AWS for Industries
Transforming healthcare enrollment with agentic AI for payors
Healthcare payors face unprecedented challenges during open enrollment periods. With millions of member records flooding in through diverse channels and formats, from EDI 834 transactions to scanned PDFs—traditional enrollment systems often buckle under the strain. Processing bottlenecks emerge, manual intervention skyrockets, and member satisfaction suffers during this critical period.
However, these challenges are not inevitable. Modern AI-powered architectures are transforming how payors handle enrollment, reducing processing times from days to minutes, while also lowering manual interventions. Let’s explore a solution where agentic AI and Amazon Bedrock are making this possible.
The open enrollment challenge
Every October, healthcare payors in the United States experience their most demanding operational period. During this annual enrollment window, millions of member records arrive through multiple channels:
- Structured EDI 834 transactions
- FHIR-compliant API submissions
- Unstructured data (scanned forms, PDFs, CSVs)
- Direct portal submissions
Legacy systems typically struggle with this seasonal surge, resulting in processing bottlenecks, extensive manual intervention, and difficulty adapting to evolving regulatory requirements. These challenges directly impact operational efficiency and member satisfaction during this critical period and are symptoms of outdated architecture.
Modern enrollment processing leverages AI agents that understand context rather than just matching patterns. They can adapt to regulatory changes without code deployments, and scale automatically from thousands to millions of submissions. This modern approach transforms enrollment from a labor-intensive seasonal burden into an automated, intelligent process.
Payors implementing our solution architecture can benefit by reducing enrollments processing times to minutes instead of days, while potentially reducing manual intervention by over 60%. The solution can handle the October surges without the capacity planning headaches of traditional systems.
Understanding agentic AI
Before we explore the solution architecture, let’s clarify what we mean by Agentic AI and why it matters for the enrollment processing. Agentic AI refers to artificial intelligence systems that can act autonomously as agents—taking independent actions to achieve goals with minimal human intervention.
Traditional automation works like a flowchart—if this happens, do that. These systems are fast and consistent, but struggle when data doesn’t match expectations. An unexpected format or missing field can halt processing entirely, for example.
Agentic AI works differently. Think of it as deploying a team of AI specialists, each with a domain of expertise:
- A Validation Specialist who reviews submissions for completeness and accuracy
- A Compliance Expert who ensures regulatory requirements are met
- A Data Enrichment Analyst who enhances records with additional context
- An Exception Handler who troubleshoots problems and routes complex cases
Each agent uses foundation models (large AI models trained on vast amounts of data) to understand context, recognize patterns, and make intelligent decisions. They can collaborate, share information, and even explain their reasoning—capabilities far beyond traditional rule-based systems.
For healthcare enrollment, this means systems that can handle the messy reality of real-world data: inconsistent formats, missing information, ambiguous entries, and exceptions that don’t fit predetermined categories.
Our innovative solution combines the AI capabilities of Amazon Bedrock with the specialized healthcare-focused agents. It streamlines enrollment processing, reduces manual workloads, and maintains strict compliance standards—all while improving member satisfaction during peak enrollment periods.
Architecture overview: agentic AI for healthcare enrollment
The big picture: How it all works together
Before diving into the technical details, let’s understand what happens when an enrollment submission enters this system. Think of it like a smart mail sorting and processing facility. Imagine a postal facility that doesn’t only sort mail by zip code, but reads each letter, understands its contents, makes intelligent decisions about handling, and even corrects addresses when they’re incomplete. That’s essentially what the solution architecture does for healthcare enrollments.
Difference from traditional systems
Traditional enrollment systems are like assembly lines—each station performs one specific task in sequence, and if anything doesn’t fit the expected pattern, the entire line stops. Our solution architecture is more like a team of intelligent specialists who collaborate, share information, and adapt to what they encounter.
The solution leverages Amazon Bedrock, Amazon Bedrock AgentCore, and AWS B2B Data Interchange for EDI. It creates an intelligent, scalable solution capable of handling enrollment data at massive scale, while maintaining compliance and data integrity.
