EXL outlines AWS-based AI pipeline for medical claims review

EXL has detailed an AWS-based AI pipeline for medical claims review that combines document processing with a domain-specific insurance model and general-purpose foundation models. Designed for adjusters, underwriters and care coordinators, the system turns medical records into extracted fields, summaries, answers and structured files, while sending low-confidence results to human reviewers.

Medical IDP brings together two EXL applications. Xtrakto.AI handles ingestion, document splitting, classification, extraction, enrichment and postprocessing. EXL Insurance LLM handles medical summarization, natural-language queries, traced reasoning and output generation. EXL serves its domain model through Amazon SageMaker AI, while models available through Amazon Bedrock handle broader language and reasoning tasks.

AWS and EXL describe an 11-step regional architecture. API Gateway accepts requests and returns results, Cognito authenticates them, and Step Functions coordinates the workflow. Textract and Lambda prepare the documents; SageMaker endpoints run extraction and domain-model inference; and DynamoDB and RDS support enrichment. Bedrock supplies general-purpose models, while Lambda, CloudWatch and S3 handle output generation, monitoring and storage. A separate SageMaker environment is reserved for model development.

The intended review workflow spans seven stages: splitting and classifying documents, extraction, enrichment, summarization, querying, traced reasoning and output delivery. According to the companies, each extracted field receives a score from 0 to 100, and fields below a configurable threshold go to validators in Xtrakto.AI. Reviewers can trace summaries and answers to source passages, edit outputs and rate them. The companies say those corrections can inform later fine-tuning. Results can be delivered as PDF reports or as JSON, XML, CSV and flat-file exports.

EXL says its training set contained more than 13,500 structured and unstructured records collected across nine years of insurance claims operations. Its white paper names Llama 3.1 8B Instruct as the base model and describes supervised, parameter-efficient fine-tuning with LoRA and NVIDIA NeMo. EXL evaluated the model using automated text metrics and blind ratings from three insurance specialists, but described the comparison as an internal benchmark, not a public leaderboard or external academic study.

The reported performance and deployment outcomes remain vendor assertions. AWS and EXL say Medical IDP cut medical-record review from days to hours and cite manual work exceeding 100 minutes per case. They also say an unnamed healthcare payer achieved faster outreach, greater clinical capacity without added headcount, improved accuracy and better compliance. The published material does not identify the customer or provide a controlled baseline, sample size, field-level production accuracy or independent validation of those outcomes.

AWS lists the services in the design as eligible for use with protected health information, subject to a business associate agreement and customer configuration responsibilities. That eligibility does not establish compliance for EXL’s implementation. HHS recognizes Safe Harbor and Expert Determination as HIPAA de-identification methods, but the available sources do not establish that EXL’s process satisfies either method or has passed an external compliance assessment.