Scalable Intelligent Document Processing with Multi-Tenant Serverless Architecture

ricoh · Healthcare · 2026 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Healthcare

Ricoh needed to scale healthcare document processing for variable, multi-document packets while meeting healthcare compliance and human-review requirements.

Constraints:

documents contain variable layouts and multiple document types
healthcare compliance standards apply
low-confidence extracted fields need human verification
monthly volume was projected to grow sevenfold

Architecture

Ricoh standardized healthcare IDP around AWS GenAI IDP Accelerator Processing Pattern 2: event-driven S3 uploads trigger Lambda functions, Textract extracts text/layout from long scanned packets, and Amazon Bedrock foundation models classify document sections and extract healthcare fields from Textract-augmented prompts. The IDP-Common engine integrates the new AI capabilities into Ricoh's existing high-capacity workflows, while DynamoDB tracks metadata/state, SQS queues and Lambda concurrency throttle Textract/Bedrock calls, and per-customer EventBridge rules plus SQS queues isolate tenants. Confidence scoring compares Textract and Bedrock outputs, routes low-confidence fields to human review queues, and stores final outputs in S3. AWS SAM/CloudFormation creates configurable per-customer deployments with CloudWatch/CloudTrail monitoring for HIPAA/HITRUST/SOC 2 controls.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  s3-document-ingest["s3-document-ingest"] --> lambda-processing["lambda-processing"]
  lambda-processing --> amazon-textract["amazon-textract"]
  lambda-processing --> amazon-bedrock["amazon-bedrock"]
  amazon-textract --> confidence-scoring["confidence-scoring"]
  amazon-bedrock --> confidence-scoring
  confidence-scoring --> human-review-queues["human-review-queues"]
  confidence-scoring --> s3-final-outputs["s3-final-outputs"]
  lambda-processing --> dynamodb-state["dynamodb-state"]
  sqs-rate-control["sqs-rate-control"] --> lambda-processing
  sam-cloudformation-deployment["sam-cloudformation-deployment"] --> lambda-processing
  sam-cloudformation-deployment --> sqs-rate-control
  lambda-processing --> cloudwatch-cloudtrail-audit["cloudwatch-cloudtrail-audit"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`s3-document-ingest`	Receives scanned documents from fax, scan, and SFTP sources and triggers event-driven processing.	`lambda-processing`
`lambda-processing`	Invokes AI services, handles document classification/extraction orchestration, and emits captured attributes.	`s3-document-ingest`, `amazon-textract`, `amazon-bedrock`, `dynamodb-state`, `sqs-rate-control`
`amazon-textract`	Converts scanned healthcare documents into structured text/layout used to augment Bedrock prompts.	`lambda-processing`, `amazon-bedrock`, `confidence-scoring`
`amazon-bedrock`	Foundation-model layer for intelligent document classification and healthcare-specific field extraction.	`lambda-processing`, `amazon-textract`, `confidence-scoring`
`confidence-scoring`	Compares extraction outputs and assigns per-field certainty scores used to decide human review.	`amazon-textract`, `amazon-bedrock`, `human-review-queues`, `s3-final-outputs`
`human-review-queues`	Routes low-confidence fields to operators for verification, correction, and compliance validation.	`confidence-scoring`
`sqs-rate-control`	Controls Textract/Bedrock request velocity, supports retries, and sends failed work to dead-letter queues.	`lambda-processing`
`dynamodb-state`	Stores document metadata, processing state, and version tracking to prevent conflicting updates.	`lambda-processing`
`sam-cloudformation-deployment`	Creates configurable per-customer serverless stacks for fast onboarding without code changes.	`lambda-processing`, `sqs-rate-control`
`cloudwatch-cloudtrail-audit`	Logs processing, metrics, errors, API calls, and audit events for compliance and operations.	`lambda-processing`
`s3-final-outputs`	Stores final extracted outputs after confidence scoring and human validation.	`confidence-scoring`

Data flow

A scanned healthcare document lands in S3 and triggers Lambda. Lambda calls Textract to extract structured text/layout, and Bedrock classifies the document type and performs structured extraction from tailored prompts enriched with Textract output and metadata. Business rules and confidence scoring compare multiple extraction sources, write metadata and processing state to DynamoDB, store outputs in S3, and route below-threshold fields to human review queues. SQS controls Textract/Bedrock request velocity and retries, while per-customer EventBridge/SQS resources isolate tenants and CloudWatch/CloudTrail record processing and audit events.

