AWS Podcast · Amazon Web Services

#754: Accelerating healthcare decisions with agents

April 7, 2026·36 min·2 clips

Gigi Yuen reveals how 20-30% of healthcare spending goes to administrative tasks, with half being low-value waste.

1. AWS Podcast episode 754 features Cohere Health's Gigi Yuen and Kenji Fujita discussing healthcare AI agents in production, released March 20, 2026. 2. Host Jillian Ford interviews Gigi Yuen (Chief Data and AI Officer) and Kenji Fujita (Staff AI Platform Engineer) at Cohere Health, a clinical intelligence company operating at the intersection of payers and providers. 3. The episode's thesis is that deploying AI agents in a regulated, high-stakes domain requires evaluation-driven development, domain expert involvement from day one, and clear ethical boundaries around automation. 4. Cohere Health's core business problem is the estimated $500 billion in low-value healthcare administrative costs — roughly half of the 20-30% of healthcare spending (cited from Health Affairs) consumed by administrative tasks between payers and providers. 5. Gigi frames trust and transparency as the most critical design constraints, noting there is 'no tolerance for hallucination' in clinical settings and that it took Cohere one month to iteratively define how to quantify hallucination in their specific context. 6. Cohere Health embeds licensed clinicians into every single development project from the start rather than bringing them in for end-stage validation, arguing this changes the quality of metrics being measured. 7. The company has drawn a firm ethical line: AI will never be used to deny a patient's care or a provider's request; any case that cannot result in automatic approval is routed to a specialist of the same clinical specialty. 8. This line is operationalized with risk tiers: a diagnostic imaging prior authorization can be auto-approved if no contraindications exist, the patient is covered, and policy criteria are met, while a surgical authorization always requires a human confident enough to authorize an invasive procedure. 9. On domain-specific data, Gigi distinguishes between incorporating knowledge systems (medical society guidelines, standard of care ontologies) for consistency versus accessing case-specific individual notes for contextual accuracy — and emphasizes data use rights as a non-negotiable constraint. 10. Kenji Fujita explains that Cohere Health chose Amazon Bedrock Agent Core because of its speed of innovation, built-in tenancy separation for memory clients (critical in HIPAA-regulated environments), and the ability to implement MCP server targets with a few lines of code rather than building auth proxies from scratch. 11. Cohere had already started building its own MCP servers with custom auth proxies; Agent Core gateway replaced that custom work with an identity-layer configuration that matched their existing authentication patterns. 12. After adopting Bedrock Agent Core, Cohere's planned agent count for Q1-Q4 2026 went from one or two to a full pipeline of agents and multi-agent systems across clinical use cases. 13. Cohere generates 50,000-60,000 labels per day from its licensed clinicians to build ground truth data sets, which it uses alongside an automated LLM-as-judge approach for model evaluation. 14. On model selection, Gigi describes maintaining a private leaderboard tracking accuracy, cost, latency, reliability, and eval data requirements across model options — noting that as ground truth data accumulates, the optimal choice for each use case changes. 15. Concrete results from the prior authorization system: 85% of decisions are now automated within minutes, and for the 15% requiring human review, agentic systems have improved clinician productivity by 30-40%. 16. Clinicians report improved job satisfaction because the agent surfaces all relevant information and clinical criteria, allowing them to focus on judgment rather than information retrieval. 17. Gigi cites McKinsey data that 95% of AI POCs never reach production and only half of those that do stay in production, framing evaluation-driven development as the single most important practice to close that gap. 18. Her second success factor is defining 'what must be true for the agent to succeed at scale' before launch — she notes these criteria are usually about people and process (training, transition plans, advocates) rather than AI technology. 19. Her third factor is data integration, arguing teams consistently underestimate integration complexity and need to confirm that people, process, and data are all ready before deploying. 20. The episode closes with a debate about whether agentic development should be centralized in a specialist team or federated across all development teams — both guests describe the answer as still unresolved.

As heard by us

A practical look at healthcare agents, with security and production pressure front and center.

AWS Podcast frames healthcare agent adoption as a practical problem, not a headline. The discussion works best when it stays on security, memory, and the pressure of putting agents into production, especially in Cohere Health's account of building with Bedrock Agent Core.

Read the full review in PlayNext →

Why you'd press play

Want a real-world agent story from healthcare, not a hype reel?

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from AWS Podcast