I mentioned in My AI Journey So Far that I crammed through a lot of AI courses in early 2024. Recently, I enrolled in an entry-level course—not for myself, but to evaluate whether I could recommend it to others who aren’t as far along in their AI journey. These courses get updated constantly, so even an introductory one can contain current material worth passing along.
While taking the course, something caught my attention.
What Is Med-PaLM?
The course mentioned Med-PaLM, a large language model developed by Google Research specifically for medical applications. I’d never heard of it. I spend most of my time in AWS land, so Google’s healthcare AI offerings weren’t on my radar.
Med-PaLM builds on Google’s PaLM language model architecture but fine-tuned for medical knowledge. It combines medical comprehension, knowledge retrieval, and reasoning to process complex clinical scenarios. The second version, Med-PaLM 2, reached 86.5% accuracy on the MedQA medical exam benchmark—the first AI system to surpass the USMLE passing threshold and reach human expert-level performance on standardized medical exams.
That got my attention.
Med-PaLM 2 now powers MedLM, a family of foundation models available to Google Cloud customers. There are two models under MedLM: a larger one designed for complex tasks, and a medium one that can be fine-tuned for scaling across different use cases. HCA Healthcare has been piloting MedLM with Augmedix to help physicians with medical notes in emergency departments—physicians use a hands-free device to capture conversations, and the platform converts them into draft notes.
There’s an important caveat. MedLM has not been designed to be used as a medical device. Any output should be verified by a healthcare professional, and no direct diagnosis should be made. It’s currently available only to allow-listed customers in the U.S., so access requires contacting a Google Cloud Customer Engineer.
The Competitive Landscape
Since I work in healthcare IT but currently live in the AWS ecosystem, I wanted to understand what alternatives exist. Turns out there’s a competitive marketplace.
AWS HealthScribe
AWS HealthScribe is Amazon’s answer to clinical documentation. It’s powered by Amazon Bedrock—which includes Claude as one of its foundation models—and uses speech recognition and generative AI to automatically generate preliminary clinical documentation.
Using a single API, HealthScribe identifies speaker roles, classifies dialogues, extracts medical terms, and generates transcripts and notes. Every sentence in the AI-generated note includes references to the original transcript, which addresses the explainability concern that’s top of mind for anyone deploying AI in clinical settings.
HealthScribe is HIPAA-eligible and doesn’t retain inbound audio or output text. It also doesn’t use your data to train AI models. Pricing is pay-as-you-go: $0.001667 per audio second, which works out to about $1.50 for a 15-minute consultation.
Microsoft Dragon Copilot
On the Microsoft side, Dragon Copilot—developed by Nuance, which Microsoft acquired—is embedded directly in Epic. It uses ambient listening technology to capture entire clinical encounters and generate structured notes. It’s trained on over 10 million patient-doctor interactions and connects with over 200 hospital systems.
Clinicians report draft notes appearing within 2-3 minutes after a visit. Studies show physicians using AI scribes save 2-3 hours daily on documentation and see 15% more patients per hour. Burnout decreased from 51.9% to 38.8% after just 30 days of use in one study of 263 physicians across six health systems.
The tradeoff is cost. Dragon Copilot runs $600-800 per month per provider with 1-3 year contracts. It’s built for hospital-scale deployments, not specialty practices.
Three Approaches, Three Philosophies
What’s interesting is how these three offerings reflect different strategic philosophies:
Google MedLM is a foundational model—a building block for healthcare applications rather than a finished product. It’s designed for developers and organizations that want to build custom solutions. Access is restricted and requires a relationship with Google Cloud.
AWS HealthScribe takes the infrastructure approach AWS is known for: a HIPAA-eligible API with transparent pricing, no data retention, and flexibility to integrate with whatever you’re building. It’s designed for healthcare software vendors and organizations that want to build their own clinical applications.
Microsoft Dragon Copilot is a finished product, deeply integrated into Epic and designed for hospital-scale deployment. It’s the most expensive but also the most turnkey for large health systems that want to reduce clinician burnout without building anything themselves.
What This Means for Healthcare IT
If you’re in healthcare IT, keeping up with this space requires ongoing attention. It’s moving fast. The AI medical scribe market has matured dramatically over the past two years. What started with Nuance’s early work has evolved into a competitive marketplace with solutions ranging from enterprise-grade systems to mobile-first apps.
There’s no single “best” solution. The right choice depends on your practice size, EHR environment, budget constraints, and specialty needs. But the productivity gains—hours saved daily, more patients seen, reduced burnout—are real and documented.
As for whether I’d recommend the course to someone just getting started? Check back here periodically to find out. The landscape shifts constantly, and even experienced practitioners can find something hiding just outside their field of view.