📈 economics

The partial adoption trap: Coordination failure, trust, and cultural lock-in in health AI adoption

This paper employs an evolutionary game theoretic model to demonstrate that health AI systems often become trapped in a suboptimal partial adoption equilibrium due to the interplay of coordination failures, trust deficits, and cultural norms, a phenomenon that standard incentive-based policies fail to address and which paradoxically affects the most transformative technologies the most severely.

Original authors: Ari Ercole

Published 2026-05-19

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Ari Ercole

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Half-Hearted" Problem

Imagine a hospital decides to install a new, super-smart AI system to help doctors. The goal isn't just to make one doctor's job easier; the goal is to completely reorganize how the whole hospital works so that patients get seen faster and costs go down.

The paper argues that while doctors are happy to use the AI for small, easy tasks (like typing up notes), they often refuse to change their entire workflow to make the system work as intended. The result? The AI is "installed" everywhere, but the hospital never actually gets the big benefits it promised. This is the Partial Adoption Trap.

The author uses a mathematical model (like a game theory simulation) to show why this happens and why standard solutions (like giving doctors bonuses) often make it worse.

The Three Ways Doctors Play the Game

The model suggests doctors have three choices when a new system arrives:

Genuine Adoption (The "All-In" Player): The doctor completely changes their routine to fit the new AI. They change how they see patients, how they schedule appointments, and how they talk to staff.
- The Catch: This is hard work and disruptive at first.
Partial Adoption (The "Smart Saver"): The doctor uses the AI to do the easy stuff (like saving time on paperwork) but keeps their old routine for everything else.
- The Catch: This feels good immediately because it saves time without the hassle of changing the whole system.
Rejection (The "Luddite"): The doctor ignores the tool entirely.

The Trap: Because "Partial Adoption" feels easier and more rewarding in the short term, most doctors choose it. But the system only works if enough people choose "Genuine Adoption." If everyone is "Partially Adopting," the system is like a car with the engine running but the wheels in neutral—it's making noise, but it isn't going anywhere.

The Four Reasons the Trap Happens

The paper identifies four specific reasons why the hospital gets stuck in this "half-hearted" state.

1. The "Too Small to Matter" Problem (Threshold Failure)

Analogy: Imagine trying to start a campfire. You need a certain number of sticks burning at once to create a flame that heats the whole tent. If you only light a few sticks, they just smoke and die out.
The Reality: The AI only creates system-wide benefits (like faster patient flow) if a critical mass of doctors changes their workflow. If you are the only one changing, you are just an outlier fighting the current. The benefits don't kick in until you cross a "tipping point." Since no single doctor can cross that point alone, they wait for others, and nothing happens.

2. The "Broken Promise" Problem (Trust Failure)

Analogy: Imagine a boss tells a team, "If you work harder and save the company money, I'll give you a bonus." The team works hard, saves the money, but the boss says, "Actually, I'm keeping the savings for the company. No bonus for you."
The Reality: Doctors know that if they genuinely restructure their work, the hospital will save a lot of money. But doctors don't trust that the hospital will share those savings with them. They fear that if they do the hard work, the hospital will just take the credit and the cash. So, they choose the "Partial" route where they keep their own time savings, rather than risking the "Genuine" route for a promise that might be broken.

3. The "Old Habits Die Hard" Problem (Cultural Lock-in)

Analogy: Imagine a group of friends where everyone agrees that "we don't do that new dance." If one person tries to do the dance, the others make fun of them or ignore them. Eventually, the person stops dancing to fit in.
The Reality: If most doctors are "Partially Adopting," the culture of the team becomes "we use the tool, but we don't really change how we work." If a doctor tries to go "Genuine," they face social pressure and friction from colleagues who aren't on board. The group culture locks everyone into the mediocre "Partial" state.

4. The "Paradox of Value"

Analogy: The more valuable a treasure is, the harder it is to get people to dig for it together.
The Reality: The paper finds a cruel irony: The technologies that have the biggest potential to save lives and money (System-Change AI) are the hardest to get adopted. The technologies that are easy to adopt (Point-Solution AI, like a simple spell-checker) are the ones that spread easily. The more valuable the change, the more likely it is to get stuck in the trap.

The "Cost Ratchet": A Glimmer of Hope

There is one interesting twist in the model called the Cost Ratchet.

Analogy: Imagine trying to push a heavy boulder up a hill. It's incredibly hard at first. But if you push it just a little bit past a certain point, the ground changes, and the boulder becomes easier to push forever. Even if you stop pushing and the boulder rolls back down, the ground stays changed. The next time you try, it's easier than the first time.

The Reality: If a hospital tries to implement the AI and fails (they don't reach the tipping point), they might think they wasted their time. But the paper argues that even a failed attempt "decays" the cost. The doctors learned a little, the system got a little used to it. The barrier to entry is permanently lower.

However: This only helps if the doctors don't lose trust too quickly. If the hospital breaks promises too fast, the doctors' trust erodes faster than the "learning" happens, and the trap gets deeper.

How to Escape the Trap (The Solution)

The paper argues that standard policies (like giving every doctor a small bonus to use the tool) actually make the trap worse because they encourage "Partial Adoption" without fixing the underlying trust or culture issues.

