The Gap Nobody Talks About

There is a version of AI in healthcare that the industry presents at conferences, demonstrates in a boardroom, or describes in press releases. In this version, clinicians move fluidly through intelligent systems, and information surfaces exactly when it’s needed. Decisions are made faster, documentation is lighter, and the cognitive burden of a demanding profession is meaningfully reduced.

Then there is the AI version that gets deployed. Training happens once, maybe twice. A few early adopters engage with the AI, then, quietly, people go back to the way they were doing things before. These people are not resistant to change or technology, but the new system—despite everything it is capable of doing—made their actual jobs harder.

This conundrum describes the adoption gap. Despite rapid investment, AI uptake in healthcare has lagged behind other sectors of the economy, as Thuy Nguyen et al. showed in a 2025 article on the adoption of artificial intelligence in the healthcare sector. In healthcare, this gap costs organizations time and money, and it costs them the thing they intended to improve in the first place.

The AI conversation in healthcare has been dominated by capability: what AI systems can do, how accurate their models are, and how much data they can process. Although this conversation is necessary, it is not sufficient. The difference between what an AI system is capable of and what a healthcare worker will actually use every day is not a technology problem; rather, it is a design problem. Until the industry treats it as one, the pattern will repeat.

The Healthcare Environment

Before we discuss what AI should do in healthcare, it’s worth being honest about situational factors.

Healthcare environments are not clean, optimized, or waiting for a new system to slot neatly into place. Years of decisions made under pressure accumulate; tools are adopted because of convenience and availability, workflows build around the limitations of existing systems, and habits form because consistency, even when imperfect, is safer than experimentation when someone’s health is on the line. According to an article by Asgari et al. (2024), the result is that most healthcare workers operate across a fragmented digital landscape: different logins, different interfaces, and different data sources that don’t talk to each other. A clinician searching for a single piece of patient information might touch three or four systems before finding it. This is not a niche problem but rather the baseline condition into which most healthcare technology is deployed.

Legacy systems persist in this environment not because staff agree the systems are ideal, but because people have learned to use them despite their quirks. Users know what they can trust and what needs to be double-checked. They know how to get the information they need, even when the system makes it harder than it should. That knowledge has real value that doesn’t show up on a product roadmap.

When AI enters the healthcare environment without accounting for legacy practices, one of two things tends to happen. Either the AI becomes another fragmented system layered on top of the existing fragmentation—one more tab to manage, one more workflow to maintain in parallel—or it attempts to replace existing systems without earning the trust that would make replacement feel safe. As Moy et al. documented in 2023, workflow fragmentation is a primary driver of clinician documentation burden.

The real opportunity AI offers in healthcare is not any single algorithm or capability. It is unification, the ability to bring fragmented information into a coherent, navigable workspace. Unification reduces the number of context switches a clinician must make between tasks; the AI makes information findable without making finding a task itself. AI requires technical integration, and more importantly, understanding how people actually work, not how a process diagram indicates they should, but through UX research and human-centered design.

When the System Becomes the Problem

There is a failure in healthcare AI that isn’t discussed enough because it doesn’t appear to be a failure at the high level. The system launches, people use it, and numbers go into dashboards. Yet, something is wrong. The skepticism is well-founded. In a survey of hospital C-suite leaders, just 12% believed current AI algorithms were robust enough to rely on.

The problem is that the AI product isn’t reliable enough to be trusted without verification. Outputs are sometimes accurate and sometimes not; they might not be dramatically wrong, just wrong enough that no healthcare worker with professional judgment and patient responsibility would stake anything on them without checking. So they check every time. And in doing so, healthcare workers are now doing two jobs: their original job in addition to auditing the AI.

This is not an edge case according to Moy et al. Auditing happens when systems are designed around ideal performance rather than real-world performance, such as when the question “Is the AI accurate enough?” gets answered with benchmark data rather than with contextual observation. A system that is 85% accurate in a controlled evaluation might deliver something closer to meaningful friction in a clinical ward because the 15% failure rate isn’t evenly distributed, and healthcare workers have no way of knowing in advance which outputs land in which category. The result is a verification loop that adds cognitive load instead of reducing it. Cognitive load is not an abstract usability concern in healthcare because it is directly connected to the quality of care.