5 supporting quotes

“(Amazon S3), triggering event-driven workflows.”article, stated

“functions invoke Amazon Bedrock models to determine document types such as claims, appeals, faxes, grievances, prior authorization requests, and clinical documentation.”article, stated

“Amazon Textract parses text and layout, and the results are combined with Amazon Bedrock models for structured data extraction.”article, stated

“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”article, stated

“rule and SQS queue, enabling multi-tenant isolation”article, stated

Why this architecture

Processing Pattern 2 was chosen because Ricoh needed both OCR/layout extraction and model-driven classification/extraction for multi-part healthcare packets that exceed image-context limits. Serverless and pay-per-use infrastructure matched bursty customer volumes without idle cost. Per-customer queues support tenant isolation, SQS throttling prevents quota exhaustion, and confidence-scored human review preserves healthcare accuracy while reducing manual review. Configuration-driven SAM/CloudFormation deployments turn bespoke per-customer builds into reusable onboarding.

4 supporting quotes

“combines Amazon Textract for OCR—the technology that converts images of text into machine-readable text—with Amazon Bedrock FMs”article, stated

“This pattern is specifically designed for complex, multi-part documents that require both text extraction and AI-powered understanding.”article, stated

“The serverless approach was chosen to remove infrastructure management overhead and align costs directly with usage”article, stated

“Configurable confidence thresholds proved essential for meeting customer requirements—helping teams keep wrongly captured attributes below agreed SLAs while minimizing the cost of manual review.”article, stated

Models used: Amazon Bedrock foundation models

Scale

Ricoh serves customers in over 200 countries and processes hundreds of thousands of critical healthcare documents each month.
“Ricoh USA, Inc. is a global technology leader serving a diverse client base in over 200 countries.”article, stated
“processing hundreds of thousands of critical documents each month, including insurance claims, grievances, appeals, and clinical records for their clients.”article, stated
Healthcare packets ranged from 15-50 pages and could include multiple document types.
“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”article, stated
Human-in-the-loop review maintains 98-99% accuracy while reducing manual review costs by 60-70%.
“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.”article, stated
The reusable framework reduced customer onboarding from 4-6 weeks to 2-3 days.

"Onboarding Time

4–6 weeks

2–3 days

>90% reduction"

— article, stated

Monthly throughput moved from about 10,000 documents toward more than 70,000 documents, a sevenfold increase.

"Monthly Throughput

~10,000 documents

>70,000 documents

7-fold increase"

— article, stated

The architecture can process traffic spikes of 1,000 documents in minutes.
“Processing Capacity Limited 1,000 documents in minutes Handles traffic spikes”article, implied

Alternatives considered (roads not taken)

Sending scanned documents directly to Bedrock multimodal models

Rejected because: Large multi-part healthcare documents exceed context limits and need classification before extraction rules can be applied.

“a simple API call to Amazon Bedrock—that is, sending a scanned document along with a prompt—would not suffice for complex workflows.”article, stated

“large language models (LLMs) have context window limits and experience declining performance in following instructions as the context fills.”article, stated

Bespoke manual engineering for each customer deployment

Rejected because: Custom builds consumed weeks of onboarding and 40-60 developer hours per customer, limiting expansion.

“Each new healthcare customer implementation required unique development and tuning by specialized engineers.”article, stated

“these workflows demanded 40–60 hours of developer time per customer to set up”article, stated

Fully manual document review

Rejected because: Confidence-scored automation lets high-confidence fields proceed while humans focus on uncertain cases, preserving accuracy at lower review cost.

