By Brent Dover, Chief Executive Officer, Carta Healthcare
For several years the central question about AI in healthcare was whether it worked at all. That question has largely been answered. Health systems have run the pilots, measured the results, and confirmed the value. The surprising part is what happened next, which in most cases is not very much.
In an April 2026 survey of 32 healthcare leaders, a striking pattern emerged. Among organizations that had already realized measurable value from AI, 71 percent were not expanding it at pace. Validation is happening. Scale is not. The proof of concept works, and then the rollout stalls.
It would be easy to assume the holdup is skepticism, but the data points elsewhere. When leaders were asked what most slows adoption, the top barrier was electronic health record integration, named by 44 percent. Lack of executive sponsorship or budget followed at 37 percent, and competing organizational priorities at 33 percent. Clinician trust, ongoing cost, and regulatory concern were tied lower down, each at 26 percent.
Read that order carefully. The people closest to these tools are not the ones holding them back. The friction is structural. A model can prove its value in a contained pilot and still be enormously difficult to wire into the daily workflow of an organization that runs on its EHR. Integration, not intelligence, is the gating issue.
A pilot is a forgiving environment. Scope is narrow, the team is motivated, the data is curated, and someone is watching closely. Scale removes every one of those advantages. Now the tool has to live inside existing systems, serve people who did not volunteer, handle the messy long tail of real documentation, and keep working when no one is paying special attention.
This is why so many promising results never travel. The pilot answered the question the organization asked, which was whether the tool could work. It did not answer the question that actually governs adoption, which is whether the tool can fit. Those are different problems, and the second is harder.
The same survey asked leaders what would most accelerate adoption. The answers were refreshingly concrete. Forty-eight percent wanted a solution that fits inside the EHR. Thirty-six percent wanted peer case studies with measurable outcomes. Twenty-eight percent wanted a vendor willing to assume performance risk.
Taken together, these signal a buyer who has moved past the demo. They are not asking whether the technology is impressive. They are asking whether it will disappear into the work they already do, whether someone like them has seen real results, and whether the vendor has enough conviction to share the downside. Peer evidence and a willingness to be paid on outcomes both point in the same direction. Healthcare buyers want proof, not slideware.
There is one more finding worth sitting with. When asked who owns AI strategy, 41 percent of organizations pointed to clinical leadership. Nineteen percent said the executive team and 15 percent said the IT department. The number that should give everyone pause is the 26 percent who reported no clear owner at all.
Scale is an organizational act, not a technical one. It requires someone with the authority to change workflows, fund integration, and hold a standard across departments. Where ownership is clear, it tends to sit with clinical leaders, which fits the nature of the work. Where ownership is absent, even a tool that works will drift, because no one is accountable for carrying it from one department to the next.
The lesson for anyone building or buying healthcare AI is that the hard part starts after the technology is proven. The first mile, getting a model to deliver value, is increasingly solved. The second mile, getting that value to spread across an enterprise, is where most of the effort now belongs.
Practically, that means treating EHR integration as a first-class design problem rather than a final-step afterthought. It means publishing real outcomes from real peers, because abstract capability claims no longer move a market that has seen plenty of them. It means being willing to tie payment to results, which both proves conviction and aligns incentives. And it means insisting on a clear owner before the rollout begins, not after it stalls.
Part of what stalls these rollouts is a category error. Organizations run a pilot to answer the question can this work, the pilot says yes, and the result is mistaken for a strategy. It is not. A successful pilot is evidence, not a plan. The plan has to answer a harder set of questions about integration, ownership, funding, and the change management required to move a tool from a motivated test team to an entire department that did not volunteer. None of those questions are technical, which is exactly why a technically successful pilot does not resolve them.
The buyer signals in the survey point toward a more honest sequence. When 28 percent of leaders say they want a vendor willing to assume performance risk, they are describing a desire to move the burden of proof off their own shoulders. This is the logic behind a Pay for Value approach, where payment is tied to results that actually materialize rather than to promises made in a sales cycle. Tying economics to outcomes does two things at once. It signals conviction, and it aligns the vendor's incentives with the slow, unglamorous work of making a tool succeed at scale rather than merely demonstrate well.
Treating electronic health record integration as a first-class design problem is the other half of the answer. The 44 percent who name integration as the top barrier are not asking for a more capable model. They are asking for a solution that disappears into the systems they already run. A tool that demands the workflow bend around it will always lose to the gravity of the EHR. A tool designed to fit inside that gravity has a chance to travel from one department to the next, which is the only path from a promising pilot to enterprise value.
The sequence that works tends to invert the usual order. Instead of selecting a model and then worrying about integration, the organizations that scale start by naming an owner and treating integration as the first design constraint, then choose technology that fits the workflow they already run. Put the unglamorous problems first, because they are the ones that actually decide whether a proven tool spreads. Capability is the easy part to acquire. The discipline to solve for ownership, fit, and funding before the rollout begins is what separates a pilot that travels from one that stalls after a promising start.
AI in healthcare has cleared the bar of working. The organizations that pull ahead now will be the ones that treat scaling as its own discipline, with its own barriers and its own answers. The proof is in. The work is to make it travel.