What makes this failure particularly damaging is that it often goes unaddressed for too long. Users don’t file bug reports, and they don’t escalate to product teams. Users adapt and develop their own workaround, which is usually some version of “use the AI for the easy stuff, and ignore it for anything that matters.” Eventually, they stop opening it at all, thus rendering the system shelfware while the dashboard continues to show engagement metrics.

Engagement and trust are not the same thing. This UX lesson is uncomfortable but important: Good design cannot fix fundamentally unreliable technology. Poor design can make unreliable technology significantly worse. An unclear interface on an inaccurate system doesn’t just frustrate users; it accelerates the point at which they give up entirely.

The Rationale for “If It’s Not Broke”

When healthcare workers resist new technology, the instinct in product teams is often to frame it as an organizational change management (OCM) problem, something to be solved with better training, better communication, or better onboarding. Sometimes that’s right, but often, it misses reality.

Consider what it costs a healthcare worker to adopt a new system. There is a time investment, including learning a new interface, building new muscle memory, and understanding a new logic. There is the cognitive investment, including the period in which the new system requires active thought instead of automatic action. And there is risk; that is, the window of time during which errors are more likely, not because of incompetence, but because any new system introduces uncertainty before it supports fluency.

In most industries, that cost is manageable. In margin-thin healthcare, it is borne against a backdrop of patient care, clinical decisions, team communication, documentation demands, and constant interruption. Asking a healthcare worker to absorb the learning curve of a new AI system is not a small request; it is asking workers to take on additional risk in an unforgiving environment. “If it’s not broke, don’t fix it” is not technophobia; it is a rational calculation that people make based on understanding the cost of being wrong in their work.

OCM intersects with user experience in a way the industry consistently underestimates. OCM is neither a launch event nor a training session on go-live day. It is a sustained, intentional process of bringing people along, and it must be informed by the same research that informs design. Understanding what users are afraid of losing, what they don’t yet trust, and what they need to verify before they will change their behavior is UX research. As Sutton and Rao (2024) argue in The Friction Project, smart leaders make the right things easier and the wrong things harder, reducing the friction that blocks good outcomes. Treating adoption as a communications problem produces newsletters; treating it as a design problem produces adoption.

Leadership endorsement matters in a way that product teams sometimes dismiss. In high-stakes environments, people take cues from the people above them. The gap is real: In one survey, only 13% of healthcare executives reported having a clear strategy for integrating AI into clinical workflows. When visible, genuine organizational commitment to a new system exists—when it’s not just a mandate but a demonstrated belief—it changes the risk calculus for individual users. Commitment signals that the organization considered the change carefully, that it is not an experiment being run on their time, and that there is accountability above the employees if something goes wrong.

What Actually Works

After all of this—the fragmentation, the verification loops, the rational resistance—we need to be specific about what human-centered design should do when it is applied well to healthcare AI. Fundamental principles often get deprioritized in the rush to deploy capability.

Reduce cognitive load ruthlessly. Every context switch costs something. Every additional tab, every login prompt, every moment in which a user has to remember which system holds which information is a small withdrawal from a cognitive account that is already overdrawn, according to Asgari et al. The heart of what Sutton and Rao (2024) call friction is small, accumulated forces that make the right action harder than it should be; the work of design is to remove friction. The systems that get adopted in healthcare are almost always the ones that make a user’s day measurably simpler, not more powerful, but simpler. The question to ask is not “what can we add?” but “what can we eliminate?”

Make search and discovery frictionless. In a high-pressure environment, users do not have time to learn search syntax or navigate complex information hierarchies. The goal is not to teach clinicians to write better queries; the goal is to build systems that interpret imperfect queries. People interact with AI search much as they do with a web search bar, using short, approximate, and sometimes misspelled queries. Rather than pushing the burden of correction onto the user through prompt engineering, well-designed systems absorb it. Natural-language approaches that account for synonyms, misspellings, and alternate phrasings were shown in 2022 to help clinicians retrieve information faster and with lower cognitive burden than literal keyword matching. If finding information requires more than a few seconds of active effort, it will not be found, and if it is not found, it will not be used. If the path to an insight is unclear, the insight might as well not exist.

