How Ricoh constructed a scalable clever doc processing resolution on AWS

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

This submit demonstrates how enterprises can overcome doc processing scaling limits by combining generative AI, serverless structure, and standardized frameworks. Ricoh engineered a repeatable, reusable framework utilizing the AWS GenAI Clever Doc Processing (IDP) Accelerator. This framework lowered buyer onboarding time from weeks to days. It additionally elevated processing capability for brand spanking new AI-intensive workflows that required advanced doc splitting. The capability is projected to develop sevenfold to over 70,000 paperwork monthly. Moreover, the answer decreased engineering hours per deployment by over 90%.

Ricoh USA, Inc. is a world expertise chief serving a various shopper base in over 200 international locations. Inside its healthcare observe, Ricoh serves main medical insurance payers, managed care organizations, and healthcare suppliers—processing tons of of 1000’s of important paperwork every month, together with insurance coverage claims, grievances, appeals, and medical information for his or her shoppers. They confronted a problem widespread to enterprises modernizing document-heavy workflows: reliance on {custom} guide engineering. Every new healthcare buyer implementation required distinctive growth and tuning by specialised engineers. Moreover, deployment required {custom} immediate engineering, mannequin fine-tuning, and integration testing that couldn’t be reused throughout clients. Though this offered an distinctive, bespoke expertise for Ricoh clients, the effort and time concerned created bottlenecks that restricted enlargement. With an anticipated sevenfold enhance in quantity, Ricoh seized the chance to innovate.

The problem was not simply to automate processes. It was to construct a scalable resolution that might ship state-of-the-art AI for doc extraction and agentic workflows. This resolution wanted to satisfy strict compliance requirements, together with HITRUST, HIPAA, and SOC II. These necessities typically stand at odds with fast AI innovation. Compliance frameworks usually limit information sharing that limits mannequin coaching capabilities. In addition they mandate rigorous safety controls that may impede the agility wanted for iterative AI growth and deployment. Regardless of these challenges, Ricoh made it a precedence to beat this pressure for his or her clients. Constructing upon basis fashions (FMs) accessible by Amazon Bedrock and mixing them with Amazon Textract, Ricoh made it doable for purchasers to learn from cutting-edge automation that aligns with the strictest compliance requirements.

This submit explores how Ricoh constructed a standardized, multi-tenant resolution for automated doc classification and extraction utilizing the AWS GenAI IDP Accelerator as a basis, reworking their doc processing from a custom-engineering bottleneck right into a scalable, repeatable service.

Buyer overview

Ricoh USA, Inc. is a world expertise chief delivering digital office companies, doc administration, and enterprise course of automation options to organizations in over 200 international locations. Inside its healthcare observe, Ricoh serves main medical insurance payers, managed care organizations, and healthcare suppliers—processing 1000’s of important paperwork every month, together with insurance coverage claims, grievances, appeals, and medical information.

“Inside the Ricoh Clever Enterprise Platform, the workflows that required the very best ranges of intelligence for key IDP duties skilled explosive development. We wanted to maneuver from bespoke builds to a platform,” says Jeremy Jacobson, AI Architect, Portfolio Resolution Improvement at Ricoh. “For our clients, we combine, function, and evolve AI so that they don’t need to. Aligning our proprietary IDP patterns and applied sciences with the AWS GenAI IDP accelerator amplified this benefit. So outfitted, we delivered a HITRUST CSF-certified configurable IDP platform that ties our clients to the frontiers of AI.”

Healthcare paperwork typically arrive unstructured and extremely variable. A single packet may embody a number of doc varieties—fax covers, medical notes, and attraction kinds—every with completely different layouts and naming conventions. Paperwork ranged from 15–50 pages, with some containing cowl letters whereas others didn’t. Totally different healthcare suppliers used various doc constructions, subject naming conventions, and placement of important info throughout completely different healthcare suppliers. Template-based extraction approaches proved ineffective.

