Healthcare has no shortage of impressive AI pilot projects. Hospitals have demonstrated that AI can assist with clinical documentation, automate administrative tasks, analyse patient data and improve operational efficiency. The harder part is no longer proving that AI can work; it is proving that it can become a reliable part of everyday hospital operations.
That is the key message emerging from a new survey organised by Carta Healthcare. According to the findings, many healthcare organisations are already seeing measurable benefits from AI, yet they still struggle to move beyond isolated pilots and deploy those capabilities across the wider enterprise.
The AI Pilot Is No Longer The Hard Part
One of the most striking findings is that 71% of healthcare organisations reporting measurable value from AI are still not expanding those initiatives at pace. That suggests the underlying technology is often not the main problem.
Pilot projects usually succeed because they are deliberately controlled. They tend to have a specific use case, a committed project champion, a well-defined workflow and plenty of attention from the organisation. Once that same technology needs to work across several departments, hospitals suddenly have to deal with different clinical processes, competing priorities, budgets, infrastructure limitations and unclear ownership.
In other words, proving that an AI model performs well in one department does not automatically mean the hospital is ready to operationalise it everywhere.
Integration With Existing Clinical Workflows Is Critical
The survey identified EHR integration as the biggest barrier to AI adoption, cited by 44% of respondents. That ranked considerably higher than clinician trust and regulatory concerns, which were each identified by 26%.
The reason is fairly straightforward. A healthcare AI tool can perform brilliantly during a demonstration and still fail in practice if clinicians need to leave their normal workflow every time they want to use it. Every additional screen, login or manual step creates friction, and that friction quickly reduces adoption.
Successful healthcare AI therefore needs to work where clinicians already work. Information should appear inside existing systems at the appropriate point in the workflow instead of forcing doctors, nurses or administrative staff to operate a completely separate application.
Integration also involves more than connecting APIs. Clinical information can be highly contextual, and documentation may vary significantly between departments and practitioners. An AI system has to understand what that information means clinically rather than simply transferring data between platforms.
Hospitals Should Demand Integration Before Buying
This has important implications for CIOs, CMIOs and digital health leaders evaluating AI vendors. Integration should be considered a core requirement from the beginning rather than something to resolve after a contract has already been signed.
Hospitals should ask vendors to demonstrate the technology using their environment, their data and their actual workflows whenever possible. A polished product demo using carefully prepared sample data proves relatively little about how the system will perform inside a busy hospital.
Health systems should also consider who carries the operational burden after deployment. An AI platform that requires significant internal development, constant manual intervention or extensive workflow redesign may create as much work as it eliminates.
The strongest solutions are likely to be those that deliver trusted information directly into existing clinical processes while minimising additional effort for hospital teams.
Clinical Leadership Is Taking Greater Ownership Of AI
Another important change is who now owns AI strategy inside healthcare organisations.
The survey found that clinical leaders are increasingly taking the lead, ahead of IT departments and executive leadership. That is a significant development because it shows AI is gradually moving away from being treated as an experimental technology project.
Once AI influences clinical workflows, the consequences are also clinical. Decisions about where AI should be used, what level of automation is acceptable and when humans must remain involved cannot be made by technology teams alone.
However, the survey also found that 26% of organisations still have no clearly defined AI owner. That is a much bigger concern than it may initially appear.
Without someone accountable for outcomes, a successful pilot can easily lose momentum after the initial excitement disappears. Nobody owns the long-term roadmap, nobody is responsible for resolving workflow problems, and eventually the technology becomes another isolated system that never reaches wider adoption.
AI Governance Needs Clear Accountability
A practical governance structure should involve several parts of the organisation. Clinical leaders should own the clinical outcomes and determine where the technology is appropriate. IT teams should handle integration, infrastructure, cybersecurity and technical reliability, while executive leadership provides investment, strategic oversight and organisational support.
Most importantly, there should be a clearly identified individual or leadership group accountable for whether the AI initiative actually succeeds.
Healthcare AI governance should not become a committee where everyone participates but nobody owns the final outcome.
That ownership becomes even more important as AI starts influencing increasingly important decisions and workflows.
Healthcare AI Vendors Need More Than Good Models
The strongest consensus in the survey concerned vendor selection. Ninety-two percent of respondents said deep clinical expertise is critical when evaluating an AI provider.
That makes sense. Healthcare is filled with ambiguity, incomplete documentation, unusual clinical situations and exceptions that do not fit neatly into a generic workflow.
A powerful general-purpose AI model may understand language extremely well, but that does not automatically mean it understands how a particular clinical process works or why certain information is significant.
Hospitals therefore need to look beyond impressive demonstrations and ask more difficult questions. Has the vendor achieved measurable results at comparable healthcare organisations? Does it understand real clinical workflows? Can it demonstrate performance under normal operating conditions rather than only controlled pilots?
Clinical expertise is increasingly becoming just as important as AI capability.
The Industry Is Moving From Proof Of Concept To Proof Of Operations
Healthcare appears to be entering a more mature stage of AI adoption.
A few years ago, the conversation centred around whether generative AI could safely deliver useful results in healthcare. Now many organisations have enough pilot experience to know that it can.
The next question is whether those capabilities can be operated consistently across dozens of departments, thousands of employees and potentially millions of patient interactions.
That requires significantly more than a good model. Hospitals need integration, governance, training, monitoring, ownership and measurable outcomes.
The shift is essentially from proof of concept to proof of operations.
Healthcare Buyers Are Becoming More Disciplined
The survey also suggests healthcare organisations are becoming less impressed by AI capability for its own sake.
Hospital leaders increasingly want evidence that a solution will reduce manual work, fit naturally into existing workflows and continue producing reliable results after the launch team has moved on to another project.
That is a healthy development.
AI projects should eventually be evaluated like any other major clinical or enterprise investment. What problem does it solve? How much work does it remove? Does it improve outcomes? What does it cost to operate? Can staff actually use it?
The novelty of AI should not exempt it from those basic questions.
Human Trust Still Matters, But Workflow Comes First
Interestingly, clinician trust did not rank as the biggest barrier in the survey. Integration was considerably more prominent.
That does not mean trust is unimportant. Clinicians still need confidence that an AI system is reliable, transparent and appropriate for the task.
But trust is difficult to build when a tool itself is inconvenient to use.
A well-designed AI capability that appears naturally inside an existing workflow, provides useful information at the right moment and allows clinicians to remain in control is much more likely to gain acceptance than something requiring staff to fundamentally change how they work.
Good integration and good clinical design can therefore become part of building trust.
Final Thoughts
Healthcare AI has reached a point where successful demonstrations are no longer particularly unusual. The harder challenge is turning those successes into dependable enterprise capabilities that clinicians actually use every day.
Hospitals that succeed are likely to be the ones that assign clear ownership, prioritise clinical leadership, demand deep integration and choose vendors based on proven operational outcomes rather than impressive demonstrations.
AI itself may already be capable enough for many healthcare applications. The bigger question now is whether healthcare organisations are operationally prepared to use it at scale.
For hospitals, the next phase of AI will not be defined by who runs the most pilots. It will be defined by who can take a useful pilot, integrate it properly, govern it responsibly and make it part of routine clinical practice.
Comments 0