Why AI in Healthcare Often Fails to Deliver System-Wide Change

Artificial intelligence is frequently hailed as a solution to the growing pressures on healthcare systems, from capacity constraints to rising costs. Many expect AI to not only streamline individual tasks but to fundamentally transform how healthcare is delivered at a system level. However, evidence suggests that while AI can boost productivity in specific, controlled settings, it often fails to produce measurable system-wide improvements. A new paper offers a compelling explanation for this gap: the limiting factor is not the technical capability of AI, but how it interacts with the incentive structures that govern coordination among healthcare providers. This perspective shifts the focus from optimizing tasks to understanding the game-like dynamics that shape behavior in complex systems.

Researchers have identified three archetypal forms of AI deployment in healthcare, each with distinct for system outcomes. The first type, effort-reducing AI, includes tools like ambient voice documentation or AI-assisted discharge summaries that lower the time and effort required for coordination tasks. The second, observability-oriented AI, encompasses technologies such as predictive models for discharge timing or analytics that make delays more visible. The third, mechanism-level AI, involves interventions that restructure how local actions map to consequences, such as systems that absorb downside risks for providers who expose capacity. The key finding is that only mechanism-level AI can reliably change system-level behavior by altering the underlying incentives, while the other two types often leave stable patterns unchanged despite improving local efficiency.

Ology employs a game-theoretic framework to analyze these archetypes through a stylized example of inpatient capacity management. In this scenario, individual wards choose between exposing available capacity to the wider system or buffering it locally. Exposing capacity can improve overall patient flow but carries higher local costs, such as increased workload and risk, while buffering protects local interests but may contribute to system congestion. Using a minimal formalization, the paper models this as a coordination game where each ward's payoff depends on both their own action and those of others. The analysis then examines how each AI archetype modifies this baseline game, focusing on whether it changes the best-response strategies that define stable equilibria.

From the analysis show that in the baseline game without AI, buffering is a Nash equilibrium because it is individually rational for each ward to protect local interests when others do the same, even though collective exposure would be socially preferable. Effort-reducing AI, by lowering the cost of actions without altering the relative disadvantage of exposure, does not change this equilibrium; the inequality favoring buffering persists. Observability-oriented AI adds expected consequences to buffering when detected, but unless these consequences outweigh the baseline advantage, the equilibrium remains stable. In contrast, mechanism-level AI bounds or redistributes the local cost of exposure, making it individually rational for wards to expose capacity. This shifts the equilibrium to cooperative behavior, as shown in the condition where exposure becomes a best response to buffering across all actors.

Of this analysis are profound for healthcare leaders and policymakers. It explains why many AI deployments, despite delivering real value at the task level, fail to produce system-wide transformation: they optimize within existing incentive structures rather than changing them. For example, AI tools that reduce documentation effort may improve clinician productivity but do not address the coordination failures that lead to capacity bottlenecks. Mechanism-level interventions, though harder to implement due to requirements for system-level ownership and risk redistribution, offer a path to genuine change by making cooperation individually rational. This suggests that procurement decisions should prioritize AI that reshapes incentives, not just workflows, to achieve broader goals like improved patient flow and reduced costs.

Limitations of the analysis include its reliance on a simplified, static model that may not capture all complexities of real-world healthcare coordination. The paper acknowledges that actual scenarios often involve asymmetric costs, veto power, or threshold effects, which could alter dynamics but do not invalidate the core insight about incentives. Additionally, the analysis is theoretical and based on a minimal example, requiring empirical validation in diverse healthcare settings. Future work could extend this to evolutionary game theory to account for adaptive behaviors over time, but the current framework provides a crucial lens for distinguishing between AI interventions that merely optimize within equilibria and those that can shift them.