*Written following the HLTH Virtual Roundtable, 27 May 2026*

During a recent HLTH roundtable, we spent an hour talking about what actually determines whether AI succeeds in a clinical environment. Not whether it can detect a finding, summarize a note, or flag a deteriorating patient. Whether it survives contact with the real world.

That's not a comfortable thing to say in a room full of people who have invested years building better models. But the data backs it.

My answer, which I've been refining for a while now: the model is rarely the problem. The infrastructure around it almost always is.

90% Deployed. 19% Working.

Healthcare AI has come a long way. The question is no longer whether AI can perform on narrow clinical tasks. In controlled environments, many tools hit accuracy benchmarks that match or exceed specialist performance. That bar has been cleared.

The gap that remains is between what works in a validation study and what works in a real hospital. A survey of 43 major U.S. health systems, published in JAMIA in mid-2025, found that 90% had deployed imaging AI. Only 19% reported high success. Ambient documentation tools were the only category hitting 100% adoption activity across respondents. Everything else showed a version of the same pattern: deployed, but not working.

That 71-point gap between deployment and success in imaging AI is the number I can't stop thinking about. It means health systems are buying and installing AI at scale while a large share of it is sitting underused, ignored, or partially trusted. That's a system failure, not a technology one.

Part of this is the pilot problem, which the industry has been talking about for years without solving. Pilots are designed to succeed: motivated site, supportive clinical lead, clean data, defined metric, 90 days, done. What they don't test is whether the tool survives a weekend call team that wasn't trained on it, or whether ownership is clear when something goes wrong at 2am. None of that shows up in a pilot. All of it shows up at scale. Tom Mihaljevic, CEO of Cleveland Clinic, put it well in a recent podcast conversation:

Sprinkle AI on a poorly organized health system and you get a poorly organized health system with bad software."

 

Cleveland Clinic spent years standardizing processes and data before expecting AI to deliver value. Most health systems haven't done that work.

Why Ambient AI Is the Exception

One technology keeps standing out in every deployment conversation: ambient documentation.

It's not that AI scribes are technically superior to every other clinical AI tool. They stand out because they solve a real, acute problem (documentation burden is the most consistently cited driver of physician burnout) by slotting into existing workflows without forcing practice change, and prove their value fast. A clinician finishes a note that would have taken 20 minutes in under five. The feedback loop is immediate and personal in a way that most clinical AI never achieves.

A multicenter study published in JAMA Network Open in 2025 followed 263 physicians across six U.S. health systems for 30 days after AI scribe adoption. Burnout prevalence dropped from 51.9% to 38.8%, a 74% reduction in adjusted odds. The signal was consistent across academic medical centers and community hospitals.

But I'll resist the hype. A JAMA study published this year across 8,581 ambulatory clinicians at five major academic health systems found the average time savings were 13.4 minutes of EHR time per day. Real, but modest. More telling: after-hours documentation time didn't change. And only 32% of adopters used the tool frequently enough to get meaningful benefit.

So even the poster child for successful deployment is more complicated than the marketing suggests. The tools work when used consistently and when the surrounding workflow actually supports them. When adoption is fragmented, results are fragmented. That's the lesson, not just "ambient AI works."

Deploying AI Raises the Legal Bar, Not Lowers It.

Trust in AI systems is asymmetric. You build it slowly and lose it fast. A health system can run a diagnostic AI tool for two years, process 50,000 cases, build real clinical confidence in the workflow, and then watch that confidence evaporate after one high-profile miss. Rebuilding takes longer than building did the first time.

Most governance conversations treat this as a reputational risk. The legal dimension is sharper, and almost nobody is talking about it clearly.

Research published in NEJM AI in 2025 ran a randomized study with simulated jurors assessing radiologist liability in cases where AI was present. When the AI flagged a finding and the radiologist missed it, jurors sided with the plaintiff in 72.9% of brain bleed cases and 78.7% of lung cancer cases. The AI's presence didn't split responsibility. It raised the standard the clinician was held to.

Think about what that means operationally. Deploying AI doesn't protect a health system from liability. It increases the evidentiary bar the clinician is expected to clear. Every missed finding now carries an implicit question: did the AI flag it? If yes, why was it ignored? If not, was the tool functioning correctly, and who was responsible for knowing?

Most health systems have not modeled this exposure. The legal frameworks were built before AI entered clinical workflows at scale. And the liability ratchet works in one direction: AI agreement with the clinician doesn't reduce liability, but AI disagreement, when the AI was right, increases it substantially.

That asymmetry changes what operational discipline means. Monitoring AI performance post-deployment isn't an implementation detail. It determines whether you're still allowed to use the tool in three years, and whether your legal team has any defensible ground when something goes wrong. Alert fatigue, unexplained output changes, one case that makes it into a mortality review. Those are the events that end programs and open organizations to liability they hadn't priced in.

A 2026 survey of healthcare IT leaders named governance as the top concern, ahead of technology selection, integration, and cost. The technology arrived before the structure to manage it did.

Who Actually Owns This

The roundtable kept circling back to a theme that I think is the actual core question of healthcare AI right now: who owns this?

Not who built the model. Who owns the deployment, monitors performance, makes the call when the tool starts behaving differently six months in, and is accountable when a clinician acts on AI output that turns out to be wrong.

In most health systems, that question doesn't have a clear answer. AI governance sits somewhere between IT, clinical quality, and the office of the CMO, with no single team holding real accountability for what happens after go-live.

The health systems that have cracked this aren't waiting for the answer to emerge organically. They're naming it. According to the AHA's 2025 Market Insights report on AI in healthcare delivery, Cleveland Clinic appointed its first Chief AI Officer reporting directly to the CIO, paired central infrastructure with local use case generation, and built a cross-disciplinary vetting process. The result: their AI sepsis tool identified 3x more cases early enough for treatment. Corewell Health stood up a formal AI Center of Excellence with multidisciplinary leadership across compliance, physicians, nurses, and IT, reviewing every pilot for alignment with data governance and patient safety before it scales. HCA built a five-person Digital Transformation Committee at the CEO level with a four-part test for every AI initiative: is it digitizable, automatable, AI-advanceable, and genuinely impactful?

What those models share isn't a specific org chart. It's that someone has a name on the door, accountability is explicit, and governance is built into the deployment lifecycle rather than bolted on afterward. Named accountability structures. Stage gates with real criteria. Monitoring that continues after go-live, not just during it.

None of that is glamorous. None of it gets announced at a conference. But it's the difference between a health system that can actually absorb AI over a decade and one that accumulates a graveyard of promising pilots.

 

What I Actually Think

AI capability is no longer the limiting factor in healthcare. That sentence should change how every health system allocates its next dollar.

The question isn't whether a better model exists. It's whether the organization can carry it, whether the workflow actually absorbs it, and whether trust is being built systematically rather than eroded through avoidable failures.

The tools that will matter in five years aren't the ones with the best benchmarks. They're the ones that fit cleanly into clinical environments, prove their value fast, and can be monitored and managed by teams that don't require a PhD to operate them.

We've cleared the capability threshold. The bottleneck is everywhere else. That's not a pessimistic assessment. It's a useful one. It tells you exactly where the work is.