By Matthew Holt
It wasn’t that long ago that I could create one of those digital health or health tech maps and actually get it right. I used to do it myself back in the Health 2.0 days, with categories like “The Rebel Alliance of New Provider Tech” and “The Frontier of Patient Empowerment Technologies.”
But the days when you could neatly match a SaaS product to its target user and differentiate it from others are gone. The whole map has been upended by the hurricane that is generative AI, leaving the industry in a state of confusion.
Over the last few months, I’ve been trying to figure out who is going to do what in AI health tech. I’ve had lots of formal and informal conversations, read a ton, and attended three conferences—all focused on this subject. It’s clear: nobody has a solid answer.
That hasn’t stopped people from trying to map it out—like this example from Protege. As you can see, there are hundreds of companies building first-generation AI products for every part of the healthcare value chain (or its lack thereof).
But this time is different. It’s not clear that AI will stop at a user interface or even have a clearly defined function. It’s not even clear that there will be a “health tech AI sector” at all.
A Multidimensional Problem
The major LLMs—ChatGPT (OpenAI/Microsoft), Gemini (Google/Alphabet), Claude (Anthropic/Amazon), Grok (X/Twitter), and LLaMA (Meta/Facebook)—are all capable of incredible work in healthcare and, of course, beyond it. They can now write in any language, code, create videos, music, and images—and they keep getting better.
They are outstanding at interpretation and summarization. I literally threw a dense 26-page CMS RFI document into ChatGPT recently, and in seconds it told me what they were asking and what they were really looking for (the subtext). The CMS official who authored the document was very impressed—and a little annoyed that they weren’t allowed to use it. If I wanted to help CMS, I could have written the response for them too.
These major LLMs are also developing “agent” capabilities—meaning they can execute multi-step business or human processes.
Right now, they are being used directly by healthcare professionals and patients for summarizing, communication, and even companionship. They’re increasingly being used for diagnosis, coaching, and therapy. And of course, many healthcare organizations are using them to redesign workflows.
The EHR Battlefield
The main battleground of healthcare AI is still the EHRs (Electronic Health Records) used by providers, with Epic being the biggest of them all. Epic has a relationship with Microsoft, which has its own AI game, and also a major stake in OpenAI (investing $13 billion in the nonprofit). Epic now uses Microsoft’s AI for note summarization, communications, and also Nuance’s DAX (an AI scribe environment). Epic is also partnered with DAX rival Abridge.
But Epic isn’t stopping there—it’s clearly building its own AI capabilities. In a great Healthcare IT Today piece, John Lee describes Epic’s non-trivial integration of AI into clinical workflows:
- Epic now offers tools to reorganize text for readability, generate patient-friendly summaries, hospital course summaries, discharge instructions, and narrative translations of clinical data.
- It can automatically replace stigmatizing language in notes (e.g., “narcotic abuser” becomes “patient has opioid use disorder”).
- Even physicians sometimes struggle to understand shorthand. Epic showed how its AI could translate something like “Pod 1 s/p CABG. HD stable. Amb w/ assist” into plain English: “Day 1 post-coronary bypass graft. Hemodynamically stable. Ambulating with assistance.”
- For nurses, AI-generated environmental documentation and shift notes will reduce manual input and free up time for patient care.
Of course, Epic isn’t the only EHR in town. Meditech is also in the race. In a wide-ranging HIStalk interview, Meditech’s COO Helen Waters discussed their work with Google and AI, highlighting:
- Their initial product was based on Google’s BERT language model, with “Condition Explorer” offering intelligent insights from the patient chart and presenting a longitudinal view of the patient’s health.
- With Vertex AI and newer Gemini models, Meditech now generates hospital stay narratives and nursing shift summaries automatically.
- They’ve introduced ambient scribes using multiple provider platforms.
- The Google partnership remains strong, and the results are clear with their vision for Expanse Navigator. The goal is to use AI not just to reduce burnout and automate administrative tasks but to enhance clinical confidence.
Voice AI is becoming the next frontier. Ironically, we’ve come full circle—from speech devices to typing on tablets and phones, and now back to voice. But this time, it may actually work well and efficiently for clinicians.
AI is Everywhere—But Who Owns It?
Both Epic, which dominates academic medical centers and large non-profits, and Meditech, used by most large for-profit systems like HCA, are embedding AI into clinical and administrative workflows.
I raised this issue during a meeting hosted by Commure—a general provider-focused AI company backed by General Catalyst. Commure has gone through several iterations but is now an AI platform offering products for admin, revenue cycle, inventory and staffing tracking, ambient listening/scribing, clinical workflow, and summarization. With HCA as both an investor and major Meditech customer, there’s a strategic dilemma: HCA must decide where to place its AI bets—Commure or Meditech.
Meanwhile, academic medical centers (AMCs), providers, payers, and plans are all using big LLMs internally. Many bought external tools like Epic, but are now customizing AI to their workflows. Some are even building their own: Stanford, for instance, developed a homegrown tool using Anthropic’s Claude to communicate lab results to patients. These are often PhD or thesis projects happening across major healthcare institutions—everyone needs something for their data scientists to do!
So What Does This Mean?
- EHRs see themselves as massive data stores and expect AI tools to take over current workflows.
- Tech trends show consolidation. Just as WordPerfect, Lotus 1-2-3, and Persuasion became Microsoft Office and Google Suite, many niche AI tools will become features in bigger products. For example, Brellium recently raised $16 million to summarize and analyze clinical notes for compliance. But doesn’t everyone already offer AI note summarization? It’s likely this will be a feature in a broader product soon.
(Note: One potential exception is conversational AI—interpreting human speech and dialog is hard, and may stay separate for longer.)
Vince Kuraitis, Girish Muralidharan, and the late Jody Ranck recently published a 3-part series suggesting EHRs will evolve into unified digital health platforms. Clinical data will integrate with systems managing staff, supplies, finance, marketing, etc.—and with medical devices across hospitals and the broader health ecosystem.
This integration could lead to AI-dominated “super-systems” making many decisions automatically, such as following care protocols or optimizing staff scheduling. The move toward deep research and agent-based AI from big LLMs has even led figures like Satya Nadella to suggest that SaaS is dead.
It’s not hard to imagine a future where AI scrapes everything and runs healthcare systems through autonomous agents.
The Real Problem for Everyone
- If you’re buying AI: You don’t know if the tool will soon be cannibalized by your EHR or already exists internally.
- If you’re selling AI: You don’t know if your product is just a feature of someone else’s platform, or if your client already built it in-house. Worse, there’s little penalty for customers waiting for something better and cheaper.
And all of this is happening while new LLMs and AI models are released every few months.
Final Thoughts
For now, the problem is that—until we have a clearer sense of how this will all play out—there will be many false starts, funding rounds that go nowhere, and AI deployments that don’t deliver. Reports like the one by Sofia Guerra and Steve Kraus at Bessemer, identifying 59 “Jobs to be Done,” can help. But I worry no one really knows what tool is best for which job.