I've spent 25 years inside healthcare organizations. I was a VP at One Medical when we were scaling past 100 offices across 20 markets. I was CIO at a PE-backed dermatology group, where I owned every technology decision from EMR configuration to cybersecurity compliance. Before that, I spent more than a decade at Epocrates, the company that introduced digital point-of-care tools to physicians before smartphones existed.

I've watched a lot of technology cycles in healthcare. I've seen the ones that genuinely changed the underlying economics of care delivery. I've seen more that were forgotten inside of three years.

That history shapes how I evaluate what's happening in healthcare AI right now. The hype cycle is loud. The vendor pitches are everywhere. But the actual deployment data, the honest accounting of where AI delivers and where it doesn't, is harder to find.

Here's my mid-year read.

What's actually working

The clearest, most consistent results in healthcare AI right now are coming from workflow automation in high-volume, rules-based processes: scheduling, inbound call handling, outbound patient outreach, referral intake, cancellation recovery.

These are the workflows where AI has moved from pilot to infrastructure at a meaningful number of practices. Why here? Because the tasks are bounded, the rules are explicit, and the failure modes are recoverable. The economics are also clear: every unscheduled referral, every cancellation that doesn't get backfilled, every no-show that goes unaddressed is a direct P&L event.

Consider referral conversion. A top-3 national dermatology practice we work with was converting 51% of its inbound referrals. That means nearly half of the patients already sent to them by a referring physician, patients who had already taken the step of getting a referral, never made it onto the schedule. At roughly 7,000 referrals per month, that's approximately 3,400 patients per month walking out the door before they ever walked in.

This is not a marketing problem. It's a follow-through problem. And it's exactly the kind of bounded, rules-based workflow that AI handles well.

Fully autonomous AI referral processing, fax arrives, AI extracts demographics and diagnosis, creates the patient record in the EHR, contacts the patient by phone and SMS, matches availability and insurance, and confirms the appointment directly in the scheduling system, moves conversion from 51% to 88%. Zero staff involvement in the process. The gap between those two numbers represents millions of dollars in annualized revenue from patients who were already in the pipeline.

That's one workflow. The same pattern holds across cancellation recovery, no-show re-engagement, and outbound recall. A large specialty practice running outbound campaigns across all of these workflows is seeing a 39% success rate filling last-minute cancellations and a 14% overall resolution rate across all campaign types. Again: no staff involvement. The AI manages the outreach, handles the conversation, and writes the appointment to the EHR.

This is real. It's measurable. And it's happening at scale, not in one-clinic pilots.

What's still mostly noise

The broader clinical AI category is moving fast technically, but it's at a different stage of the maturity curve in terms of actual deployment.

Diagnosis support. Treatment recommendation. Predictive risk stratification. The model capabilities are genuinely impressive, and some of the research results are promising. But the deployment reality is more complicated than the demos suggest.

The evidence base for most clinical AI tools is still thin or narrowly validated. The liability frameworks have not kept pace with the capabilities. Real-world workflow integration, across EMR configurations that vary by location, provider preferences that vary by physician, and regulatory requirements that vary by state, is harder than it looks in a conference presentation.

This doesn't mean these tools won't matter. Some of them will be transformative over the next five to ten years. But healthcare operators who are being pitched clinical AI as a near-term operational solution should be asking hard questions: What does the peer-reviewed evidence actually say? What does the contract say about liability when the model is wrong? How long did integration take at the reference site, and what EMR were they on?

That's not reflexive skepticism. That's how healthcare technology should be evaluated.

What surprised me this year

The thing I didn't fully anticipate was how quickly the operator conversation shifted in tenor.

Eighteen months ago, the primary question from COOs and CFOs evaluating AI was some version of "does this actually work?" The skepticism was high, and it was justified. The category was young, the vendor claims outpaced the evidence, and operators had been burned before by technology that promised transformation and delivered dashboards.

That conversation has changed.

The operators who deployed early, who went through the implementation learning curve and built confidence in the system, are no longer debating ROI. They're asking how to extend it. The question I hear most often from senior ops leaders at large specialty practices today isn't "should we do this?" It's "how do we add billing? When can we automate referral intake for our other service lines? What's the roadmap for labs?"

That shift in the nature of the question is one of the clearest leading indicators I've seen that a technology has crossed from experiment to infrastructure.

What practice leaders should actually be evaluating right now

If you're a COO, CFO, or VP of Operations at a large specialty practice trying to cut through the noise, here's the framework I'd use.

Start with the workflows that are both operationally painful and rules-based. Referral intake, scheduling, inbound access, outbound recall. These are where the ROI is clearest and the AI solutions most mature. Get a real win here before you try to solve harder problems.

Evaluate vendors on outcomes, not demos. Ask for case studies from practices of comparable size and complexity. Ask specifically how performance is measured, what the baseline was before deployment, and who owns the definition of success. And look carefully at the pricing model: a vendor who gets paid per booked appointment has a fundamentally different incentive structure than one who gets paid per call placed.

Ask about the failure rate. Every honest AI vendor will tell you their system has one. The question isn't whether failures happen. It's how the system handles them, how they're logged, and how the feedback loop drives continuous improvement. A vendor claiming near-zero failures is either not measuring carefully or not being straight with you.

Think about integration depth before you think about AI. The best model in the world doesn't overcome a shallow EHR integration. Ask exactly which integrations your shortlisted vendors have built, how deeply they write back to the scheduling system, and what the implementation timeline looks like for your specific EMR environment.

Where this goes in the next 12 to 18 months

The practices that invested early in AI-powered patient access workflows are already seeing compounding effects. Better access drives better retention. Better retention drives better lifetime value. That changes the unit economics of the practice over time in ways that are hard to replicate with any other operational investment.

The vendors who delivered real results in scheduling and access are expanding into adjacent workflows. Billing. Referral intake. Prescription refills. Labs. The platform logic holds: each new workflow added builds on integrations and institutional configuration already in place, which is why the second and third workflows come online faster than the first.

The category will also consolidate. The next 12 to 18 months will make it easier to separate vendors with real deployment data and real client outcomes from those who have been selling the promise of AI without the operational receipts to back it up.

If you're an operator evaluating the space now, that consolidation is your friend. The proof points are accumulating. The signal is getting clearer.

You just have to ask the right questions.