By Betsy Castillo, Vice President, Clinical Data Abstraction, Carta Healthcare
Healthcare organizations are rushing to adopt AI, and clinical data abstraction is a natural target because the promise of cost savings is real. A growing number of vendors now offer AI abstraction tools that promise to accelerate the work while cutting costs. The promise is attractive. It is also, too often, built on a quiet assumption that turns out to be wrong, which is that you can build AI for healthcare without deep input from the people who actually do clinical work.
The blunt version of the lesson is this. Good AI does not mean good abstraction. A model can be fast, fluent, and impressive and still be unsuited to the complex, high-stakes world of clinical registries, because speed and fluency are not the same thing as clinical correctness. It does not matter how quickly a model answers if it answers incorrectly, and without clinical input, wrong answers are not an occasional risk. They are inevitable.
There are two main reasons clinical expertise is so often absent from these tools. The first is naivety, the mistaken assumption that you can develop AI for healthcare without people who have clinical backgrounds, as if abstraction were a generic text problem. The second is cost. Bringing nurses and physicians onto a team to consult during product development and testing is expensive, and for many startups that expense is a genuine obstacle. The result is a market full of tools that are, underneath a polished interface, retooled general-purpose models trained on generic data, with little clinical grounding in how they were built.
Technology alone is not a solution in the healthcare realm. You need clinical understanding for the technology to work, and that understanding has to be present in two places, in how the tool is developed and in how its output is overseen. A tool built without clinicians will encode misunderstandings that no amount of after-the-fact checking fully removes. A tool used without clinicians will produce confident answers that no one qualified is positioned to question.
Consider a real scenario. Clinicians asked an AI model for a patient's most recent ejection fraction, a critical measure of heart function. The model returned three documents from the record, each with a different value, all technically correct. Take the most recent and there appeared to be no cause for concern. An experienced clinician knew better, because an earlier and lower reading was the very reason the patient was undergoing a procedure. The values were all accurate. Only one made sense in the clinical context, and that nuance is exactly what a model cannot capture on its own.
Another example. Asked whether a patient returned to the operating room, an AI model answered yes, based on the patient's post-surgical visit to an endoscopy suite. The suite is affiliated with the operating room, but it is not the same thing, and the distinction changes the meaning of the record. A clinician noticed the misinterpretation immediately. The model had retrieved a real event and drawn the wrong conclusion from it, which is precisely the failure that pure retrieval invites.
The pattern is consistent. AI can process data, but it cannot, on its own, infer or apply judgment. A random lab value has no meaning to a model until someone puts it in context. The model gives you information. Turning that information into a correct answer requires clinical knowledge, and that is the step where good AI, by itself, falls short.
Clinical judgment matters most where the record is silent. Suppose the documentation does not confirm that a particular physician gave a patient aspirin. A purely literal reading concludes the patient did not receive it. A clinician who knows from experience that this physician always gives aspirin recognizes a documentation gap rather than a clinical absence. Both readings are defensible from the text alone. Only one is correct, and only experience distinguishes them. Left to itself, a model will conclude that the patients are not getting aspirin, simply because the confirming note is missing.
This is also where abstraction quietly improves care rather than merely recording it. When a clinician spots the gap, the documentation can be corrected and definitions clarified, and the underlying record gets better. A provider with imperfect documentation still benefits, because the process forces the organization to evaluate definitions and validate inputs. Good AI can help by highlighting inconsistencies, but it takes a human to recognize the pattern and act on it.
The clinicians using these tools describe the same truth from the inside. A nurse with more than three decades of experience wrote about her own transition to an AI-assisted workflow, and her initial questions were not about whether the technology was impressive. They were whether AI could grasp the way clinicians think and document, and whether it would notice the small details. What changed her mind was not raw speed. It was discovering that the tool surfaced things she might have missed, a lab value buried deep in a note, a discharge medication mentioned in the documentation but left out of the summary, each answer linked to its source. The tool earned her trust by extending her judgment, not by substituting for it.
She is equally clear about the limit. There are moments, she wrote, when something in a chart simply does not feel right, an instinct honed over years that no system can replicate. That sentence is the entire argument in miniature. The value of a good tool is that it clears away the routine so this kind of judgment can be spent where it matters most. The danger of treating good AI as good enough is that it quietly removes the very instinct that catches what the model cannot.
Some vendors view abstraction as a process of data retrieval, a matter of matching labels to values. It is far more than that. Abstraction is clinical storytelling. Quality abstraction translates raw clinical data into a coherent view of a patient's journey, and doing that well requires interpreting registry questions, knowing where relevant data tends to hide, and recognizing when documentation is incomplete or misleading. Those are clinical acts, not extraction steps, and they are the reason a credentialed abstractor sits at the center of the work rather than at its edge.
None of this is an argument against AI. Used well, it is a powerful instrument that makes skilled people faster and even improves the documentation it draws from. The argument is narrower and more important. An instrument is only as good as the expertise guiding it. The vendors and health systems that build clinical judgment into both the tool and its oversight are the ones who turn good AI into good abstraction. The rest will keep mistaking speed for quality, until an audit or a single misread chart reminds them of the difference.
