Something quietly enormous happened in healthcare this spring, and almost nobody outside the industry noticed. In a span of weeks, Mount Sinai, Cedars-Sinai, and more than a dozen other major US health systems stopped "experimenting" with AI and started signing system-wide enterprise contracts — the kind that put AI into the hands of every doctor, nurse, and pharmacist at once. Mount Sinai inked the first enterprise deal of its kind with a company called OpenEvidence on June 8.
Here's why I'm writing this for you, even if you've never set foot in a hospital boardroom: this is one of the clearest, most measurable examples of AI turning into real money anywhere in the economy. Not hype. Not a flashy demo. Hard dollars — a single 500-clinician health system can recover $25 to $37 million a year from one category of AI alone. Let me show you exactly how that math works, who's getting rich off it, and what it means for you whether you work in healthcare, invest, or just want to understand where AI actually pays.
From pilots to production: the dam just broke
For two years, hospitals treated AI the way a cautious swimmer treats cold water — one toe at a time. A pilot here, a small trial there, endless committees. That era just ended. In 2026, roughly 75% of US health systems are using or planning to use an enterprise AI platform, and they've shifted from isolated pilots to system-wide rollouts spanning clinical documentation, decision support, virtual care, and workforce management.
The reason for the sudden urgency is simple and it's the theme of this whole site: the ROI finally got undeniable. More than half of the health systems that could actually measure their AI returns reported at least 2x ROI, and most expect 200–400% return within 3–5 years. When the numbers get that clear, the cautious-swimmer approach stops being prudent and starts being expensive. Every quarter a system waits is money left on the table — the exact opposite of the Gartner finding that AI layoffs create budget room but not returns. Here, the returns are real because the AI is amplifying expensive clinicians, not replacing them.
The $37 million number, explained
Let me make the headline figure concrete, because it's the heart of the story.
The single biggest win in healthcare AI right now isn't some exotic diagnostic robot. It's boring, and that's why it works: ambient documentation — AI that listens to a doctor-patient conversation and writes the clinical note automatically, so the physician isn't typing into an EHR until 9 PM. Doctors call that after-hours typing "pajama time," and they hate it.
Here's the math that's making CFOs sign contracts:
| Metric | Value |
|---|---|
| Time an AI scribe saves per clinician | 60–90 min/day |
| Fully-loaded clinician cost | ~$250/hour |
| Recovered value per clinician | $50,000–$75,000/year |
| Across a 500-clinician system | $25M–$37M/year |
| Documentation hours saved (one platform, one year) | 15,700+ hours (≈1,794 working days) |
| Clinical note-taking adoption | 68% (+62% YoY) |
| Health systems reporting ≥2x ROI | More than half |
Sit with that table for a second. You take a tool that costs a system a few million a year, and it hands back $25–37 million in recovered clinician time — time those doctors can spend seeing more patients (more revenue) or simply not burning out (less turnover, and turnover is brutally expensive in medicine). This is the same "amplify the expensive human" pattern that drives every real AI return, the one I keep coming back to in which jobs AI actually reshapes. Nobody's firing the doctor. They're giving the doctor 90 minutes a day back, and that's worth a fortune.
The OpenEvidence rocket — and why it matters to you
The company at the center of this week's news is the clearest sign of how fast this money is moving. OpenEvidence is an AI clinical decision-support tool — think of it as a doctor's research assistant that answers natural-language medical questions with answers grounded in peer-reviewed literature, right inside the EHR.
Its growth is staggering. It's already used by roughly half of all US physicians, reaches 860,000+ clinicians, and its valuation rocketed from $3.5 billion in July to $12 billion in January — more than tripling in six months. That's not healthcare-slow. That's frontier-AI-fast, the same trajectory I traced in Anthropic's run to a $965 billion valuation.
And here's the detail that ties it all together for this site: OpenEvidence is reportedly building a non-ad-supported enterprise model "similar to Anthropic's" for big systems like Mount Sinai and Cedars-Sinai. The business model is converging on the same thing across all of AI — give the powerful tool broad distribution, then charge enterprises for the premium, private, compliant version. It's the exact playbook I broke down in how AI companies turn 'free' into billions, now playing out in hospitals.
