Health AI risks stalling in ‘partial adoption trap’

Health systems may fail to capture the biggest productivity gains from artificial intelligence (AI) because doctors often adopt tools in limited ways that help individual workflows but do not transform care pathways, according to a new study submitted on arXiv.

The working paper, The partial adoption trap: Coordination failure, trust, and cultural lock-in in health AI adoption, develops an evolutionary game theory model to explain why point-solution AI spreads more easily than system-change AI, even when the latter offers greater value for hospitals, clinics and patients.

Why health AI adoption can look successful while transformation fails

AI tools are spreading across health systems, including diagnostic support tools, administrative automation, clinical decision aids and ambient voice technologies that document consultations in real-time. Conventional adoption metrics may suggest progress because clinicians are using the tools. However, the paper argues that usage alone can conceal a deeper failure: the system may adopt the technology without changing the clinical pathway it was meant to improve.

The key difference is between point-solution AI and system-change AI. Point-solution tools help clinicians perform a specific task better or faster. Their benefits are usually visible to the individual user, so adoption can spread through normal incentives. System-change AI is different. It is designed to restructure workflows, appointment patterns, referral pathways, documentation practices and administrative coordination. Its benefits often appear only when enough clinicians adopt the tool deeply enough for the wider system to reorganize around it.

That difference creates the partial adoption trap. In the model, doctors choose among three strategies: genuine adoption, partial adoption and rejection. Genuine adoption means integrating AI into clinical practice in a way that changes fundamental working patterns. Partial adoption means using the tool for immediate private gains, such as saving time on documentation or decision support, without changing the broader workflow. Rejection means not using the tool.

Partial adoption is a critical issue - it's not the same as refusal. Doctors may appear to be using an AI tool, and adoption rates may look high. Yet the technology still fails to produce system-wide benefits because too few clinicians change their workflows in a coordinated way. The result is a health system that has adopted AI in form but not in function.

Ambient voice technology illustrates the problem. AI scribes can reduce documentation burden and help doctors save time during consultations, but if each doctor uses the tool only to ease paperwork while keeping existing appointment structures unchanged, the system does not gain the full productivity benefit. To unlock pathway-level gains, enough clinicians must restructure clinic flow, booking systems, administrative support and patient pathways together.

The paper argues that this threshold structure is not captured by traditional technology adoption theories. Models that treat adoption as a binary individual choice fail to explain why technologies can be widely installed but weakly embedded. In system-change AI, the outcome depends not only on whether doctors use the tool, but whether enough of them use it in the deeper way required for collective transformation.

This makes partial adoption a stable outcome, not a temporary implementation flaw. Doctors can rationally choose partial use because it offers private benefits with lower disruption costs. Genuine adoption is more costly because it requires changing routines and bearing workflow friction. Unless enough colleagues also change, the genuine adopter may carry costs without receiving the system-level benefits.

The model shows that health AI adoption can become bistable. One stable outcome is full genuine adoption, where the system crosses the threshold needed for transformation. The other is the partial adoption trap, where the whole population uses AI superficially while systemic gains fail to materialize. In many conditions, partial adoption becomes the default attractor.

Trust, coordination and clinical culture deepen the trap

The study identifies three compounding failure modes behind stalled system-change AI adoption: threshold coordination failure, trust failure and cultural lock-in.

• The threshold failure arises because no individual doctor can deliver system-wide gains alone. Genuine adoption becomes worthwhile only when enough colleagues also restructure their workflows. Below that threshold, partial adoption is safer and more attractive. Doctors can capture immediate benefits without taking on the disruption of broader change. This creates a coordination problem: everyone may be better off if genuine adoption succeeds, but each individual has reason to avoid moving first.
• Trust: System-change AI often promises productivity gains, but doctors may doubt whether those gains will be shared fairly. If clinicians believe that saved time will simply be converted into more workload, tighter schedules or managerial targets, they have less reason to genuinely adopt the tool. In the model, the organization's inability to credibly commit to sharing productivity gains reduces the perceived value of genuine adoption. This trust problem is especially severe for high-value AI technologies. The larger the possible productivity gains, the stronger the organization's temptation to capture them. That creates a self-reinforcing cycle. Doctors expect the organization to renege on promised benefits, so they avoid genuine adoption. Because genuine adoption fails, the system never fully realizes the gains that could have rebuilt trust.
• Cultural lock-in: Clinical adoption is not just an individual calculation. Doctors observe peers, imitate successful patterns and respond to professional norms. If partial adoption becomes the accepted local behavior, genuine adopters may face friction for deviating from the group. They may be seen as disrupting established workflows, increasing pressure on colleagues or aligning too closely with managerial demands.

