AI in Healthcare: What Actually Happened (And Why the Best Is Still Stuck in Pilots)

Three years ago, the AI conference circuit was full of presentations about curing cancer. The slides were beautiful. The timelines were bold. The science, in many cases, was real.

The cure is not here yet.

What is here is harder to put on a conference slide but more immediately consequential: AI that catches the stroke you almost missed, closes the care gap that was slipping through the cracks, and keeps the physician in the room instead of staring at a screen. The revolution in healthcare AI is not the one anyone predicted. It is quieter, more specific, and in the places where it has been allowed to actually run — genuinely remarkable.

The problem is that almost nowhere has it been allowed to actually run.

I. What Everyone Expected vs. What Arrived

The narrative around AI and healthcare in 2022 and 2023 was built around moonshots. Drug discovery in months instead of decades. Cancer diagnosis with superhuman accuracy. Personalized genomic medicine at scale.

Those promises were not false. They were premature.

What arrived first was not the dramatic breakthrough but the operational win. AI that reduces physician burnout by eliminating documentation overhead. AI that spots the imaging finding a radiologist might catch on their third review but not their first. AI that flags the patient most likely to miss their appointment so a human can intervene before the no-show happens.

These are not the outcomes that make the cover of a magazine. They are the outcomes that keep health systems functional and patients alive.

II. The Wins That Are Actually Happening

The results in the places where healthcare AI has been meaningfully deployed are not incremental. They are structural.

One major health system deployed ambient AI scribes from Abridge across their ambulatory environment. 2,500 active users. More than 30,000 clinical notes generated per week. Measurable impact on burnout reduction, provider satisfaction, on-time chart closures, and clinician productivity — at system scale, not pilot scale.

AI-powered radiology detection is being operationalized across health systems to identify intracranial hemorrhage, vessel occlusion, pulmonary embolism, and cervical spine fractures earlier and more consistently than traditional review workflows. These are not edge cases. These are conditions where the difference between catching it in the first read and the third is measured in survival rates.

PopEVE, an AI system developed at Harvard Medical School, was applied to roughly 30,000 patients with severe developmental disorders and surfaced probable diagnoses for approximately one-third of them — patients who had been undiagnosed, in many cases, for years.

AI-driven scheduling and pre-visit workflow tools have measurably reduced no-show rates, collected past-due balances, closed care gaps, and reduced call center volume across the health systems that have deployed them at scale.

None of this is curing cancer. All of it is saving lives, reducing waste, and giving clinicians back the cognitive capacity that documentation and administration had consumed.

III. The Communication Layer That Makes Everything Else Matter

The best care in the world doesn't help if the patient doesn't show up.

That sounds obvious. It rarely gets treated that way. The ambient scribing, the imaging AI, the diagnostic models — all of it fails if the patient misses the appointment, doesn't respond to the recall notice, or falls through the cracks between one communication system and another.

Patient communication in most health systems is a patchwork. A hospital might have separate platforms managing email, SMS, RCS, and voice calls, each pulling from a different slice of the patient record. A patient who changed their phone number two years ago is still getting texts to a dead number. A patient who responded to an email reminder is still being called three times by the automated phone system. A patient flagged for a care gap in one system is invisible to the outreach tool running in another. The systems don't talk to each other because they were never designed to.

AI changes both sides of this problem.

The communication layer itself — email, SMS, RCS, voice — can be automated and personalized at scale. AI determines the right channel for the right patient at the right time, adjusts messaging based on prior response patterns, and sequences outreach across multiple touchpoints without human coordination at each step.

But the more consequential capability is what happens upstream: AI normalizing disparate patient databases. Reconciling identity records across systems that evolved separately. Resolving conflicting contact information. Creating a unified patient record that every communication channel draws from consistently — so the left hand knows what the right hand is doing.

GetScaled deployed this infrastructure for a major healthcare organization. The result was a 55% increase in patient recall rates — patients who had been falling through the cracks of fragmented communication systems returning for the care they needed.

That number is not primarily a technology story. It is an identity and infrastructure story. The AI did not change what was being said to patients. It changed whether the message reached the right patient, through the right channel, at the moment it would actually land.

In healthcare, that is not a marginal improvement. Missed appointments are one of the primary mechanisms through which chronic disease progresses undetected, preventable hospitalizations occur, and health systems absorb costs that early intervention would have eliminated. Closing the communication gap is not a patient convenience feature. It is a clinical outcome driver.

