By Brent Dover, Chief Executive Officer, Carta Healthcare
Most conversations about AI in healthcare are conversations about technology. Which model, which capability, which benchmark. These are reasonable things to discuss, but they tend to crowd out a less glamorous question that turns out to be more predictive of success. Inside a given health system, who actually owns the AI strategy? The answer, more often than it should be, is no one in particular.
When healthcare leaders were asked who owns AI strategy in their organization, 41 percent pointed to clinical leadership, 19 percent to the executive team, and 15 percent to the IT department. The number worth pausing on is the remaining share. Roughly one in four organizations reported no clear owner at all.
That is a striking admission for a category of technology described as transformative. It means a meaningful portion of health systems are pursuing AI without anyone clearly accountable for where it goes, how it is governed, or whether it scales. The consequence is predictable. Without an owner, a tool that works in one department has no one whose job it is to carry it to the next. Pilots succeed and then stall, not for lack of merit but for lack of a person with the authority and the mandate to push them forward.
Where ownership is clear, it tends to sit with clinical leadership, and that is not an accident. The decisions that determine whether clinical AI succeeds are clinical decisions. What level of accuracy is acceptable, how edge cases are handled, who validates outputs, what happens when the model and the record disagree. These are not questions an IT function can answer alone, because they are judgments about clinical risk and clinical meaning. When clinicians own the strategy, governance is anchored where the consequences actually fall.
This also reflects a broader finding. Ninety-two percent of leaders say deep healthcare domain expertise is critical in an AI vendor, the most lopsided result in the survey, with half giving it the maximum possible score. The same instinct that puts clinicians in charge internally pushes organizations toward partners who understand the clinical domain rather than vendors who treat healthcare as a generic application of a model. Ownership and expertise are two expressions of the same conviction, that clinical context is not optional.
Here is the strategic point that the technology-first framing misses. Model capability is improving and will keep improving, and on a long enough horizon it trends toward becoming a utility. What does not commoditize at the same pace is the governance around the model. The oversight workflows, the inter-rater reliability management, the adjudication processes, the audit interface, the registry-specific expertise, and the integration into the operational fabric of a health system. Even an organization with an excellent model still has to solve all of that, and solving it is harder than acquiring intelligence.
This is why the durable advantage in clinical AI is not the model underneath. It is the system around it. An organization that owns the workflow, the governance infrastructure, the domain expertise, and the accountability interface can let the model evolve indefinitely without losing its footing, because the model was never the thing holding everything together. The model is replaceable. The risk-governed operating system around it is not. The system is the moat.
It follows that the real risk for a health system is not that AI improves and outpaces its plans. AI will improve. The risk is treating AI as purely a technology function and assuming that governance is something the organization can pick up casually along the way. Capability without governance is exactly the configuration that produces an expensive surprise, because it optimizes for what the model can do and neglects who is accountable when it is wrong.
An organization with a clear owner, clinical leadership at the center, and a deliberate governance structure can absorb new capability as it arrives and put it to use safely. An organization without an owner will keep generating promising pilots that never become enterprise capability, and will mistake the absence of a clear failure for the presence of a strategy.
It is worth being concrete about what governance means here, because the word invites the image of a policy document filed and forgotten. Real governance in clinical AI is operational. It is the oversight workflow that runs on every case, the inter-rater reliability management that catches drift, the adjudication process that resolves disagreement, the audit interface that makes data defensible on demand, and the registry-specific expertise that knows what each program actually requires. None of that lives in a binder. It lives in the daily operation of the work, which is why it cannot be acquired as an afterthought once a model is already in production.
This connects directly to why domain expertise ranks so highly when leaders evaluate vendors. The 92 percent who call deep healthcare expertise critical, with half awarding it the maximum possible score, are expressing the same conviction that puts clinicians in charge internally. A partner who treats healthcare as a generic application of a model will build governance that looks plausible and fails at the edges, because the edges are where clinical knowledge is required. A partner grounded in the domain builds governance around the cases that actually break, which are the only cases governance exists to handle.
Clear ownership is what allows all of this to scale. When a clinical leader owns the strategy, there is someone with the authority to set the accuracy standard, fund the integration, and carry a consistent approach across departments. When ownership is absent, governance becomes everyone's concern and therefore no one's responsibility, and the organization accumulates pilots instead of capability. The ownership question is not bureaucratic. It is the thing that determines whether governance is a living system or a document no one reads.
It is worth being honest that naming an owner is harder than it sounds, because AI cuts across the territory of clinicians, executives, and IT, and each has a legitimate stake. That is precisely why ambiguity is so common and so costly. The resolution is not to hand the decision to whoever is most enthusiastic about the technology. It is to anchor ownership where the clinical consequences land, with a leader who has both the authority to change workflows and the standing to set an accuracy standard the organization will respect. Shared enthusiasm is not the same thing as clear accountability, and only the second one scales.
Before the model selection, before the pilot, before the integration roadmap, the question to settle is ownership. Who is accountable for AI strategy, and do they have the authority and the clinical grounding to govern it across the organization? Settling that question is unglamorous next to the technology, and it is more determinative of the outcome than any single capability choice. The systems that get the most from AI will not necessarily be the ones with the best model. They will be the ones that knew, from the start, exactly who was driving.
