 

1. Extraction Accuracy Improved to 92–95%

• Field extraction accuracy rose from 70–75% with rule-based OCR to 92–95% with the Textract and Bedrock pipeline.

 

2. Processing Time Cut to Under 30 Seconds

• Per-document processing dropped from 15–20 minutes of manual review to under 30 seconds through the automated AI pipeline.

 

3. Manual Review Rate Reduced to Under 8%

• The share of documents requiring manual review fell from approximately 35% to under 8%.

 

4. Document Coverage Expanded

• The platform moved from handling pre-defined templates only to processing unstructured and semi-structured documents across all formats.

 

5. Infrastructure Cost Reduced by an Estimated 35%

• Replacing always-on EC2 with serverless Lambda on a pay-per-document model removed idle compute spend and lowered infrastructure cost by an estimated 35%.

 

6. Results Validated Across 500 Documents

• Accuracy was benchmarked across a diverse 500-document test set, confirming the pre and post-deployment improvement before go-live.

 

7. High Customer Satisfaction

• The engagement closed with a customer satisfaction score of 4.25 out of 5.

 

Key Differentiators & Business Impact

This is a production-grade Generative AI implementation that pairs two AWS services to do what neither does as well alone. Amazon Textract provides structured extraction, and Amazon Bedrock reasons over that output to classify documents, validate fields, and fill gaps — compensating for the cases where OCR alone falls short on complex or poorly formatted documents. The combination is what lifted accuracy from the low seventies into the low-to-mid nineties.

The architecture addresses the Customer's core challenges directly: it raises accuracy, cuts the manual review rate from roughly a third of documents to under a tenth, and extends coverage well beyond pre-defined templates. Investing in structured prompt design during the proof-of-concept phase was central to this, producing the consistent JSON output the downstream platform depends on.

The business impact extends to both cost and compliance. Moving from always-on EC2 to a serverless, pay-per-document model removed idle compute spend and lowered infrastructure cost by an estimated 35%, while a per-document audit trail in DynamoDB — designed around the Customer's compliance requirements — keeps the pipeline traceable and auditable end to end.

 

Technical Capabilities → Business Outcomes

Technical Capability 

Business Outcome

Amazon Textract 

High-accuracy OCR, form, and table extraction across complex documents

Amazon Bedrock (Claude Haiku) 

Classification, validation, and gap-filling for 92–95% accuracy

Event-driven Amazon S3 + AWS

Lambda pipeline 

Fully automated, hands-off document processing

Serverless architecture (no EC2) 

An estimated 35% lower infrastructure cost; no idle spend

Amazon DynamoDB with audit trail 

Traceable, compliance-ready storage of every document result

Amazon SES 

Timely completion and manual-review notifications

Amazon CloudWatch 

Visibility into error rates, processing times, and Bedrock

 

Conclusion

Service Partnership

CloudTry helped the Customer move the platform's document processing pipeline from a rule-based OCR approach to a Generative AI one, integrating Amazon Textract and Amazon Bedrock into a fully serverless AWS architecture. The result raised extraction accuracy, sharply reduced manual review, and replaced always-on infrastructure with a pay-per-document model — without disrupting the Customer's existing AWS setup.

 

Partner Support Services (Pre- and Post-Implementation)

Before implementation, CloudTry led discovery and scoping, producing a statement of work with commercial terms and a total cost of ownership analysis comparing the existing EC2-based setup against the proposed serverless architecture, including per-document cost estimates. The team designed the end-to-end architecture and ran a proof-of-concept phase focused on structured prompt engineering for classification and validation. The DynamoDB audit trail schema was agreed with the Customer up front, since its clients' compliance requirements shaped the design significantly.

After deployment, CloudTry benchmarked accuracy across a diverse 500-document test set to confirm the improvement from the previous pipeline, which gave both teams confidence in the results and made customer sign-off straightforward. Amazon CloudWatch dashboards were delivered as part of the engagement, giving the Customer's team visibility into error rates, processing times, and Bedrock invocation metrics from day one. The engagement ran from February 2026 to March 2026 and closed with a customer satisfaction score of 4.25 out of 5.

 

About CloudTry

CloudTry is a cloud and AI services company headquartered in Lucknow and with offices in Noida, and Kathmandu (Nepal), specializing in the design and deployment of cloud-native and generative AI solutions on AWS. As an AWS Advanced Tier Services Partner, CloudTry brings a team certified across the AWS stack, holding the AWS Certified Cloud Practitioner, AWS Certified AI Practitioner, AWS Certified Machine Learning – Specialty, and AWS Certified Solutions Architect credentials at both Associate and Professional levels. The company works with organizations to move from manual, staff-dependent processes to scalable, production-grade AI systems, with a focus on retrieval-augmented generation, conversational AI, and cost-efficient serverless architectures, among other ready-to-deploy AI solutions, fully built on AWS.