Platforms
A Word That Explains Healthcare AI’s Future
Many people – certainly everyone from my generation – will remember the iconic scene in the 1967 movie, The Graduate, in which a befuddled Benjamin Braddock (played by Dustin Hoffman) receives unbidden career advice from one of his parents’ friends. “I just want to say one word to you,” says the man, his voice hushed as if he were spilling nuclear codes. “Plastics. There’s a great future in plastics. Think about it.”
As I survey the current healthcare AI ecosystem, my one word is “Platforms.” In recent months, we’ve witnessed the start of a big shakeout, as companies that initially pitched their AI tools as solving single problems have recognized the imperative to become a one-stop shopping platform that addresses multiple problems for health systems and clinicians. This shift carries enormous implications for healthcare AI’s future.
From Scribes to Platforms: The Great Pivot
The transformation is most visible among companies – including Abridge, Ambience, Augmedix, Nabla, Nuance, and Suki – that began by tackling clinical documentation through AI scribes. Their choice to start with scribes wasn’t accidental – it reflected hard-won lessons from healthcare AI’s troubled past.
In the 1970s and 80s, when AI first became viable (though far less powerful than today’s generative AI), most healthcare AI start-ups focused on a different problem: diagnosis. In retrospect, this was a fatally flawed decision – when you’re trying to gain buy-in from skeptical end-users, tackling the hardest and highest risk problem is the last place you should start. “We were not naïve about the complexity,” Stanford’s Larry Fagan, a leading early AI researcher, told me. “It’s just that it was the most exciting question.” The problem was that the tools were not only hard to use; they sometimes produced absurd outputs. The result was that AI was never embraced by clinicians or health systems, and the AI companies of the day all went belly up, ushering in a 40-year period known as healthcare’s “AI Winter.”
With the renewed interest in AI catalyzed by the advent of gen AI, developers, investors, and entrepreneurs rushed back into healthcare, drawn both by our field’s human importance and the fact that it accounts for nearly 20 percent of U.S. GDP. Having absorbed the historical lessons, the new generation of AI entrepreneurs concentrated on solving important but low-risk problems for clinicians. Win doctors’ hearts and minds first, the thinking went, then pursue more ambitious (and profitable) solutions. AI scribes fit the bill perfectly.
At UCSF Health, where I work, we now offer AI scribes to all 2,000 of our outpatient physicians. Studies demonstrate high levels of satisfaction on the part of both clinicians and patients, accompanied by modest productivity gains. While those gains have been a bit disappointing, the tools have nevertheless hit a tipping point. A few years ago, virtually no one used AI scribes (in 2022, UCSF Health employed dozens of human scribes, mostly pre-med students on a gap year). I predict that, by next year, outpatient practices everywhere will struggle to recruit and retain doctors without offering them a digital documentation wingman.
The Commodity Trap
With that kind of adoption curve, this should be break-out-the-champagne time at the AI scribe companies. But something surprising happened: AI scribes (now rebranded as “ambient clinical intelligence”) evolved from gee-whiz novelties into ho-hum commodities. Bryan Roberts, a leading healthcare investor at Venrock, predicts that “scribes will not be a defensible moat… they’re going to be a dime a dozen.” Sara Murray, UCSF’s chief health AI officer, told me, “As this technology becomes a commodity and performance differences between vendors narrow, pricing and EHR integration will likely drive decision-making.”
The scribe companies see the writing on the wall. Suki’s ads proclaim, “More than just an AI scribe; an AI assistant.” Augmedix was acquired by VC General Catalyst’s platform company Commure1 to serve as one tool in a broader suite of AI products. Abridge, the space’s leading start-up, has raised $800 million (including $300 million last month), boosting its valuation to above $5 billion. The company, like all the others, plans to tackle more ambitious use cases. “It’s so exciting right now because we’re seeing overnight impact,” Abridge CEO Shiv Rao told me. “But over time we’ll expand into those other [areas].”
And the commoditization of scribes hit its natural culmination last week when Doximity, a kind of LinkedIn for health professionals, announced the launch of its own scribe product. The cost of using Doximity’s scribe is, you guessed it, nada.2
The Real Prize: AI Clinical Decision Support
The trend reflects a crucial insight. While AI tools like scribes, coding assistants, prior authorization generators, and call center chatbots are useful, the real action in healthcare involves clinical decisions made by doctors and other non-physician providers like NPs. There’s an old adage that healthcare’s most expensive piece of technology is the doctor’s pen, since approximately 80 percent of healthcare costs flow from a clinician’s decisions. The electronic health record may have turned that pen into a keyboard, but the point remains: truly game-changing clinical and financial impact from AI will come through influencing the billions of diagnostic and therapeutic decisions clinicians make annually.
Given where the money lies, you’d think healthcare AI start-ups, VCs, and tech giants (Google, OpenAI, Microsoft, and the like) would target diagnosis and clinical decision-support directly – building AI tools to guide which tests, medications, and procedures should be offered to patients. But they’re too smart for that. They know the history of healthcare’s AI winter, the ongoing challenges with alert fatigue, physicians’ allergic reactions to “cookbook medicine,” and Epic’s problematic 2021 rollout of a not-ready-for-prime-time sepsis decision support tool. Moreover, nobody wants to deal with liability issues from flawed AI recommendations. “Fail fast and iterate” works fine for restaurant apps, but in healthcare, it can be deadly. All of this makes clinical decision-support too risky to start with – but too important to ignore.
