The healthcare industry has reached something it rarely achieves: near-total agreement. Surveys show 97 percent of healthcare professionals believe AI should support clinical expertise rather than replace it. Human review and validation rank as the top driver of trust in AI outputs. The governance question — who stays at the helm, who holds accountability — has been settled faster than most predicted.
But consensus on governance doesn’t mean the harder problem is solved. Nearly 75% of those same respondents cite misinterpretation of complex clinical data as the leading risk when AI operates without clinical oversight.
That number points to something deeper. Misinterpretation is often treated as a problem caused by AI. In reality, it starts much earlier. Clinical data is documented, then translated, then entered into different systems. At each step, details can be lost, changed, or interpreted differently. By the time data reaches a model, it may already be incomplete or inconsistent.
Answering One Question, Surfacing Another
Healthcare organizations have moved quickly toward the right architecture. Hybrid models, where AI accelerates the work and clinicians validate the outputs, are becoming the standard. That’s a good start. But it resolves the governance question while leaving a structural problem underneath largely untouched.
Most AI strategies focus on who reviews the output. That assumes the output is where risk shows up. Increasingly, it isn’t.
The harder question sits one layer deeper: what is that person reviewing against?
Clinical Data is Messy
Clinical data doesn’t arrive clean. By the time information from a patient encounter enters a model or feeds a quality report, it has passed through several layers of translation.
A clinician documents the encounter. That documentation gets interpreted during abstraction. The abstracted data feeds a registry. The registry data flows into analytics. Each step involves judgment calls, definitions that evolve, and systems that don’t always exchange information cleanly. At every step, there is an opportunity for misinterpretation — not as an exception, but as part of the process itself.
That distance between what happened in a care setting and what eventually gets measured creates real risk. A condition documented in a narrative note may not satisfy a registry measure. Timing details that seem incidental to a clinician may determine whether a case counts at all. These aren’t unusual situations. They’re the normal conditions of clinical documentation at scale.
When AI is introduced into this environment, it works with whatever data it receives. If the inputs reflect care accurately, the model performs well. If they don’t, the model will work efficiently toward the wrong answer. Human oversight can catch when the model is wrong. But it cannot correct when the data itself has already drifted from reality.
Quality Scores Often Reflect Data Problems
Health systems know that quality scores reflect documentation accuracy as much as they reflect care quality. Research shows that significant shares of patients lack structured diagnoses in the EHR even when the clinical data to support those diagnoses already exists. Improving problem list completeness doesn’t always change quality scores. Meanwhile, mortality indices and case-mix indices have shifted meaningfully after health systems improved their clinical documentation — without any change in clinical practice or patient population. The metric moved, but the medicine didn’t.
A hybrid model running at speed through high volumes of clinical records will surface and amplify inconsistencies that manual workflows might have caught slowly or absorbed quietly. AI doesn’t introduce misinterpretation. It accelerates how quickly it spreads. That’s not an indictment of the technology. It’s an argument for fixing the layer the technology depends on.
The failure pattern to watch isn’t models producing clearly wrong answers. It’s models producing confident, efficient answers built on data that drifted from reality somewhere upstream.
The Question AI Strategies Tend to Skip
Healthcare organizations are investing heavily in AI governance. Oversight frameworks and validation protocols are being built across health systems at a pace that would have seemed unlikely a few years ago. That work is necessary. But governance applied to the output layer doesn’t compensate for problems in the input layer.
A clinician reviewing an AI result can only assess what they’re shown. If the underlying data was incomplete or didn’t align with measure requirements before the model processed it, the validation step isn’t a safeguard. It becomes a review of a representation, not a verification of reality.
The organizations building durable AI strategies aren’t just adding oversight. They’re shortening the distance between care delivery and the data used to measure it. They’re treating data capture as part of the quality infrastructure, not a precondition for it. The standard for AI performance can’t be “the output looks reasonable.”
It has to be “the output reflects what actually happened.”
The Next Frontier
Healthcare’s AI conversation will keep advancing. Models will improve, automation will expand, and the governance frameworks being built today will mature. All of that matters.
But the limiting factor in healthcare AI isn’t model capability, and it isn’t oversight design. It is data fidelity: how accurately clinical reality survives the journey from documentation to data. Keeping clinicians at the helm is the right answer to the governance question.
The harder question is whether the data those clinicians rely on is trustworthy enough to scale. Until that is addressed, AI will continue to operate with precision on top of approximation.
About Brent Dover
Brent Dover is the CEO at Carta Healthcare, the leader in enterprise clinical data management. With a deep commitment to improving healthcare outcomes, Brent leads Carta Healthcare’s mission to unlock the power of clinical data for hospitals, health systems, and life sciences organizations.
His strategic vision drives the company’s focus on leveraging advanced artificial intelligence and expert clinical professionals to streamline data abstraction, enhance operational efficiency, and generate actionable insights that ultimately elevate patient care. Under his leadership, Carta Healthcare has experienced significant growth and recognition for its innovative approach to solving complex healthcare data challenges.
