United StatesHealthcare and life sciences2025 – ongoingMyriad Genetics
Image Genius: generative AI document classification for prior authorization
Myriad Genetics cut document classification cost by 77% after moving from Textract and Comprehend to Nova models on Amazon Bedrock
Problem
Myriad Genetics handles thousands of medical documents a day to support prior authorization. Its first classification system, built on Amazon Textract and Amazon Comprehend, cost about 3 cents a page and took 8.5 minutes per document. Staff extracted key data from forms by hand because the layouts varied.
Approach
Myriad and AWS rebuilt the pipeline on the open-source GenAI Intelligent Document Processing Accelerator. Amazon Nova Pro classified documents, with prompts refined from document samples and negative examples. Amazon Nova Premier extracted data, including checkboxes. The output went into Myriad's authorization system through an event-driven pipeline.
Results
Takeaway
The first build proved the use case but was too costly to keep. Switching to a cheaper model and tuning the prompts changed the cost per page.
Categories
- Industry
- Healthcare and life sciences
- Business function
- Production and quality
- Innovation type
- Process innovation
- AI technology
- Generative AI
- Adoption
- Partner co-development
- Evidence
- Vendor case study
Sources
Source summaries
AWS and two Myriad engineers describe moving document classification and extraction from Textract and Comprehend to Bedrock with Nova models. They report accuracy of 98% against 94% before, a 77% cost cut and classification time of 1.5 minutes against 8.5. The post does not describe its measurement method in full.
Full details are in the original sources.