リコーがAWS上で構築したスケーラブルなインテリジェント文書処理ソリューション

*This post is cowritten by Jeremy Jacobson and Rado Fulek from Ricoh.*

This post demonstrates how enterprises can overcome document processing scaling limits by combining generative AI, serverless architecture, and standardized frameworks. Ricoh engineered a repeatable, reusable framework using the AWS GenAI Intelligent Document Processing (IDP) Accelerator. This framework reduced customer onboarding time from weeks to days. It also increased processing capacity for new AI-intensive workflows that required complex document splitting. The capacity is projected to grow sevenfold to over 70,000 documents per month. Additionally, the solution decreased engineering hours per deployment by over 90%.

Ricoh USA, Inc. is a global technology leader serving a diverse client base in over 200 countries. Within its healthcare practice, Ricoh serves major health insurance payers, managed care organizations, and healthcare providers—processing hundreds of thousands of critical documents each month, including insurance claims, grievances, appeals, and clinical records for their clients. They faced a challenge common to enterprises modernizing document-heavy workflows: reliance on custom manual engineering. Each new healthcare customer implementation required unique development and tuning by specialized engineers. Additionally, deployment required custom prompt engineering, model fine-tuning, and integration testing that couldn’t be reused across customers. Although this provided an exceptional, bespoke experience for Ricoh customers, the time and effort involved created bottlenecks that limited expansion. With an anticipated sevenfold increase in volume, Ricoh seized the opportunity to innovate.

The challenge was not just to automate processes. It was to build a scalable solution that could deliver state-of-the-art AI for document extraction and agentic workflows. This solution needed to meet strict compliance standards, including HITRUST, HIPAA, and SOC II. These requirements often stand at odds with rapid AI innovation. Compliance frameworks typically restrict data sharing that limits model training capabilities. They also mandate rigorous security controls that can impede the agility needed for iterative AI development and deployment. Despite these challenges, Ricoh made it a priority to overcome this tension for their customers. Building upon foundation models (FMs) available through Amazon Bedrock and combining them with Amazon Textract, Ricoh made it possible for customers to benefit from cutting-edge automation that aligns with the strictest compliance standards.

This post explores how Ricoh built a standardized, multi-tenant solution for automated document classification and extraction using the AWS GenAI IDP Accelerator as a foundation, transforming their document processing from a custom-engineering bottleneck into a scalable, repeatable service.

Customer overview

Ricoh USA, Inc. is a global technology leader delivering digital workplace services, document management, and business process automation solutions to organizations in over 200 countries. Within its healthcare practice, Ricoh serves major health insurance payers, managed care organizations, and healthcare providers—processing thousands of critical documents each month, including insurance claims, grievances, appeals, and clinical records.

“Within the Ricoh Intelligent Business Platform, the workflows that required the highest levels of intelligence for key IDP tasks experienced explosive growth. We needed to move from bespoke builds to a platform,” says Jeremy Jacobson, AI Architect, Portfolio Solution Development at Ricoh. “For our customers, we integrate, operate, and evolve AI so they don’t have to. Aligning our proprietary IDP patterns and technologies with the AWS GenAI IDP accelerator amplified this advantage. So equipped, we delivered a HITRUST CSF-certified configurable IDP platform that ties our customers to the frontiers of AI.”

Healthcare documents often arrive unstructured and highly variable. A single packet might include multiple document types—fax covers, clinical notes, and appeal forms—each with different layouts and naming conventions. Documents ranged from 15–50 pages, with some containing cover letters while others did not. Different healthcare providers used varying document structures, field naming conventions, and placement of critical information across different healthcare providers. Template-based extraction approaches proved ineffective.

For Ricoh’s Intelligent Business Platform services, functional requirements included capturing data attributes from scans of unstructured or semi-structured documents and assigning to each data attribute a confidence level that reliably identifies when human review is needed. Every attribute with a confidence level below a predefined threshold is reviewed by a person to verify accuracy and compliance. Human reviewers verify extracted data, correct errors, and validate that critical healthcare information—such as member IDs, diagnosis codes, and claim amounts—meets the quality standards required for regulatory compliance alignment and claims processing. This human-in-the-loop approach achieves two key business outcomes: maintaining the high accuracy levels (typically 98–99%) required by healthcare payers while reducing manual review costs by 60–70% compared to fully manual processing.