For Ricoh’s Clever Enterprise Platform companies, purposeful necessities included capturing information attributes from scans of unstructured or semi-structured paperwork and assigning to every information attribute a confidence degree that reliably identifies when human evaluate is required. Each attribute with a confidence degree under a predefined threshold is reviewed by an individual to confirm accuracy and compliance. Human reviewers confirm extracted information, appropriate errors, and validate that important healthcare info—reminiscent of member IDs, analysis codes, and declare quantities—meets the standard requirements required for regulatory compliance alignment and claims processing. This human-in-the-loop strategy achieves two key enterprise outcomes: sustaining the excessive accuracy ranges (usually 98–99%) required by healthcare payers whereas lowering guide evaluate prices by 60–70% in comparison with totally guide processing.

The answer wanted to extract key information reminiscent of member IDs, supplier info, and declare particulars from varied sections of paperwork, with the potential to look by medical notes and different sections when info was not present in cowl letters. Non-functional necessities addressed a number of important operational wants:

• Efficiency and scalability – Deal with site visitors spikes to course of as much as 1,000 paperwork in minutes whereas avoiding wasted computational sources throughout low-traffic durations
• Accuracy and high quality – Meet strict service degree agreements (SLAs) for supply deadlines and information accuracy
• Value optimization – Allow configurable confidence thresholds that stability accuracy necessities with guide evaluate prices—conserving wrongly captured attributes under the agreed SLA whereas minimizing costly human evaluate
• Operational effectivity – Allow fast buyer onboarding by configuration adjustments reasonably than code adjustments

Challenges with advanced doc processing workflows

For a while, the Ricoh crew had mixed conventional optical character recognition (OCR)—which detects and extracts textual content from scanned paperwork—with multimodal AI fashions that may perceive each textual content and pictures concurrently. This strategy helped tackle advanced challenges reminiscent of distinguishing between comparable fields when extracting information from paperwork with a number of names and addresses.

After multimodal FMs grew to become accessible on Amazon Bedrock, it quickly grew to become clear {that a} easy API name to Amazon Bedrock—that’s, sending a scanned doc together with a immediate—wouldn’t suffice for advanced workflows. When paperwork are composed of a number of elements or sections, reminiscent of cowl sheets, contracts, or authorization responses, extraction guidelines typically depend on first efficiently classifying the part kind.

The answer wanted to deal with advanced doc classification, distinguishing between claims, disputes, emails, and fax cowl sheets with out breaking down packets into granular doc varieties. Moreover, giant language fashions (LLMs) have context window limits and expertise declining efficiency in following directions because the context fills. Doc web page dimension limitations required the Ricoh crew to make use of various approaches for bigger paperwork.

The Ricoh crew additionally required flexibility to combine with their present high-capacity doc processing workflows—together with doc routing techniques, case administration companies, and downstream enterprise purposes—whereas sustaining management over processing steps and mannequin choice. This included distinctive necessities reminiscent of splitting paperwork based mostly on healthcare supplier or affected person info.

To enhance accuracy, the Ricoh crew utilized extra subtle technique of dynamically inserting context into prompts—a way the place related doc metadata, beforehand extracted fields, and doc construction info are programmatically added to the AI mannequin’s directions based mostly on the precise doc being processed. This context-aware prompting improved extraction accuracy by 15–20% in comparison with static prompts, serving to the mannequin perceive doc relationships and subject dependencies.

Though these beneficial properties have been substantial, when making an attempt to recreate this success, the Ricoh crew ran right into a persistent hurdle: these workflows demanded 40–60 hours of developer time per buyer to arrange, as an example to include newly launched options of the underlying fashions. Ricoh coordinated with the AWS Generative AI Innovation Heart on the IDP Accelerator to deal with these scalability challenges.

Resolution overview

Ricoh partnered with AWS to implement the GenAI IDP Accelerator, a reference framework designed that will help you deploy production-grade doc processing options. The accelerator offers a number of processing patterns optimized for various doc varieties and workflows.