- Amazon Bedrock provides access to foundation models—these are large AI models trained on vast amounts of data that power intelligent decision-making across the enrollment process. Think of foundation models as the brain that gives agents their ability to understand context and meaning.
- Amazon Bedrock AgentCore orchestrates multiple specialized AI agents, enabling them to work together seamlessly, while maintaining context throughout complex workflows. It’s the coordinator that confirms agents collaborate effectively and share information.
- AWS B2B Data Interchange handles the parsing, validation, and transformation of EDI 834 transactions, verifying HIPAA-compliant processing of structured enrollment data. It eliminates the need for custom EDI parsing code, which is notoriously difficult to build and maintain.
Together, these services create a solution architecture base that can process millions of enrollment records during peak periods, while adapting to regulatory changes and continuously improving through machine learning.
Figure 1: Agentic AI healthcare enrollment architecture
The architecture consists of five key layers:
- Data ingestion layer
- AI processing layer
- Orchestration and processing
- Storage and state management
- Human-in-the-loop integration
1. Data ingestion layer
The ingestion layer is designed to accept enrollment data from multiple sources:
- AWS B2B for EDI: Processes and validates EDI 834 transactions according to HIPAA standards
- Amazon API Gateway: Handles FHIR API submissions and portal data
- Amazon Simple Storage Service (Amazon S3): Stores incoming PDFs, CSVs, and other documents
- Amazon Bedrock Data Automation: Extracts data from scanned forms and documents
This multi-channel approach verifies payors can meet members where they are, regardless of submission methods, while maintaining a consistent processing pipeline.
2. AI processing layer
At the core of our architecture is the AI processing layer powered by Amazon Bedrock. It employs four specialized agents running on Amazon Bedrock AgentCore:
- Validation Agent: Verifies data completeness, format compliance, and adherence to business rules
- Compliance Check Agent: Confirms regulatory compliance with healthcare requirements
- Enrichment Agent: Enhances enrollment data with additional information from connected systems
- Error Handling Agent: Identifies, categorizes, and addresses exceptions
These agents leverage foundation models within Amazon Bedrock, along with specialized APIs through the Model Context Protocol (MCP) as toolsets for domain-specific tasks. The agent architecture enables complex decision-making capabilities that go beyond traditional rule-based approaches.
3. Orchestration and processing
AWS Step Functions coordinate the workflow between agents and services, verifying each enrollment follows the appropriate processing path:
- Routing submissions based on format and content
- Managing validation workflows
- Handling exceptions and human review processes
- Coordinating integration with core systems
AWS Lambda functions power the serverless processing components, allowing the system to scale instantly during peak periods and economically handle the dramatic volume fluctuations inherent to enrollment seasons.
4. Storage and state management
The solution uses purpose-optimized storage services:
- Amazon S3: Stores raw documents and artifacts
- Amazon DynamoDB: Maintains processing state and metadata
- Amazon ElastiCache: Provides high-performance caching for common reference data
This hybrid approach balances performance needs with cost efficiency across varying workload patterns.
5. Human-in-the-loop integration
Despite advanced AI capabilities, human expertise remains essential for complex scenarios. The architecture includes:
- Exception handling queues route cases requiring manual review with rich context, including why the exception occurred, what the AI agents attempted, confidence scores, and relevant supporting data. The system prioritizes exceptions based on business impact and urgency.
- Learning feedback loops capture human decisions and corrections, feeding this information back to the AI agents to improve future processing. When reviewers resolve exceptions, the system learns from these patterns to handle similar cases automatically in the future.
- Dashboards and analytics provide operational visibility into exception patterns, processing metrics, agent performance, and areas for improvement, enabling data-driven optimization of the entire enrollment process.
Key technical innovations
Agentic AI approach
Rather than applying a single AI model to the entire enrollment process, our architecture employs four specialized agents with domain expertise:
- Validation Agent
- Compliance Check Agent
- Enrichment Agent
- Error Handling Agent
Figure 2: Process flow diagram for enrollment submission
Each agent can access specialized tools, reference data, and APIs while maintaining context throughout the enrollment journey. This approach delivers more accurate decisions than general-purpose AI models alone.