IV. The Gap Between What Works and What Scales

Here is the number that matters: almost none of it is scaling.

The World Economic Forum published an analysis in early 2026 identifying why digital and AI solutions in healthcare systematically fail to move beyond pilot deployments. The findings were not about model quality. They were about infrastructure: scattered data, stringent regulations, limited access to the anonymized training data needed to run AI safely, and fragmented governance that creates an environment where speed consistently outstrips strategy.

Healthcare AI today looks almost exactly like enterprise AI in every other sector — a landscape of successful pilots, a small number of system-scale deployments, and an enormous gap between the two that is not closing at the rate anyone hoped.

A Nature npj Digital Medicine analysis published in 2025 found that governance was the primary variable separating healthcare AI deployments that scaled from those that did not. Not the models. Not the training data quality. Governance — who owns the system, who is accountable when it makes a wrong decision, how outputs are audited, how the system is updated when clinical evidence changes.

V. Why Healthcare Is the Hardest Environment for AI Execution

Healthcare presents the scaling problem in its most acute form.

Every other sector where AI is being deployed has integration complexity, data fragmentation, and governance gaps. Healthcare has all of those plus regulatory constraints that are still being written, liability exposure that is legally undefined, clinicians who are actively using AI outside enterprise control — creating a governance gap that compounds with every new tool — and a reimbursement model that does not yet know how to value AI-driven outcomes.

Premier Inc., a healthcare improvement company serving more than 4,400 U.S. hospitals, called 2026 "the year the AI Wild West ends" and argued that health systems must take control of their AI governance or face escalating risk as deployments proliferate without oversight.

This is the same structural failure mode as every other sector. The AI works. The architecture around the AI does not.

VI. The Infrastructure Problem Behind the Scaling Failure

The healthcare AI deployments that have moved from pilot to system scale share one characteristic that is almost never cited in the press releases: they built the execution infrastructure before they built the applications.

In the cases that scaled — ambient scribing at health system level, imaging AI deployed across an entire radiology workflow, scheduling AI integrated directly into the EHR — the AI was not bolted onto existing systems. It was built into a unified operational layer that resolved patient identity consistently, connected to clinical workflows directly, and produced outputs that the system could act on without requiring manual human handoffs at each step.

In the cases that did not scale — which is the majority — the AI produced intelligence that a human then had to carry somewhere. The model was right. The handoff broke.

This is not a healthcare-specific problem. It is the universal failure mode of AI deployment. Healthcare simply makes it more visible because the stakes are higher.

VII. What Scaling Actually Requires

The path from "our AI pilot showed remarkable results" to "our AI is operating across the entire system" has been documented in enough deployments now that the requirements are known.

Unified patient identity that resolves consistently across data sources. Direct integration into the workflows where clinical and operational decisions are made — not an adjacent system that a human bridges manually. Governance infrastructure that defines ownership, accountability, audit trails, and update protocols before the first patient interaction, not after the first incident. And a feedback loop that allows the system to learn from its own outcomes rather than requiring periodic manual retraining cycles.

These are not technically exotic requirements. They are organizational and architectural requirements that most health systems have not yet prioritized because the pilots looked so good without them.

VIII. The Lesson That Applies Everywhere

Healthcare is the sector where the gap between AI's demonstrated capability and AI's operational deployment is most consequential. But it is not the sector where the gap is unique.

The same fragmented data, the same governance vacuum, the same handoff problem between AI intelligence and human execution — these are the conditions producing the 88% AI deployment failure rate across every sector. Healthcare makes the human cost of those failures visible in ways that a failed marketing automation deployment does not.

The AI that catches the stroke. The model that surfaces the undiagnosed genetic disorder. The scheduling system that closes the care gap before the patient falls through it. These are working, in the places where they have been allowed to work.

The constraint is not the science. It has never been the science.

The constraint is the architecture between the AI and the patient. Close that gap, and the outcomes that everyone expected three years ago become possible. Leave it open, and the pilots will keep producing remarkable results that never reach the people who need them.

At GetScaled, the infrastructure problem is what we solve — not in healthcare specifically, but in the operational layer where AI intelligence becomes real-world action. The architecture that closes the gap between model output and execution is the same regardless of sector. The stakes in healthcare make the case for getting it right more urgent than anywhere else.

Learn more at GetScaled.com.