The crew chosen Processing Sample 2, which mixes Amazon Textract for OCR—the expertise that converts pictures of textual content into machine-readable textual content—with Amazon Bedrock FMs for clever classification and extraction. This sample is particularly designed for advanced, multi-part paperwork that require each textual content extraction and AI-powered understanding. The strategy provided full management over mannequin orchestration and was very best for dealing with Ricoh’s multi-part healthcare paperwork as a result of it helps sequential processing (classify first, then extract based mostly on classification) and handles paperwork exceeding typical LLM context home windows by processing them in sections.

The answer was architected to align with stringent healthcare compliance necessities. For HIPAA compliance, the Protected Well being Data (PHI) is encrypted at relaxation utilizing AWS Key Administration Service (AWS KMS) and in transit utilizing TLS 1.2+. Entry controls observe the precept of least privilege, with AWS Identification and Entry Administration (IAM) insurance policies proscribing information entry to approved personnel solely.

For HITRUST certification necessities, the structure implements complete audit logging by Amazon CloudWatch and AWS CloudTrail, capturing information entry and processing actions. SOC 2 Sort II compliance alignment is supported by the usage of AWS companies that preserve their very own SOC 2 certifications, mixed with Ricoh’s documented operational controls for change administration, occasion response, and steady monitoring.

The pay-per-use pricing mannequin removes idle infrastructure prices—Ricoh solely pays for precise doc processing, with no fees in periods of inactivity. This value predictability was essential for supporting a number of clients with various doc volumes, as every buyer’s prices scale proportionally with their utilization reasonably than requiring fastened infrastructure investments.

Paperwork enter utilizing Amazon Easy Storage Service (Amazon S3), triggering event-driven workflows. AWS Lambda features invoke Amazon Bedrock fashions to find out doc varieties reminiscent of claims, appeals, faxes, grievances, prior authorization requests, and medical documentation. Amazon Textract parses textual content and format, and the outcomes are mixed with Amazon Bedrock fashions for structured information extraction. Customized enterprise guidelines—configurable logic particular to every buyer’s necessities, reminiscent of subject validation guidelines, doc routing standards, and information transformation specs—work alongside confidence scoring to find out which fields require human evaluate.

Confidence scores are calculated by evaluating extraction outcomes from a number of sources (Amazon Textract and Amazon Bedrock) and assigning a numerical worth (0–100%) indicating the system’s certainty in every extracted subject. Fields scoring under customer-defined thresholds (usually 70–85%) are flagged for human validation. Last outputs are saved in Amazon S3, with low-confidence circumstances routed for human validation by evaluate queues the place operators confirm extracted information, appropriate errors, and supply suggestions that improves future processing.

The core IDP-Frequent engine from the AWS GenAI IDP Accelerator served as the combination layer, serving to Ricoh preserve its established workflows. The IDP Frequent Bundle is a Python library that gives shared performance for the Accelerated Clever Doc Processing resolution on AWS. This resolution helps companies robotically extract and course of info from paperwork utilizing AI companies, eradicating guide information entry and bettering accuracy.

Every buyer deployment is instantiated utilizing a configurable AWS Serverless Software Mannequin (AWS SAM) software deployed as an AWS CloudFormation stack, supporting fast onboarding. This abstracts away infrastructure particulars—together with Amazon Digital Personal Cloud (Amazon VPC) configuration, safety group guidelines, IAM function insurance policies, and repair quotas—so crew members can focus solely on the customer-dependent parameters reminiscent of Lambda reserved concurrency or database connection particulars. This targeted strategy is effective when onboarding a brand new buyer.

The modular design helped Ricoh combine particular parameters and {custom} performance reminiscent of customer-defined proprietary doc classification, {custom} information extraction for industry-specific kinds, or redaction guidelines for personally identifiable info (PII) compliance alignment into their present high-capacity workflow with out disrupting established processes. This strategy helped the crew preserve operational effectivity by automated deployment that lowered buyer onboarding time from weeks to days, whereas including superior AI capabilities for doc processing, together with clever doc classification, and automatic information extraction from unstructured kinds.