The model shows that such negative coordination norms make the partial adoption trap deeper over time. Once a culture of limited use forms, it becomes self-reinforcing. Doctors who might otherwise adopt genuinely are pulled back toward the dominant norm. The longer partial adoption persists, the harder it becomes to shift the system toward genuine transformation.

The paper also introduces a cost ratchet dynamic. Failed adoption attempts are not always wasted. If a system temporarily crosses the adoption threshold, even without achieving lasting embedding, the disruption cost of future genuine adoption can fall. Staff may learn new routines, administrative systems may adjust, and some barriers may be permanently lowered. Repeated attempts can therefore make future success more likely.

The benefit of a cost ratchet depends on whether implementation learning happens faster than trust erosion. If failed AI rollouts quickly damage clinician confidence, the trust loss can outweigh the reduced implementation cost. In that case, repeated imperfect pushes for adoption may make future attempts harder, not easier.

This is why the model predicts what it calls a value-adoption paradox. The AI tools with the greatest system-wide value are often the hardest to adopt deeply. Point-solution tools spread because individuals can capture benefits directly. System-change tools require threshold coordination, credible organizational commitments and cultural alignment. The more transformative the technology, the more severe the adoption barriers become.

Overall, a standard policy may appear to work for point solutions while failing for the technologies that matter most for productivity, access and pathway reform. A hospital may successfully deploy tools that help individual clinicians while repeatedly failing to embed tools that would reshape care delivery at scale.

Policy must move beyond individual incentives

The study states that standard health AI adoption policy is poorly matched to system-change technologies. Health systems often focus on broad individual incentives, tool access, training sessions and adoption targets. Those measures may increase visible use, but the model predicts they will often produce partial adoption rather than genuine transformation.

Escaping the trap requires addressing all failure modes in the right sequence. Trust architecture must come before adoption requirements. Clinicians need credible commitments about how productivity gains will be used and shared. If doctors believe AI-driven efficiency will only increase workload or reduce professional autonomy, they will have reason to resist deep workflow change.

Those commitments cannot be vague. The model suggests that organizations need verifiable sharing rules, protected implementation budgets, leadership continuity and transparent reporting on where productivity gains go. Contractually fixed arrangements can remove the trust game that otherwise undermines adoption.

Cultural readiness must come before individual incentives. If clinical teams are not prepared for workflow change, incentives may push some doctors toward genuine adoption while leaving them isolated within a partial-adoption culture. That can increase deviance costs and strengthen resistance. The paper argues that clinical champions, team-level preparation and shared expectations should precede the rollout of system-change tools.

The model warns against dispersing early adopters across a system. Instead, adoption should be seeded at the clinical-team level, where enough doctors can cross the threshold together. Team-based adoption can create local genuine-adoption clusters, allowing booking systems, administrative staff and clinical routines to change coherently.

Threshold subsidies should be concentrated, not universal. The paper argues that support should target the fraction of doctors needed to push a local system above the tipping point. Once the threshold is crossed and systemic benefits begin to appear, the need for subsidy falls. This approach differs from broad incentives that reward shallow use without ensuring pathway transformation.

Implementation support should be concentrated in the embedding window, when effective adoption has crossed the threshold and workflow costs can fall. That is the period when health systems should invest heavily in operational support, process redesign, administrative alignment and troubleshooting. Spreading support evenly across a rollout may miss the point at which it matters most.

The study's welfare analysis warns that the costs of partial adoption are often invisible. Traditional adoption rates measure how many clinicians use a tool, but they do not distinguish between genuine and partial adoption. A system can report high uptake while effective adoption remains low. For system-change AI, the correct metric is not tool use alone but whether enough clinicians have changed practice to produce systemic benefit.

If health systems reward installation, login rates or superficial use, they may lock in the very pattern that prevents transformation. AI then becomes another layer of digital activity without the promised reduction in waiting times, duplication, transaction costs or workforce burden, the study concludes.