1. Validation Agent
The Validation Agent serves as the system’s first line of defense, confirming data integrity and completeness.
Unlike traditional validation routines that only check field formats, this agent:
- Intelligently extracts information from varying document layouts
- Understands semantic equivalence across different terminology
- Correlates information across multiple fields and documents
- Detects logical inconsistencies that might indicate errors
- Assigns confidence scores to its validations
For healthcare payors, this means dramatically reducing manual review requirements. The agent can understand that John A. Smith and Smith, John Arthur likely refer to the same member, or that a birthdate of 01/02/1980 in one document and Jan. 2, 1980 in another represent the same information.
2. Compliance Check Agent
Healthcare enrollment is subject to complex, evolving regulations at both federal and state levels.
The Compliance Check Agent:
- Maintains awareness of current regulations including CMS requirements
- Applies appropriate compliance rules based on plan type and geography
- Identifies potential compliance risks before they become problems
- Suggests compliant alternatives when issues are detected
- Documents compliance decisions for audit purposes
For payors, this represents a significant advancement over traditional systems that require manual updates to compliance rules—who often struggle with regulatory interpretation.
3. Enrichment Agent
The Enrichment Agent elevates enrollment processing from only data capture to intelligent information enhancement:
- Cross-references external data sources for verification
- Completes missing information from available context
- Standardizes data formats for downstream systems
- Adds meta-information, such as risk factors or care program eligibility
- Identifies opportunities for member outreach or enhanced services
Payors gain the ability to create more complete member profiles at the point of enrollment, enabling better service delivery from day one.
4. Error Handling Agent
When exceptions occur, the Error Handling Agent provides intelligence far beyond traditional error handling:
- Analyzes root causes rather than just symptoms
- Applies contextual corrections based on historical patterns
- Provides detailed explanations for human reviewers
- Learns from resolution patterns to improve future processing
- Prioritizes exceptions based on business impact
This intelligent approach to error handling means payors can operate with smaller exception handling teams, while achieving higher accuracy and member satisfaction.
Intelligent self-correction
Understanding confidence scoring
One of the most powerful features of agentic AI is its ability to assess its own certainty—something traditional systems cannot do. Instead of binary pass or fail decisions, agents assign confidence scores that reflect how certain they are about their conclusions. Agents consider multiple factors when assigning confidence.
Example: Name matching scenario:
- Submission A:
- Form 1: “John Smith”
- Form 2: “John Smith”
- Confidence: 99%
- Submission B:
- Form 1: “J. Smith”
- Form 2: “John Smith”
- Confidence: 78%
The system employs confidence scoring for automated corrections:
- High-confidence corrections (less than 95%) proceed automatically
- Medium-confidence corrections (75-95%) are flagged for quick review
- Low-confidence items (greater than 75%) route to specialized reviewers
In our name matching example, why did Submission B receive a medium, 78% confidence, score? There are several factors that the agent is considering. While the last names do match, and the J likely stands for John, the J could also stand for James, Joseph, Jennifer, and so on. It is probable that they are the same person, but there is meaningful uncertainty, so additional context is needed for absolute certainty. This tiered approach dramatically reduces manual effort, while maintaining accuracy.
Regulatory adaptability
Healthcare regulations evolve constantly. Our solution’s architecture uses a flexible rule engine that can:
- Rapidly incorporate new CMS mandates
- Adapt to state-specific requirements
- Update compliance checks without code changes
The solution architecture uses a fundamentally different approach that separates regulatory knowledge from processing logic. Instead of hard-coding regulations into the system, the Compliance Agent accesses regulatory knowledge bases that can be updated independently of the processing code. Think of the traditional system as regulations baked into the recipe. To change the recipe, you must re-write and republish the cookbook.