Why this is happening now (the capability finally caught up)
You might wonder why 2026 specifically. The answer is that the underlying models finally got good enough — and cheap enough — to trust with clinical work.
A medical AI can't be charming but wrong; in healthcare, a confident hallucination can hurt someone. What changed is that the frontier models crossed the reliability threshold where grounded, cited, low-hallucination answers became dependable. I wrote about exactly this leap in the Claude Fable 5 launch — the same model family is accelerating drug design 10x, and the same reliability gains are what make a tool safe to put in front of 860,000 clinicians. And because frontier model prices keep falling, running these tools across an entire hospital system finally pencils out. Capability went up, cost came down, and the dam broke.
The risks nobody puts in the press release
I'd be selling you a fairy tale if I left it at "$37 million, sign here." Healthcare AI has real failure modes, and the smart systems are pricing them in.
The first is automation bias — the danger that an overworked clinician starts rubber-stamping whatever the AI suggests instead of thinking. A tool that's right 95% of the time can be more dangerous than one that's right 70%, because people stop checking. That's exactly why the serious deployments keep a human in the loop and ground every answer in cited, peer-reviewed literature rather than letting the model free-associate.
The second is the measurement trap. That 2x ROI is real for systems that picked a narrow, measurable win like documentation. But plenty of hospitals will buy broad, fuzzy "AI transformation" packages, fail to measure anything specific, and quietly get the budget-room-not-returns outcome from the Gartner layoffs study. The lesson holds across every industry: pick one expensive, measurable task, automate that, and count the dollars — don't buy a vague platform and hope.
The third is vendor lock-in and cost creep. Once a tool is wired into Epic across seven hospitals, the vendor has enormous pricing power at renewal. Today's bargain can become tomorrow's ransom — the same pass-through pricing dynamic that plays out across all enterprise AI. The savings are real; just go in knowing the negotiating leverage shifts the day you flip it on.
None of this kills the opportunity. It just separates the systems that get the $37 million from the ones that get a bill and a headache.
What this means for you
Let me make it practical, depending on where you sit.
If you work in healthcare, the message is the one I'd give a friend: lean into being the person who wields these tools, not the one who fears them. The systems winning here aren't cutting clinical staff — they're making each clinician dramatically more productive and less burned out. The nurses, doctors, and pharmacists who become fluent in the AI layer are the ones who'll thrive, and that AI fluency now carries a real pay premium across every field, as I track in the highest-paying AI jobs of 2026.
If you run or advise a business — any business, healthcare is your case study in how to actually get AI ROI. The winners didn't buy AI to slash headcount. They found the single most expensive, most hated, most repetitive task (after-hours documentation), automated that, and measured the recovered value in hard dollars. Copy that move in your own industry. The general version — automating expensive human services with AI — is the same engine behind the best solo AI businesses of 2026 and how companies cut support costs 40–60%.
If you invest, healthcare AI is one of the most defensible corners of the whole boom, because the ROI is measurable and the switching costs are enormous — once a tool is wired into Epic across seven hospitals, it doesn't get ripped out. OpenEvidence tripling to $12 billion is a signal, not a fluke. Just remember that, as I warned in the AI wealth divide, most of these gains are being captured in private rounds you and I can't buy into — so the realistic public play is the platforms and infrastructure underneath, plus watching which of these names eventually reaches the market.
The honest take
Healthcare is usually the slowest industry on Earth to adopt anything — it took decades to fully digitize paper charts. So when hospitals start signing system-wide AI contracts in a matter of weeks, that tells you something profound: the ROI got so obvious that even the most cautious, most regulated, most risk-averse buyers in the economy couldn't say no. A 2x return that shows up as $37 million a year in recovered clinician time is not a debate. It's a decision.
What I find genuinely encouraging here is the shape of the win. This isn't AI replacing doctors. It's AI handing exhausted clinicians their evenings back, letting them see more patients and burn out less — and the money follows precisely because the humans are amplified, not erased. That's the version of the AI future worth rooting for, and healthcare just proved it pays.
So here's the question to carry out of this, wherever you work: what's the "pajama time" in your own industry — the expensive, hated, repetitive task everyone assumes is just part of the job? Because whoever automates that first is about to find their own $37 million.