Structure particulars

The structure was designed with three main aims: allow fast buyer onboarding by configuration reasonably than code adjustments, assist align with healthcare laws (HIPAA, HITRUST, SOC 2), and supply cost-effective scalability for variable doc volumes. The serverless strategy was chosen to take away infrastructure administration overhead and align prices instantly with utilization, and the multi-tenant design with per-customer queues balances useful resource effectivity with workload isolation. The choice to make use of Processing Sample 2 (Amazon Textract and Amazon Bedrock) reasonably than Amazon Bedrock alone was pushed by the necessity to deal with paperwork exceeding LLM context home windows and the requirement for structured textual content extraction that might be selectively included in prompts based mostly on doc kind.

The implementation used a serverless structure by which Lambda features are robotically invoked upon add of scanned paperwork to Amazon S3. The Lambda features deal with calls to the AI companies—Amazon Textract and Amazon Bedrock—and output the captured attributes together with their confidence scores to an Amazon DynamoDB database.

The structure incorporates AWS Properly-Architected Framework rules throughout a number of pillars. For safety, the information is encrypted at relaxation utilizing AWS KMS with customer-managed keys and in transit utilizing TLS 1.2+. IAM roles implement least-privilege entry, separated by perform, with separate roles for doc ingestion, processing, and retrieval. CloudTrail logs the API requires audit trails, and CloudWatch Logs captures application-level occasions for safety monitoring.

For reliability, the serverless design removes single factors of failure, with automated retries and dead-letter queues (DLQs) dealing with transient errors. For efficiency effectivity, Lambda concurrency limits and Amazon Easy Queue Service (Amazon SQS) queue throttling helps forestall API quota exhaustion whereas sustaining excessive throughput. For value optimization, the pay-per-use mannequin removes idle useful resource prices, and Amazon S3 lifecycle insurance policies robotically transition processed paperwork to lower-cost storage tiers.

For operational excellence, infrastructure as code utilizing AWS SAM and CloudFormation allows constant deployments, and CloudWatch dashboards and alarms present real-time visibility into processing metrics and error charges.

A important a part of the structure is an SQS queue that makes it doable for the crew to regulate the speed at which they’re making requests to Amazon Textract and Amazon Bedrock API endpoints by controlling message processing velocity by Lambda concurrency settings and Amazon SQS visibility timeouts. This design helps them keep inside service quota limits (reminiscent of transactions per second for Amazon Textract and requests per minute for Amazon Bedrock). Moreover, Amazon SQS seamlessly facilitates retries and sending of unprocessed messages to a DLQ.

Every buyer has its personal Amazon EventBridge rule and SQS queue, enabling multi-tenant isolation (serving to forestall one buyer’s excessive quantity from impacting others) and unbiased scaling (permitting per-customer concurrency limits and throughput controls).

The structure used Amazon S3 for doc storage. Totally different buckets have been created to handle paperwork from varied sources, together with fax, scan, and SFTP techniques. DynamoDB tables saved doc metadata and processing state, monitoring doc variations and serving to forestall a number of makes an attempt to replace the identical doc concurrently. CloudWatch offered complete monitoring and logging of profitable extraction charges and processing anomalies.

The precise interplay with AI companies makes use of Amazon Textract to reinforce Amazon Bedrock prompts with structured information extracted from the scanned doc. Right here, the crew took benefit of their earlier Amazon Textract based mostly resolution and used it as one other supply of reality for the extracted values, which make it doable to compute dependable confidence scores by evaluating outcomes from each extraction strategies. This dual-extraction strategy was used throughout the preliminary deployment section to validate accuracy, with the legacy system phased out after confidence within the new system was established.

For doc processing, the answer used Amazon Textract to extract textual content from giant healthcare paperwork, addressing the problem of paperwork that exceeded the context window limitations of FMs when processed as pictures. For instance, a 50-page medical report would exceed most LLM context home windows if despatched as pictures, however Amazon Textract converts it to structured textual content that may be selectively included in prompts. Amazon Bedrock FMs dealt with the clever classification and extraction duties, with tailor-made directions for healthcare information designed to determine doc varieties and extract healthcare-specific info reminiscent of member IDs, supplier particulars, and declare info.

