By Greg Miller, Vice President of Marketing & Business Development, Carta Healthcare
There is a common assumption baked into how people talk about AI in healthcare, and it usually goes unexamined. The assumption is that any approach combining AI with human experts is a transitional arrangement, a waystation we tolerate until the models are good enough to go it alone. Under this view, human oversight is scaffolding, useful for now and destined to come down. It is worth challenging that assumption directly, because in regulated healthcare it is not just incomplete. It is backwards.
The fair version of the skeptic's case is straightforward. AI models are advancing rapidly. Automation rates are climbing. The cost of intelligence is falling. If clinical abstraction is fundamentally about reading charts and extracting data, then full automation should eventually win, and any approach that keeps humans involved should eventually lose. Stated that way, it sounds reasonable, and it deserves a rigorous answer rather than a defensive one.
The answer is that the premise is wrong. Clinical abstraction is not simply a model capability race, because it does not operate in isolation. It operates inside a regulated, audited, high-consequence system where the data feeds reimbursement, accreditation, public reporting, and population health. When the data is wrong, the result is not a technical miss. It is audit exposure, financial consequence, and reputational risk. That reality changes the optimization function entirely. The architecture that wins in an exception-heavy environment is the one designed around exception handling and accountability, not the one designed around raw throughput.
It helps to be precise about what Hybrid Intelligence actually is, because the casual description, AI plus humans, undersells it. It is not two resources layered together for comfort. It is a control architecture with a deliberate division of responsibility. AI handles pattern recognition at scale, temporal synthesis across fragmented documentation, and routine case throughput. Human experts handle ambiguity, contextual interpretation, edge cases, and final accountability. The system is built around escalation and defensibility rather than pure automation, because clinical abstraction is full of the unusual rather than the uniform. Incomplete documentation, ambiguous timestamps, conflicting records, and subtle comorbidities buried in unstructured notes are common, not rare.
Seen this way, the human role is not a safeguard against an immature model. It is a structural component that remains necessary even as the model matures, in the same way that aviation, radiology, and finance kept accountability embedded in their systems long after the technology became excellent.
The deepest reason Hybrid Intelligence is the end state rather than a phase is a mismatch in speed. AI capability and institutional risk tolerance do not move at the same pace. Capability can jump in a single release. The willingness of a health system to hand an auditor a number that no clinician validated changes slowly, deliberately, and only with cause. That gap between what a model can do and what an institution is prepared to be accountable for is not a temporary lag waiting to close. It is a durable feature of regulated, high-consequence work. Hybrid Intelligence lives inside that gap, and the gap is not going away.
Even if an AI-only approach eventually matched the raw extraction performance, it would still have to solve for embedded oversight, inter-rater reliability management, adjudication, and integration into the operational fabric of a health system. Intelligence can be replicated. A risk-governed operating system is much harder to replicate, and it is the part institutions actually depend on.
This is not only a theoretical argument. The people doing the work say the same thing. In the field, ninety-seven percent of healthcare professionals say AI should support clinical expertise rather than replace it, and seventy-four percent name misinterpretation of complex clinical data as the top risk of running AI without human oversight. The strong preference for combining AI with clinical teams is not nostalgia or fear of change. It is an accurate read, from the people closest to the consequences, of what high-stakes clinical data actually requires.
Step back and the pattern across every high-consequence field is consistent. Automation increases, and accountability stays embedded. Clinical data abstraction sits squarely in that category, which means Hybrid Intelligence is not a compromise between humans and machines, struck because the machines are not ready yet. It is the architecture that aligns automation with governance in an environment where being wrong is expensive and someone has to answer for the outcome.
If Hybrid Intelligence is the end state, it is worth being precise about where the durability actually comes from, because it is not the model. Models improve, and any advantage based purely on having a better one is temporary by construction. The durable advantage is the system around the model. An organization that owns the workflow, the governance infrastructure, the registry expertise, and the audit interface can let the model underneath evolve indefinitely without losing its footing, because the model was never the load-bearing element. The system is.
This reframes the competitive question in a way that should reassure rather than worry. The risk is not that AI gets better, which it will and should. The risk would be reframing abstraction internally as a pure technology function and assuming governance is something an organization can pick up casually on the side. Capability without an operating system around it is the configuration that produces expensive surprises. Capability inside a risk-governed system is the configuration that compounds, because every improvement in the model makes a well-governed operation better without destabilizing it.
There is a through-line from the founding to the present in all of this. The conviction that began with manual data collection at Stanford Children's Hospital, that the clinical stakes demand clinical judgment, is the same conviction that makes Hybrid Intelligence durable now. The technology has advanced almost beyond recognition since then. The reason the human stays at the helm has not changed at all, because it was never about the limits of the machine. It was always about the nature of the work.
This is also why the durability of the model is the wrong thing to worry about and the durability of the system is the right one. A health system that ties its strategy to having the best model will be unsettled every time a better one appears, which will be often. A health system that builds the governance, the registry expertise, and the accountability around the model can welcome every improvement as an upgrade rather than a threat, because the part that creates trust was never the model in the first place. The model is the engine. The system is the vehicle, and it is the vehicle that carries an organization somewhere it can defend.
That is why it is worth stating without hedging. Hybrid Intelligence is the durable end state for regulated healthcare workflows, not a transitional phase on the way to something else. The model underneath will keep improving, and it should. The system that governs it, keeps clinicians at the helm, and answers for the result is the part that lasts. The future of healthcare AI is not autonomous. It is Hybrid, by design and for good.