To escape, the hospital needs a specific sequence of actions:

Fix Trust First: Before asking doctors to change, the hospital must sign a contract guaranteeing that if the system saves money, the doctors will get a share. No trust, no change.
Prepare the Culture: Before introducing the tech, get the teams ready. Make sure the "group norm" supports change so that the first adopters aren't bullied by the status quo.
Seed the Team, Not the Individual: Don't spread the change thin across the whole hospital. Pick one specific team (a "seed") and get everyone in that team to go "Genuine" at the same time. This helps them cross the tipping point locally, creating a small fire that can eventually spread.
Support the "Embedding" Phase: When the team finally crosses the threshold, pour all the support resources into that moment to make sure the new way of working sticks.

The Bottom Line

The paper concludes that we cannot treat "System-Change AI" like a simple software update. If we try to force it on doctors with individual incentives and no trust, we will get a hospital full of doctors using the tool to type notes faster, while the hospital remains just as slow and expensive as before. We are stuck in a trap of "looking like we changed, but not actually changing."

Technical Summary: The Partial Adoption Trap

Problem Statement
The paper addresses a paradox in health artificial intelligence (AI) adoption: while "point-solution" tools (which improve individual clinician performance on specific tasks) diffuse readily, "system-change" AI (intended to restructure care pathways and generate systemic productivity gains) consistently stalls at a state of partial adoption. In this state, tools are formally installed and used for immediate private benefits (e.g., time savings on documentation) but fail to trigger the collective workflow restructuring required for systemic transformation. Standard diffusion-of-innovations and technology acceptance models fail to explain this, as they treat adoption as an individual, binary decision driven by perceived usefulness, ignoring the collective action structure where systemic benefits only materialize above a population threshold.

Methodology
The author employs an evolutionary game theoretic (EGT) framework to model clinical AI adoption in a large population of doctors. The model is justified by the observation that clinical adoption decisions are often boundedly rational, driven by peer imitation rather than complex optimization.

Strategy Space: Doctors choose among three strategies:
1. Genuine Adoption (G): Restructuring fundamental working patterns to contribute to systemic benefits.
2. Partial Adoption (P): Using the tool for immediate private gains without restructuring workflows.
3. Rejection (R): Non-adoption.
Dynamics: The system evolves via replicator dynamics, where strategy frequencies change based on relative fitness (payoffs).
Key Mechanisms:
- Threshold Benefit: Systemic benefits ( $\Phi$ ) are modeled as a sigmoid function of "effective adoption" ( $e = x_G + \gamma x_P$ ), materializing only when $e$ exceeds a critical threshold $e^*$ .
- Cost Ratchet: The disruption cost of genuine adoption ( $c$ ) is state-dependent. It decays exponentially only when effective adoption exceeds the threshold ( $e > e^*$ ), reflecting that environmental frictions dissolve only when the new workflow becomes the norm. Crucially, this cost decay is irreversible.
- Trust Game: The appropriability of productivity gains ( $\alpha$ ) is endogenized through a trust game between doctors and the organization. The organization may renege on sharing gains, creating a self-confirming trust trap.
- Coordination Norms: The model incorporates social coordination effects, where negative norms against genuine adoption (deviance costs) can reinforce partial adoption.

Key Results

Bistability and the Partial Adoption Trap: Under plausible parameters, the system is bistable. It possesses two locally asymptotically stable equilibria: full genuine adoption and the partial adoption trap (where the entire population adopts partially). The partial adoption trap is the generic attractor for most initial conditions. A separatrix divides the state space; below it, the population drifts toward the trap, and above it, toward full adoption.
The Cost Ratchet: Failed adoption attempts that temporarily push effective adoption above the threshold permanently lower the disruption cost for future attempts, even if embedding ultimately fails. This creates path dependence where "failed" pilots reduce future barriers.
Trust-Trust Interaction: The model identifies a trust trap where rational anticipation of organizational reneging prevents genuine adoption, which in turn prevents the realization of gains that would allow belief revision. The benefit of the cost ratchet is offset if trust erosion (belief updating) occurs faster than cost decay.
Cultural Lock-in: Negative coordination norms among doctors (social costs for deviating from the partial adoption norm) deepen the trap and contract the basin of attraction for genuine adoption, independent of individual financial incentives.
The Value-Adoption Paradox: A technology type index ( $\rho$ ) is defined, ranging from pure system-change ( $\rho=0$ ) to pure point-solution ( $\rho=1$ ). The model demonstrates that technologies with the highest systemic value (lowest $\rho$ ) generate the deepest partial adoption traps, the strongest temptation for organizational reneging, and the most damaging coordination dynamics. Thus, systemic value and adoptability are negatively correlated.
Welfare Analysis: The welfare loss from the partial adoption trap is largest for high-value system-change AI. Standard metrics (total adoption rate $x_G + x_P$ ) mask this failure, as they appear high in the trap, whereas the correct metric (effective adoption $e$ ) remains low.

Significance and Policy Implications
The paper argues that standard health system digital transformation policies, which typically rely on individual incentives to address threshold failures, are predicted to systematically produce the partial adoption trap. This results in widespread process compliance without pathway transformation.

To escape the trap, the paper derives a specific policy architecture that must be sequenced correctly:

Trust Architecture: Must precede adoption requirements (e.g., fixed, verifiable sharing contracts).
Cultural Preparation: Must precede individual incentives (e.g., clinical champion programs to reduce deviance costs).
Targeted Seeding: Threshold subsidies should be concentrated on seeding block adoption within clinical teams rather than dispersing incentives universally.
Embedding Support: Implementation support must be concentrated in the "embedding window" (when $e > e^*$ ) to accelerate cost decay.

The paper concludes that applying point-solution policy logic to system-change AI leads not to slow adoption, but to a permanent lock-in of a suboptimal equilibrium that mimics adoption while failing to generate systemic benefit.