The solution needed to extract key data such as member IDs, provider information, and claim details from various sections of documents, with the capability to search through clinical notes and other sections when information was not found in cover letters. Non-functional requirements addressed several critical operational needs:

• Performance and scalability – Handle traffic spikes to process up to 1,000 documents in minutes while avoiding wasted computational resources during low-traffic periods

• Accuracy and quality – Meet strict service level agreements (SLAs) for delivery deadlines and data accuracy

• Cost optimization – Enable configurable confidence thresholds that balance accuracy requirements with manual review costs—keeping wrongly captured attributes below the agreed SLA while minimizing expensive human review

• Operational efficiency – Enable quick customer onboarding through configuration changes rather than code changes

Challenges with complex document processing workflows

For some time, the Ricoh team had combined traditional optical character recognition (OCR)—which detects and extracts text from scanned documents—with multimodal AI models that can understand both text and images simultaneously. This approach helped address complex challenges such as distinguishing between similar fields when extracting data from documents with multiple names and addresses.

After multimodal FMs became available on Amazon Bedrock, it soon became clear that a simple API call to Amazon Bedrock—that is, sending a scanned document along with a prompt—would not suffice for complex workflows. When documents are composed of multiple parts or sections, such as cover sheets, contracts, or authorization responses, extraction rules often depend upon first successfully classifying the section type.

The solution needed to handle complex document classification, distinguishing between claims, disputes, emails, and fax cover sheets without breaking down packets into granular document types. Additionally, large language models (LLMs) have context window limits and experience declining performance in following instructions as the context fills. Document page size limitations required the Ricoh team to use alternative approaches for larger documents.

The Ricoh team also required flexibility to integrate with their existing high-capacity document processing workflows—including document routing systems, case management services, and downstream business applications—while maintaining control over processing steps and model selection. This included unique requirements such as splitting documents based on healthcare provider or patient information.

To improve accuracy, the Ricoh team utilized more sophisticated means of dynamically inserting context into prompts—a technique where relevant document metadata, previously extracted fields, and document structure information are programmatically added to the AI model’s instructions based on the specific document being processed. This context-aware prompting improved extraction accuracy by 15–20% compared to static prompts, helping the model understand document relationships and field dependencies.

Although these gains were substantial, when trying to recreate this success, the Ricoh team ran into a persistent hurdle: these workflows demanded 40–60 hours of developer time per customer to set up, for instance to incorporate newly released features of the underlying models. Ricoh coordinated with the AWS Generative AI Innovation Center on the IDP Accelerator to address these scalability challenges.

Solution overview

Ricoh partnered with AWS to implement the GenAI IDP Accelerator, a reference framework designed to help you deploy production-grade document processing solutions. The accelerator provides multiple processing patterns optimized for different document types and workflows.

The team selected Processing Pattern 2, which combines Amazon Textract for OCR—the technology that converts images of text into machine-readable text—with Amazon Bedrock FMs for intelligent classification and extraction. This pattern is specifically designed for complex, multi-part documents that require both text extraction and AI-powered understanding. The approach offered full control over model orchestration and was ideal for handling Ricoh’s multi-part healthcare documents because it supports sequential processing (classify first, then extract based on classification) and handles documents exceeding typical LLM context windows by processing them in sections.

The solution was architected to align with stringent healthcare compliance requirements. For HIPAA compliance, the Protected Health Information (PHI) is encrypted at rest using AWS Key Management Service (AWS KMS) and in transit using TLS 1.2+. Access controls follow the principle of least privilege, with AWS Identity and Access Management (IAM) policies restricting data access to authorized personnel only.

For HITRUST certification requirements, the architecture implements comprehensive audit logging through Amazon CloudWatch and AWS CloudTrail, capturing data access and processing activities. SOC 2 Type II compliance alignment is supported through the use of AWS services that maintain their own SOC 2 certifications, combined with Ricoh’s documented operational controls for change management, event response, and continuous monitoring.

The pay-per-use pricing model removes idle infrastructure costs—Ricoh only pays for actual document processing, with no charges during periods of inactivity. This cost predictability was crucial for supporting multiple customers with varying document volumes, as each customer’s costs scale proportionally with their usage rather than requiring fixed infrastructure investments.

Documents enter using Amazon Simple Storage Service (Amazon S3), triggering event-driven workflows. AWS Lambda functions invoke Amazon Bedrock models to determine document types such as claims, appeals, faxes, grievances, prior authorization requests, and clinical documentation. Amazon Textract parses text and layout, and the results are combined with Amazon Bedrock models for structured data extraction. Custom business rules—configurable logic specific to each customer’s requirements, such as field validation rules, document routing criteria, and data transformation specifications—work alongside confidence scoring to determine which fields require human review.

