Healthcare IT News Features TTN’s 5-Clinic Transformation Powered by SPRY
Learn More
News
2 min read
The Infrastructure Gap: What 2025 Revealed About Healthcare AI Adoption
By Brijraj Bhuptani, cofounder & CEO, SPRY Therapeutics
2025 was supposed to be the year healthcare AI went mainstream. Instead, it became the year the infrastructure gap became impossible to ignore.
According to the AMA, 66% of physicians are now using AI, up from 38% just a year ago. Black Book Research surveyed 1,158 payer IT leaders and found that 2026 buying decisions will hinge on one test: solutions must be compliance-ready AND outcomes-relevant.
But here's what the headlines miss: Most healthcare AI is being deployed on infrastructure that can't support it.
I spent 2025 talking to healthcare operators—solo practitioners, multi-clinic groups, hospital-owned practices. The pattern is consistent: Everyone's excited about AI. Almost no one's asking if their infrastructure can handle it.
This isn't a technology problem. It's an architecture problem.
U.S. healthcare isn't one $4.5 trillion market. It's 1,000+ different $4 billion markets, each with its own workflows, payer rules, compliance requirements, and definition of "working."
Most healthcare AI is being built for the illusion of horizontal scale. "Works for everyone" platforms with beautiful demos. Implementation reality that falls apart at the edges.
Because healthcare rewards vertical depth, not horizontal breadth.
The operators who actually adopt new technology aren't asking "Does this have AI?" They're asking:
What works: Understanding the specific chaos—the payer-specific prior auth forms, the CPT code modifiers that vary by state, the authorization workflows that differ by specialty—and rebuilding the infrastructure underneath.
Legacy EMR vendors bolted AI scribes onto 20-year-old infrastructure. The scribe worked, sort of. But the note it generated couldn't auto-populate billing codes because the revenue cycle module was built in 2003 and doesn't understand modern ICD-10 logic.
The AI saved 20 minutes. The broken handoffs cost 30.
New AI startups launched with beautiful demos: ambient documentation, smart scheduling, predictive analytics. But when operators tried to implement them:
According to APTA research, 41% of healthcare professionals still spend over 4 hours per day on administrative tasks despite all the "AI efficiency" promises. The industry vacancy rate sits at 11%, burnout runs between 45-71%.
Adding AI to broken workflows doesn't fix healthcare. It automates the chaos faster.
Most healthcare AI in 2025 was deployed in the wrong places.
Everyone wanted to talk about "AI doctors" and "clinical decision support." They miss where the leverage actually is.
The breakthrough isn't AI doing clinical work. It's AI eating the toil so humans can focus on work.
Work:
Toil:
The strategic choice matters:
Building AI for clinical judgment means competing on accuracy, liability, and trust in life-or-death decisions. The bar is extraordinarily high. The adoption curve is measured in decades.
Building AI to automate eligibility verification, prior auth submissions, and claim scrubbing means competing on speed and ROI. The bar is still high—if AI gets billing wrong, clinics don't get paid—but the adoption curve is measured in quarters.
Great AI doesn't replace the work. It eliminates the toil.
In early 2025, saying "we use AI" meant everything. By late 2025, it meant nothing.
The question that started mattering: Did you build for trust, or did you build for demo day?
Healthcare has constraints that consumer tech doesn't:
Most AI startups skip these. They're hard. They're boring. They take time.
But this boring work became the actual moat in 2025.
ONC certification (CEHRT): Enterprise buyers won't talk to you without it. It's the entry ticket.
FHIR-forward from day one: TEFCA went live in 2025. HTI-1 pushes USCDI v3 by January 1, 2026. Closed workflows that trap data are becoming indefensible.
Human-in-the-loop always: AI drafts, humans approve. No black boxes. This is how you earn trust from the people actually using the system every day.
This unlocks two strategic advantages: the ability to sit inside enterprise without forcing disruption (hospital-owned PT groups can run platforms for rehab while ortho stays on Epic), and the ability to sell modular infrastructure (clinics keep their clinical system, just fix revenue operations—effectively doubling addressable market).
What's working: Operators are adopting billing and prior auth engines on top of their existing EMRs. They keep their clinical workflows, fix what's costing them money, without the risk and disruption of full platform migration.
The data points to four major shifts:
January 1, 2026 brings new CMS requirements (CMS-0057-F). By March 31, impacted payers must publicly report PA metrics—approval rates, denials, turnaround times. This is the first time providers can point to hard numbers and say "this payer takes 14 days, this one takes 3."
Clinics that automate PA workflows now will have 6-month proof points when competitors are still drowning in manual fax loops.
TEFCA is live. HTI-1 mandates USCDI v3 by January 1, 2026. The question operators are asking: Is your platform ONC-certified? Does data move seamlessly across eligibility, prior auth, and clinical systems?
If the answer is no, you're not infrastructure. You're a feature. And features get commoditized.
Black Book's survey shows the 2026 test: solutions must be compliance-ready AND outcomes-relevant.
"We added AI" doesn't cut it. "We reduced denials by 40% and improved collections by 30%" does.
The "see more patients" playbook is dead. CY 2026 KX thresholds stayed flat at $2,480. Patient responsibility keeps growing. Collections are harder. The new playbook is "get paid for every visit, faster."
Epic, Cerner, and other EHR giants are bundling AI features into their platforms. If you're a point solution that only does one thing—and that one thing is now free inside the EMR—you're in trouble.
The survivors will be the ones who built modular infrastructure that works standalone OR inside existing systems.
The healthcare AI race everyone's talking about—who has the best model, the slickest UI, the most funding—is a distraction.
The real race is infrastructure.
Who's rebuilding the rails that move data, money, and work through healthcare? Who's automating toil so humans can focus on work? Who's earning trust through compliance, outcomes, and proof—not promises?
The pattern I'm seeing work: Pick one underserved market where the need is clearest. Build deep enough to prove real outcomes—not just features, but actual infrastructure that runs from patient intake to getting paid. Then prove the architecture generalizes across adjacent specialties.
Outpatient rehab was a proving ground for many—a $52B segment where clinics were using 5-7 disconnected tools, suffering low collections, watching therapists burn out. The platforms that succeeded there are now expanding into mental health, pediatric therapy, chiropractic. Not by "adding specialties"—by proving their infrastructure generalizes.
That's what infrastructure looks like: vertical depth in one market, horizontal scalability proven across specialties, modular enough to work inside complex enterprises.
The infrastructure age of healthcare is here.
Some companies will build it. Most will talk about it.
The operators who can tell the difference will be the ones who win.