How Cohere Health digitizes clinical policies using Amazon Bedrock AgentCore
AWS Oleksiy Kononenko
Cohere Health used AWS to digitize messy insurance-policy PDFs into structured data. That cut policy work time 30% and deployment from months to weeks.
Based on reporting by AWS, Oleksiy Kononenko — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Prior authorization is the paperwork health insurers demand before they'll pay for certain treatments, and it has stayed stubbornly manual for reasons that have nothing to do with medicine. The actual clinical logic behind these approvals is usually sound. The problem is that the rules governing them live in PDFs and scanned documents that nobody built to be read by machines. Multiply that across hundreds of health plans, dozens of clinical categories, and constant regulatory churn, and you get a bottleneck that touches hundreds of millions of patients a year.
Cohere Health decided the fix wasn't a smarter chatbot but better plumbing. The company built Cohere Policy Studio on top of Amazon Bedrock AgentCore, leaning on its microVM-based session isolation to keep different health-plan customers' data walled off from each other, its Gateway to unify access to internal APIs and Lambda-based tools, and its Memory service to preserve context across the back-and-forth that policy analysts need when correcting an agent's output. The team wired this into LangChain and added a layer called Agent Skills — modular, versioned bundles of domain expertise that let clinical policy experts encode their own knowledge without touching the underlying agent infrastructure.
The engineering choices are pragmatic rather than flashy. A two-tier Docker image setup lets every team inherit a shared base runtime and just drop in a YAML config for their own memory mode, caching strategy, and model settings. Skills get dual-layer versioning — semantic tags in git alongside S3 object versioning — so nothing ships to production without a paper trail and a rollback path. Evaluation runs through Arize AI, with human clinical analysts spot-checking outputs that automated metrics might miss, which keeps the human-in-the-loop promise from being just a slide in a deck.
The payoff shows up in the numbers Cohere Health is willing to publish. Per-policy digitization time dropped from two hours fifteen minutes to one hour thirty-five, a 30 percent cut, and that's before the agent framework has scaled across the company's full policy library. More striking is deployment velocity: standing up a new agent used to take three to four months and now takes two to six weeks, largely because reusable base images and modular skills mean teams aren't rebuilding runtime infrastructure every time they need a new capability.
None of this happens in a vacuum. CMS is requiring health plans to support API-based electronic prior authorization by January 2027, and AHIP wants 80 percent of electronic submissions approved in real time. Those deadlines turn what used to be an internal efficiency project into a compliance race, and Cohere Health's bet is that infrastructure built for multi-tenant isolation and fast iteration will matter more over the next two years than any single model upgrade.
My take — AI-written commentary, not fact-checked reporting
Healthcare has treated PDFs like scripture for decades, so it's almost funny that the fix here isn't a bigger model but boring infrastructure — gateways, versioned S3 buckets, isolated microVMs. That's the actual lesson: the industry's AI bottleneck was never reasoning power, it was that nobody bothered to make the rulebook machine-readable. Keeping clinicians in the loop for judgment calls while automating the grunt work of digitization is the correct split, and more health-tech vendors chasing flashy diagnosis bots should notice that the unglamorous plumbing project is the one actually shipping results before the 2027 deadline forces everyone's hand.
Read more about this at: AWS