Confidence scores are calculated by comparing extraction results from multiple sources (Amazon Textract and Amazon Bedrock) and assigning a numerical value (0–100%) indicating the system’s certainty in each extracted field. Fields scoring below customer-defined thresholds (typically 70–85%) are flagged for human validation. Final outputs are stored in Amazon S3, with low-confidence cases routed for human validation through review queues where operators verify extracted data, correct errors, and provide feedback that improves future processing.

The core IDP-Common engine from the AWS GenAI IDP Accelerator served as the integration layer, helping Ricoh maintain its established workflows. The IDP Common Package is a Python library that provides shared functionality for the Accelerated Intelligent Document Processing solution on AWS. This solution helps businesses automatically extract and process information from documents using AI services, removing manual data entry and improving accuracy.

Each customer deployment is instantiated using a configurable AWS Serverless Application Model (AWS SAM) application deployed as an AWS CloudFormation stack, supporting rapid onboarding. This abstracts away infrastructure details—including Amazon Virtual Private Cloud (Amazon VPC) configuration, security group rules, IAM role policies, and service quotas—so team members can focus only on the customer-dependent parameters such as Lambda reserved concurrency or database connection details. This focused approach is valuable when onboarding a new customer.

The modular design helped Ricoh integrate specific parameters and custom functionality such as customer-defined proprietary document classification, custom data extraction for industry-specific forms, or redaction rules for personally identifiable information (PII) compliance alignment into their existing high-capacity workflow without disrupting established processes. This approach helped the team maintain operational efficiency through automated deployment that reduced customer onboarding time from weeks to days, while adding advanced AI capabilities for document processing, including intelligent document classification, and automated data extraction from unstructured forms.

Architecture details

The architecture was designed with three primary objectives: enable rapid customer onboarding through configuration rather than code changes, help align with healthcare regulations (HIPAA, HITRUST, SOC 2), and provide cost-effective scalability for variable document volumes. The serverless approach was chosen to remove infrastructure management overhead and align costs directly with usage, and the multi-tenant design with per-customer queues balances resource efficiency with workload isolation. The decision to use Processing Pattern 2 (Amazon Textract and Amazon Bedrock) rather than Amazon Bedrock alone was driven by the need to handle documents exceeding LLM context windows and the requirement for structured text extraction that could be selectively included in prompts based on document type.

The implementation used a serverless architecture in which Lambda functions are automatically invoked upon upload of scanned documents to Amazon S3. The Lambda functions handle calls to the AI services—Amazon Textract and Amazon Bedrock—and output the captured attributes along with their confidence scores to an Amazon DynamoDB database.

The architecture incorporates AWS Well-Architected Framework principles across multiple pillars. For security, the data is encrypted at rest using AWS KMS with customer-managed keys and in transit using TLS 1.2+. IAM roles enforce least-privilege access, separated by function, with separate roles for document ingestion, processing, and retrieval. CloudTrail logs the API calls for audit trails, and CloudWatch Logs captures application-level events for security monitoring.

For reliability, the serverless design removes single points of failure, with automatic retries and dead-letter queues (DLQs) handling transient errors. For performance efficiency, Lambda concurrency limits and Amazon Simple Queue Service (Amazon SQS) queue throttling helps prevent API quota exhaustion while maintaining high throughput. For cost optimization, the pay-per-use model removes idle resource costs, and Amazon S3 lifecycle policies automatically transition processed documents to lower-cost storage tiers.

For operational excellence, infrastructure as code using AWS SAM and CloudFormation enables consistent deployments, and CloudWatch dashboards and alarms provide real-time visibility into processing metrics and error rates.

A critical part of the architecture is an SQS queue that makes it possible for the team to control the rate at which they are making requests to Amazon Textract and Amazon Bedrock API endpoints by controlling message processing velocity through Lambda concurrency settings and Amazon SQS visibility timeouts. This design helps them stay within service quota limits (such as transactions per second for Amazon Textract and requests per minute for Amazon Bedrock). Furthermore, Amazon SQS seamlessly facilitates retries and sending of unprocessed messages to a DLQ.

Each customer has its own Amazon EventBridge rule and SQS queue, enabling multi-tenant isolation (helping prevent one customer’s high volume from impacting others) and independent scaling (allowing per-customer concurrency limits and throughput controls).

The architecture used Amazon S3 for document storage. Different buckets were created to manage documents from various sources, including fax, scan, and SFTP systems. DynamoDB tables stored document metadata and processing state, tracking document versions and helping prevent multiple attempts to update the same document s