By Brent Dover, Chief Executive Officer, Carta Healthcare
When a new technology gets good enough to take over consequential work, there is always a moment when full automation looks inevitable. The machine is faster, cheaper, and tireless, and the humans in the loop start to look like a transitional inconvenience. It is worth noticing that this moment has arrived before, in several industries, and that it almost never resolved the way the inevitability narrative predicted.
The pattern that actually repeats is more interesting. Automation increases dramatically, and oversight remains embedded in the architecture. Not as a courtesy to displaced workers, but because the systems that survive contact with the real world are designed around what happens when something goes wrong.
Modern aircraft can largely fly themselves. Autopilot is extraordinarily precise, and automation has expanded across the cockpit for decades. Yet pilots never left. The reason is not nostalgia. Aviation is not optimized for maximum automation. It is optimized for risk-adjusted performance. Redundancy, escalation capability, and clear accountability stay embedded because the governing question is not whether the system can usually fly the plane. It is what happens in the rare event that it cannot. A discipline organized around that question keeps a human accountable by design.
Radiology offered an even cleaner test, because imaging is exactly the kind of perceptual task models do well. AI can detect nodules and flag abnormalities with impressive sensitivity, and as it improved, the replacement predictions grew loud. Radiologists were not eliminated. The model became a second reader that flags and prioritizes, while clinicians retain interpretation and legal responsibility. The regulatory and legal frameworks did not dissolve because the technology got better. The question of who is accountable when the read is wrong never went away, and so neither did the radiologist.
Finance ran the experiment closest to pure automation. Algorithmic trading came to dominate execution speed, far beyond human reaction time. Then volatility events exposed systemic risk, and the industry did not respond by removing humans. It added governance. Kill switches, monitoring desks, compliance layers, and circuit breakers became part of the architecture. Speed without guardrails proved dangerous in a setting where errors cascade. The lesson was that the more automated and interconnected a system becomes, the more governance it needs, not less.
Even in lower-stakes settings the pattern holds. Early assumptions held that chatbots would replace support agents outright, and for password resets and order status, they did. But when issues became nuanced, emotionally charged, or financially sensitive, full automation eroded trust. The systems that worked best used AI to handle routine inquiries at scale while designing seamless escalation to a person for the hard cases. The dividing line was not the technology. It was the stakes of getting it wrong.
Place clinical data abstraction beside these examples and the family resemblance is obvious. Registry data feeds reimbursement, accreditation, public reporting, and outcomes research. Errors do not stay contained. They cascade into dashboards, quality metrics, accreditation reviews, and payment. The environment is exception-heavy, full of incomplete documentation, ambiguous timestamps, conflicting records, and registry rule nuance. It is, in other words, precisely the kind of high-consequence, exception-rich domain where every other field chose automation with accountability rather than automation alone.
This reframes a debate that often gets stuck on capability. The interesting question was never whether the model is good. In aviation, radiology, and finance, the models got very good, and the humans stayed anyway. They stayed because accountability is a property of the system, not the model, and because institutions facing asymmetric downside risk will always demand governance where the cost of failure is high.
What all of this suggests is that Hybrid Intelligence in healthcare is not a transitional compromise waiting to be automated away. It is the same durable design that high-stakes industries keep converging on once the early enthusiasm meets the real distribution of cases. Automation handles the routine and the scalable. Accountable experts own the ambiguous and the consequential. Escalation and defensibility are built into the architecture rather than bolted on after an incident.
There is a useful way to see why this pattern is durable rather than coincidental. Different kinds of companies optimize for different things, and the objective function shapes the architecture. A horizontal, AI-only company optimizes for automation rate, scalability, and margin. A system built for regulated clinical abstraction optimizes for sustained accuracy in the ninety-eight to ninety-nine percent range, audit readiness, registry-specific nuance, governance workflows, and institutional trust. Those are not the same goals, and they do not produce the same design. One is built to maximize throughput. The other is built to be defensible when someone with authority examines the output line by line.
This is why even a dramatic improvement in raw model capability does not collapse the distinction. Suppose an AI-only approach eventually matched the extraction performance of a system designed for accountability. It would still have to solve for embedded oversight, inter-rater reliability management, adjudication processes, and integration into the operational fabric of a health system. Intelligence can be replicated. The risk-governed operating system around it is much harder to build, because it is made of workflow, accountability, and trust rather than model weights.
The cross-industry record is really a record of this same realization arriving field by field. Aviation, radiology, and finance all discovered that the hard part was never the capability. It was the architecture of accountability that capability has to live inside. Healthcare is arriving at the same place, and it has the advantage of being able to see the pattern clearly in the industries that got there first.
It is worth naming why this matters for healthcare specifically, beyond the comfort of analogy. In each of these industries, the move to embed accountability was not driven by sentiment about preserving jobs. It was driven by the asymmetry of the downside. When the cost of a rare failure dwarfs the savings from automating the routine, rational institutions invest in governance, because the expected value of preventing the failure exceeds the expected value of the marginal efficiency. Clinical data abstraction has exactly that asymmetry. The savings from automating routine cases are real, and the cost of a wrong value that reaches an audit or a benchmark can dwarf them. The architecture follows from the math, not from nostalgia.
The history here is not a warning against AI. Every one of these industries automated aggressively and benefited enormously. The point is narrower and more useful. In domains where being wrong is expensive and someone has to answer for the outcome, the winning architecture has been remarkably consistent. Healthcare has every reason to learn from it rather than relearn it the hard way.