“Every attribute with a confidence level below a predefined threshold is reviewed by a person to verify accuracy and compliance.”article, stated

“Human indexers now concentrate their time on difficult documents and extractions, with AI serving as their partner in the process rather than performing routine data entry.”article, stated

Fixed infrastructure sized for peak customer document volume

Rejected because: Serverless pay-per-use scaling removes idle costs and lets each customer's cost track actual document processing.

“The pay-per-use pricing model removes idle infrastructure costs—Ricoh only pays for actual document processing, with no charges during periods of inactivity.”article, stated

Patterns

`human-in-the-loop-gate` (governance)

Extracted fields below a confidence threshold are routed to human reviewers for verification and correction.

“Every attribute with a confidence level below a predefined threshold is reviewed by a person to verify accuracy and compliance.”article, stated

`privacy-preserving-processing` (governance)

The document-processing architecture is explicitly constrained by healthcare compliance standards.

“This solution needed to meet strict compliance standards, including HITRUST, HIPAA, and SOC II.”article, stated

`prompt-chaining` (workflow)

The summary describes a sequential classify-then-extract pipeline where Bedrock prompts are augmented with Textract output and document metadata.

“This sequential processing approach (classify first, then extract based on classification) proved essential for handling the complex, multi-part healthcare documents where extraction rules depend on accurately identifying document section types.”summary, stated

Tradeoffs

Use hybrid OCR plus foundation-model extraction for long healthcare packets.

Chose Textract layout/text extraction feeding Bedrock prompts over sending complete document images directly to a multimodal model.

Rationale: The summary states this avoided context-window limits for long clinical records.

“The decision to use a hybrid OCR-plus-FM approach rather than relying solely on multimodal models sent to Amazon Bedrock solved practical limitations around context window constraints.”summary, stated

Standardize deployments through configuration rather than bespoke code.

Chose configurable AWS SAM/CloudFormation stacks per customer over custom manual engineering for each implementation.

Rationale: Configuration-driven deployments reduced onboarding time and engineering hours while preserving customer-specific parameters.

“application deployed as an AWS CloudFormation stack, supporting rapid onboarding.”article, stated

“This focused approach is valuable when onboarding a new customer.”article, implied

Throttle AI-service calls with SQS rather than relying on unconstrained Lambda concurrency.

Chose SQS queue velocity control with Lambda concurrency settings and visibility timeouts over unbounded direct calls to Textract and Bedrock endpoints.

Rationale: The queue protects service quotas, supports retries, and improves reliability under traffic spikes.

“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”article, stated

“Amazon SQS seamlessly facilitates retries and sending of unprocessed messages to a DLQ.”article, stated

Use confidence-scored human review instead of full automation or fully manual processing.

Chose human review only for fields below customer-defined confidence thresholds over manual review of all fields or automatic acceptance of all extracted fields.

Rationale: The approach maintains payer-required accuracy while reducing manual-review cost.

“Fields scoring below customer-defined thresholds (typically 70–85%) are flagged for human validation.”article, stated

“This approach balanced automation benefits with the accuracy requirements of healthcare document processing.”article, stated

Gotchas (what bit them)

Large healthcare packets can exceed practical multimodal context limits.

Root cause: A single packet may contain dozens of pages with multiple document types and variable layouts.

Resolution: Ricoh used Amazon Textract output to augment Bedrock prompts instead of relying only on end-to-end multimodal processing.

“Healthcare documents often exceed the context windows of foundation models when processed as images—a 50-page clinical record would overwhelm most LLM context limits.”summary, stated

Evaluation

Confidence scores determine which fields require human review.

Metrics: field confidence score · human review threshold

“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.”article, stated

Outcomes

The article reports human-in-the-loop review maintains 98-99% accuracy while reducing manual review costs by 60-70% compared with fully manual processing.

The source says capacity was projected to grow to more than 70,000 documents per month.

“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.”article, stated