For doc classification and splitting, the crew used LLMs to intelligently determine doc varieties and cut up multi-document packets based mostly on supplier or affected person info.

Concerning quick onboarding for brand spanking new clients, the crew used a configurable AWS SAM software deployed as a CloudFormation nested stack for every buyer. This abstracts away infrastructure particulars—reminiscent of VPC configuration, safety group guidelines, IAM function insurance policies, and repair quotas—and so crew members can focus solely on the customer-dependent parameters when onboarding a brand new buyer.

The modular structure helped Ricoh deploy solely the parts they wanted whereas sustaining the choice so as to add further options reminiscent of doc summarization or data base integration sooner or later.

Outcomes and outcomes

Ricoh has been in a position to decrease costs for an essential healthcare buyer by measuring and attaining vital reductions in human labor required to index paperwork in manufacturing. Human indexers now focus their time on troublesome paperwork and extractions, with AI serving as their companion within the course of reasonably than performing routine information entry.

Ricoh’s Clever Enterprise Platform achieved vital operational enhancements and potential annual financial savings exceeding 1,900 person-hours by automation, dramatically lowering the guide effort required for doc processing.

The automated classification system efficiently distinguished between insurance coverage coverage holders’ grievances and appeals claims, a important functionality for healthcare compliance and workflow administration. These doc varieties have completely different regulatory timelines (grievances usually require 30-day decision, appeals require 60 days) and should be routed to completely different processing groups. Misclassification may end up in missed deadlines, regulatory penalties, and member dissatisfaction.

The answer demonstrated extraction accuracy ranges that assist reduce monetary penalties from processing errors, a vital final result within the closely regulated healthcare {industry}. The arrogance scoring capabilities enabled efficient human-in-the-loop evaluate processes, serving to confirm that paperwork requiring skilled validation have been correctly flagged whereas permitting high-confidence extractions to proceed robotically.

Ricoh efficiently created a reusable framework that may be deployed throughout a number of healthcare clients, offering a scalable basis for increasing their doc processing companies to future use circumstances. The answer now processes over 10,000 healthcare paperwork month-to-month with the infrastructure in place to scale to 70,000 paperwork as shopper wants develop.

The Clever Enterprise Platform achieved vital operational enhancements, as detailed within the following desk.

Greatest practices and classes discovered

The Ricoh implementation highlighted a number of finest practices for deploying IDP options in manufacturing environments:

