Mike Litvinenko is the Founder of Eximion.

Health AI investment has followed a familiar playbook: build models that ingest large datasets, automate workflows and output better predictions. That work matters, but it assumes the hardest part of clinical care is information retrieval.

In practice, the limiting factor is often judgment under pressure: how doctors weigh uncertainty, update a differential diagnosis, and avoid cognitive shortcuts when time is scarce and stakes are high.

If national and system-level AI strategies focus only on automation and prediction, they risk accelerating the easy part of medicine while leaving the most consequential part untouched.

The next phase of health AI should be assessed not only by efficiency gains, but by whether it strengthens clinical reasoning and reduces preventable diagnostic error.

The gap is not data. It’s decision behavior.

Clinical environments force rapid decisions with partial information. In those conditions, doctors fall back on habits: pattern recognition, mental shortcuts and assumptions formed by training and exposure. Those habits can be protective, but they can also create blind spots, especially in complex or atypical presentations where common patterns mislead.

This is why many AI deployments look impressive in pilots and uneven in real-world practice. They streamline tasks but do not change how decisions are made. They optimize the workflow around clinical reasoning rather than investing in the reasoning itself.

Health systems, regulators, funders and procurement bodies are making significant commitments to AI. But the most important question is not whether a model can answer a clinical question faster. It is whether the system can measurably improve decision quality at scale, across settings, specialties and levels of experience.

That requires policy and investment that treats clinical judgment as infrastructure: something to be built, measured and strengthened continuously, not left to informal learning and individual experience.

What health AI strategy should prioritize

If the goal is durable clinical impact, investment needs to shift from technology-first deployments to capability-building systems.

One priority is rehearsal-based training as core infrastructure. Continuing professional development is often treated as a compliance requirement rather than a performance system. Yet rehearsal, repeated exposure to complex scenarios with structured feedback, is how high-stakes professions reduce error.

Health systems can fund environments where doctors practice difficult decisions outside the moment of care, especially for low-frequency, high-impact conditions.

A second priority is evaluation standards that measure reasoning, not only outputs. Many adoption frameworks focus on operational metrics: time saved, notes generated, throughput improved. Those measures matter, but they miss whether AI is making doctors more accurate, more adaptable, and less prone to cognitive bias.

Policymakers and procurement bodies can require evaluation models that track improvements in clinical reasoning over time, including calibration under uncertainty and guideline adherence where it matters most.

A third priority is using AI to introduce challenge, not just convenience. Most tools are designed to reduce friction. But reducing friction can also reduce vigilance.

A more mature strategy invests in systems that surface edge cases, highlight uncertainty and expose disagreement between expert pathways, because that is what trains doctors to think rigorously. Helpful AI is not always safe AI. Challenge can be a safety feature when deployed responsibly.

A fourth priority is treating clinician decision data as a public research asset, with strong governance. The next frontier of health research is not only biological data. It is decision data: how doctors interpret information and where reasoning fails.

De-identified, privacy-protected datasets on clinical decision-making can support research on diagnostic error, training design and system-level risk. This requires governance frameworks, ethical standards for aggregation, and clinician protections that are built in from the start.

What this looks like in practice

A training-first approach is already emerging in vertical AI systems that model decision pathways and provide structured feedback during simulated scenarios. Rather than replacing clinical judgment, these systems aim to make judgment visible, measurable, and improvable through repetition and calibration.

Eximion is one example of this direction, using structured clinical simulations to capture decision patterns and support feedback-based learning. It illustrates a broader investment thesis: AI that strengthens decision-making capacity, not only administrative throughput.

Treating clinical judgment as infrastructure aligns with multiple health priorities: safer care, more equitable outcomes, better use of scarce specialist time, and greater resilience in clinical systems facing workforce strain.

The next generation of doctors will still hold the responsibility for decisions. AI may assist, but it will not carry accountability when complexity, ambiguity and risk converge. If health systems want AI to deliver durable value, they should invest where medicine ultimately succeeds or fails: the quality of human judgment.

The question is no longer whether AI can answer clinical questions. It is whether funders and procurement bodies are willing to prioritize the policies, evaluation standards and training investments that make clinical decision-making measurably better. 

Mary Sahagun, public relations and communications lead for Eximion, contributed communications support to this op-ed.

R$