An agentic AI system on other hand contains regulations in a reference guide. Every time an update is made to the reference guide, the agent immediately uses the new information.
How it works: A step-by-step example
Consider the following scenario of a new CMS guidance on dependent coverage. On February 1, 2025, a new CMS clarifies that stepchildren are explicitly covered under all ACA Marketplace plans, with updated documentation requirements.
Day 1: Regulatory update published
- A new CMS guidance update detected:
- Source: CMS.gov regulatory updates feed
- Topic: Dependent eligibility of stepchildren
- Effective date: Immediate
- Impact: ACA Marketplace plans
- Automated knowledge base update:
- New guidance ingested into regulatory knowledge base
- Relevant sections tagged: dependent eligibility, ACA Marketplace, stepchildren
- Previous guidance marked as superseded
- Change summary generated for compliance team review
Day 2: Compliance team reviews
- Compliance team notification:
- New CMS guidance detected and staged for activation
- Review process:
- Compliance officer reviews change summary
- Confirms interpretation is correct
- Approves activation
- System updates knowledge base status: ACTIVE
- Time required: 30 minutes
Day 2 (Afternoon): Compliance agent begins using new guidance
- Enrollment submission received:
- Dependent relationship: Stepchild
- Plan type: ACA Marketplace
- Compliance agent processing:
- Queries regulatory knowledge base
- Retrieves current guidance (updated this morning)
- Applies new requirements:
- Stepchildren explicitly allowed
- Documentation requirements: Birth certificate + custody documentation
- No additional restrictions
- Checks submission:
- Birth certificate provided
- Custody documentation provided
- Applies new requirements:
- Decision: APPROVED
- Confidence: 97%
Previously, this would have required manual review and taken potentially longer than two days.
Handling PHI and PII data
The solution incorporates comprehensive safeguards for protected health information (PHI) and personally identifiable information (PII) in compliance with HIPAA regulations. It exclusively leverages AWS services that are HIPAA-eligible, these services process, store, and transmit PHI securely in accordance with Business Associate Agreements (BAA). All data is encrypted both in transit and at rest using AWS encryption services, with strict access controls implemented through AWS Identity and Access Management (IAM) and fine-grained permissions.
The solution maintains detailed audit trails of all PHI and PII access and processing activities for compliance reporting. Data tokenization and masking techniques are employed for sensitive information when full data access is unnecessary.
The AI agents are designed to process only the minimum necessary PHI and PII required for enrollment functions, adhering to data minimization principles. Additionally, the solution supports configurable data retention policies to verify PHI and PII are not stored longer than required by regulations or business needs. These comprehensive security measures confirm that member data remains protected throughout the enrollment process, while maintaining the efficiency benefits of the AI-driven architecture.
Benefits
A key architectural benefit is the ability to scale elastically during peak enrollment periods, maintaining consistent performance even under high volumes. The solution’s automatic regulatory updates enhance compliance and help mitigate risks in an ever-changing healthcare regulatory landscape. Furthermore, the solution improves member satisfaction through expedited onboarding and reduced enrollment errors, while payors gain deeper insights into their member population from the initial enrollment. It facilitates personalized service delivery from day one.
Conclusion
The healthcare payor member enrollment solution demonstrates how the AI services of AWS, combined with specialized agents, can transform complex, document-heavy processes in highly regulated industries. By leveraging the foundation models of Amazon Bedrock, Amazon Bedrock AgentCore for orchestration, and AWS B2B for EDI processing, healthcare payors can build enrollment systems that scale elastically during peak periods, while maintaining compliance and improving the member experience.
This solution represents a shift from batch-oriented, rules-based systems to intelligent, agentic processing that adapts to changing conditions and continuously improves through feedback loops. As healthcare payors prepare for future enrollment periods, this approach offers a clear path to operational excellence and enhanced member satisfaction.
To learn more about implementing this architecture in your organization, contact your AWS account team or visit AWS for Healthcare & Life Sciences. You can also contact an AWS Representative to learn how we can help accelerate your business.