• Select the suitable processing sample – Choosing Sample 2 from the AWS IDP Accelerator offered the pliability wanted for advanced healthcare doc necessities whereas sustaining management over mannequin choice and processing steps. This selection was important for dealing with distinctive doc splitting necessities and integration with present workflows.
• Use a hybrid strategy combining OCR with FMs – The crew discovered that utilizing Amazon Textract to reinforce Amazon Bedrock prompts with structured information offered each scalability and accuracy for paperwork of various sizes and complexity. This hybrid strategy of mixing OCR with FMs addressed sensible limitations round context window sizes when processing paperwork as pictures—Amazon Textract handles paperwork of various sizes, and Amazon Bedrock offers clever understanding of the extracted textual content, enabling each scalability (no doc dimension limits) and accuracy (AI-powered subject extraction and validation). Profiting from the earlier Amazon Textract based mostly resolution as one other supply of reality throughout the validation section helped the crew compute dependable confidence scores with out incurring vital further prices, as a result of Amazon Textract was already getting used for textual content extraction within the new structure.
• Combine confidence scoring from the start – Integrating confidence scoring from the start enabled efficient human-in-the-loop workflows, permitting the system to robotically flag unsure extractions for skilled evaluate. This strategy balanced automation advantages with the accuracy necessities of healthcare doc processing. Configurable confidence thresholds proved important for assembly buyer necessities—serving to groups hold wrongly captured attributes under agreed SLAs whereas minimizing the price of guide evaluate.
• Implement fee limiting with SQS queues – Implementing an SQS queue to restrict the speed of API calls to Amazon Textract and Amazon Bedrock endpoints helped the crew keep inside quota limits whereas seamlessly facilitating retries and DLQ dealing with. This architectural determination helped forestall throttling points and improved general system reliability.
• Standardize utilizing configuration reasonably than code – Standardizing utilizing configuration reasonably than code adjustments was a key enabler of fast buyer onboarding. The configurable AWS SAM software deployed as a CloudFormation nested stack for every buyer abstracted away infrastructure particulars, so crew members might focus solely on customer-dependent parameters. This strategy lowered upkeep efforts and enabled fast onboarding for brand spanking new clients.
• Use a modular structure for integration – The modular structure of the GenAI IDP Accelerator proved useful for integration with present techniques. Fairly than changing established workflows, the core IDP-Frequent engine helped Ricoh improve their present infrastructure with AI capabilities—together with doc classification, clever subject extraction, confidence scoring, and pure language understanding.
• Plan for scalability from the outset – Planning for scalability from the outset enabled easy development from proof of idea to manufacturing volumes. The serverless structure’s automated scaling capabilities and pay-per-use pricing mannequin aligned infrastructure prices with enterprise development, offering predictable economics as doc volumes elevated. The structure dealt with spikes in site visitors to seamlessly course of as much as 1,000 paperwork in a couple of minutes whereas not losing computational sources in periods of low or no site visitors.

Getting began

Able to construct your individual IDP resolution? The AWS GenAI IDP Accelerator offers a confirmed basis for deploying production-grade doc automation:

• Discover the accelerator – Go to the GenAI IDP Accelerator repository to evaluate the structure patterns, deployment guides, and pattern code
• Select your sample – Evaluation the a number of processing patterns accessible and choose the one that most closely fits your doc varieties and workflow necessities
• Begin small, scale quick – Start with a proof of idea utilizing your most difficult doc varieties, then use the modular structure to increase throughout your group
• Leverage AWS experience – Join with AWS Options Architects and the GenAI Innovation Heart to debate your particular use case and implementation technique

For organizations processing excessive volumes of advanced paperwork, the mix of serverless structure, FMs, and standardized frameworks provides a path to fast deployment and scalable development.

Conclusion

Ricoh’s implementation of the AWS GenAI IDP Accelerator demonstrates how enterprises can overcome scaling limits by combining generative AI, serverless structure, and compliance frameworks. The result’s sooner onboarding, greater accuracy, and lowered operational overhead—all with out compromising compliance requirements (HIPAA, HITRUST, SOC 2) or operational management. By creating a reusable framework reasonably than single-use options, Ricoh reworked doc processing right into a scalable service.

The Clever Enterprise Platform’s potential to deal with advanced healthcare doc variations, present confidence scoring for human-in-the-loop workflows, and scale from 10,000 to doubtlessly 70,000 paperwork month-to-month showcases the sensible advantages of IDP powered by generative AI on AWS. The reusable framework Ricoh created can now be deployed throughout a number of healthcare clients, offering a basis for increasing their doc processing companies.

For organizations dealing with comparable doc processing challenges, the GenAI IDP Accelerator provides a confirmed path from proof of idea to production-ready options. The mixture of serverless structure, a number of processing patterns, and integration flexibility helps groups construct doc automation tailor-made to their particular wants whereas utilizing the most recent advances in generative AI and AWS companies. Their story is proof that with the appropriate basis, AI doesn’t simply automate work—it could possibly speed up development.

To get began with the GenAI IDP Accelerator, go to the challenge repository and discover the documentation and deployment guides.

Acknowledgments

Particular thanks to Bob Strahan for his management of the GenAI IDP Accelerator challenge. We might additionally prefer to thank Guillermo Tantachuco, Saeideh Shahrokh Esfahani, Mofijul Islam, Suresh Konappanava, and Yingwei Yu for his or her contributions and steerage all through.

In regards to the authors