OpenEvidence: Attacking the Platform Imperative from Another Angle
While scribe companies try to broaden their offerings into becoming patient-care platforms, I was also struck this month by a new $210 million investment in OpenEvidence, which places the current value of the company at $3.5 billion. I hadn’t heard of the company 18 months ago; it’s now become the go-to AI question-answering resource for many clinicians, including our UCSF internal medicine residents. Trained on the medical literature and used primarily by doctors, its output and recommendations tend to be more sophisticated, better referenced, and, in my experience, more accurate than those delivered by general-purpose AI tools like GPT, Gemini, or Claude.
Although OpenEvidence may be content with becoming this era’s UpToDate (a digital knowledge base that became ubiquitous, seemingly overnight, about 30 years ago), I suspect that the company has grander ambitions: integration into the all-important Platform of Platforms, the electronic health record. Picture this: after AI agents review and summarize a patient’s past record and create the clinical documentation for today’s visit, OpenEvidence delivers real-time diagnostic and clinical decision support via the EHR. This vision no longer seems far-fetched, and it’s doubtless responsible for some portion of the company’s sky-high valuation. (UpToDate isn’t conceding – in October 2024, it signed a partnership agreement with Abridge, clearly with the same playbook in mind.)
Epic’s Platform Advantage
As scores of start-ups compete to become health systems’ platform of choice, one company already occupies the pole position: Epic. As America’s leading EHR vendor, the Wisconsin behemoth has several natural advantages when it comes to convincing health systems that they should use Epic’s AI products rather than straying into the messy world of start-up tools. First, there’s certainty that Epic will be around in five years, something no start-up can guarantee. Second, while integration of third-party tools into EHRs has improved, it still requires time and money. Epic can promise tools that are integrated by design into its tech stack and workflow.
But Epic faces a problem when it comes to AI: some of its early AI tools (like that problematic sepsis decision support algorithm and its EHR in-box answering assistant) were not very good. This shouldn’t be too surprising – despite enormous recent investments in AI, Epic is primarily an EHR company. A start-up whose sole focus is building AI for specific use-cases should have advantages over an EHR company creating dozens of AI tools spanning scribing to billing to clinical trials matching.
Yet as start-ups try to innovate their way into health systems, they often hit a wall: Many systems prefer to stick with tools built by their EHR vendor. A 2025 survey showed that roughly four in five U.S. hospitals using AI predictive models were using ones supplied by their EHR vendor. Health systems favor Epic’s AI offerings for reasons beyond economics, integration, and codependency: Most healthcare IT leaders don’t really want to deal with 120 decibels of “innovation noise.” As one CIO lamented, “I have fifty vendors reaching out a week. Which one do I listen to?” Choosing Epic for AI needs has become the safe choice, echoing the maxim about a colossus from a previous era: “Nobody ever gets fired for buying IBM.”
Epic recently introduced a new pricing option that, while helping its clients employ the company’s AI tools at scale, also discourages marketplace wandering. Beyond its traditional a la carte model, health systems can now pay a single fee to access Epic’s entire AI portfolio. One health system IT executive described Epic’s approach as like the kid who “licks all the cookies” before friends arrive at her birthday party.
Until now, Epic has steered clear of building an AI scribe itself, instead inking partnership agreements with Abridge and Nuance/DAX. Perhaps unsurprisingly, just yesterday, Politico reported that Epic will roll out its own scribing tool later this month. And with that, yet another cookie is licked.
The Bottom Line
While it’s great to see companies tackling each of healthcare’s myriad problems – scribing, billing, reading x-rays, assisting with surgeries – keep your eye on the quest to become healthcare’s AI platform of choice. Epic is sure to have a seat at the winner’s table, since the company already controls so many of the chips. For many start-ups, the focus will first be on getting through the door of health systems by successfully delivering a solution to a single important problem, such as clinical documentation or billing and coding, then branching out to become the platform of choice for addressing an array of problems, ultimately including clinical decision support.
As Julia Adler-Milstein, the AI expert who leads UCSF’s Division of Clinical Informatics and Digital Transformation (DoC-IT), told me, “It feels like the market is rewarding the unit of scale that is general enough to apply to a broad group of specialties, but also solves a very well-defined problem so that you can put it into a budget somewhere.” Her observation captures the current market dynamics beautifully: Start-ups must build trust with focused products, then rapidly pivot to creating platform-like offerings, either independently or through strategic partnerships.
All of this will shake out in the next few years, and the dynamics will have little to do with whether OpenAI or Gemini achieves superintelligence, how many GPUs are in the Abilene, Texas data center, whether the U.S. embraces or shuns AI regulation, or whether AI is “woke.” While the titanic battles in Washington, D.C. and Silicon Valley will influence healthcare’s direction, the overall shape of our AI transformation will be determined largely by inside baseball: the clinical and business dynamics of the healthcare field itself.
I’ll be back next month to recap key developments from August. If you’d like to go deeper, you can pre-order my book on healthcare AI, A Giant Leap (Feb 3, 2026).
I serve as an advisor to Commure.
Doximity’s AI scribe doesn’t integrate with the EHR, though, which limits its attractiveness.