Surface relevant information proactively. The most valuable thing a well-designed AI system can do is anticipate what a user needs before they ask for it, not in a way that is intrusive or cluttered, but in a way that makes it feel like the system understands the context of the work being done. This design requires deep knowledge of actual workflows, which is why it cannot be designed by assumption.

Design for clarity, not capability. The instinct in product development is to show what the system can do. In healthcare, this instinct produces interfaces full of features that users do not have time to explore. A clean interface that surfaces what matters, while hiding what doesn’t, is not a lack of ambition. It is a sophisticated design choice that requires enough research to know what matters and what doesn’t.

Build trust through research, not assumptions. Every design decision that cannot be traced to something a real user said or did is a guess. In most industries, guesses are recoverable, whereas in healthcare, they compound. Real, sustained, contextual research is not a phase in a product timeline. It is the foundation that determines whether a product will be used or ignored and must be integrated into the inner workings of a project life cycle from start to finish.

UX Research Is Not a Phase

There is a version of UX research that gets practiced in healthcare technology: a round of user interviews before kickoff, a usability test before launch, and a satisfaction survey afterward. This type of research is better than nothing, but it is not enough. Documentation burden and usability problems in clinical systems are consistently traced to designs that failed to account for real-world workflows.

The gap between a designed workflow and an actual workflow is almost always bigger than teams expect, and in healthcare, it is rarely surfaced by interviews alone. People describe how they think they work, or how they are supposed to work, or how they worked before the last reorganization. Observation reveals something different: The workarounds, the informal knowledge-sharing, and the judgment calls that happen in seconds carry significant weight.

Research also validates AI performance in context, which differs from validating it in a lab. A system performing well on benchmark datasets may perform very differently with the messy, incomplete, and inconsistently formatted documents of real clinical environments; controlled studies of clinical search tools show that performance and usability must be measured against real tasks, not idealized inputs. Distinguishing the difference between digital modernization, which updates existing systems, and digital transformation, which fundamentally changes how work gets done, reduces the cost of getting it wrong, which compounds the later it is caught. The widely cited principle attributed to the IBM Systems Sciences Institute holds that a problem caught early can cost a fraction of what the same problem costs once a system is in production, on the order of 100 times less (NIST, 2002). Understanding the gap before deployment makes the difference between a successful launch and an expensive reversion.

Most importantly, user-informed design fosters ownership. When users see their feedback reflected in a product—when the thing that frustrated them in the first round of testing has been addressed, or when the feature they asked for is implemented—they become advocates instead of resistors. Their stake in the product’s success creates a psychological shift that reliably drives adoption.

The talented and motivated teams building usable AI for healthcare are often not the clinicians, coordinators, or the people who use these systems under pressure. UX research builds in the people doing the work at the center of every decision.

The Bridge That Doesn’t Build Itself

AI will continue to transform healthcare. The models will get better, the integrations will get deeper, and the capability gap between what is technically possible and what currently exists in clinical environments will continue to close. None of that is in question. What is in question is whether the adoption gap closes with the support of UX research.

The history of technology in healthcare is littered with systems that were genuinely capable yet unused, not because they were bad technology, but because the people they were built for were never truly centered in the process of building them, according to Menlo Ventures (2025). We can learn from the lessons of an assumed workflow, an underestimated cognitive load, and change management that wasn’t properly implemented. Users, when faced with a system that made their demanding job harder in exchange for an unseeable future benefit, made the only rational choice available to them.

AI does not change this dynamic around adoption but intensifies it: The more capable the system, the more disorienting a failure is.

Human-centered design provides more than a finishing touch on an AI product. Human-centered design forms the essential infrastructure that determines whether the product reaches the intended users. In healthcare, that infrastructure differentiates between technology that changes outcomes and technology that changes nothing except the budget.

The question every UX practitioner who is working in or adjacent to healthcare should ask is not “How might we build the right features?” but “How might we be in the room early enough to make sure the right features are the ones that get built?”

Resources

Asgari, Elham, Japsimar Kaur, Gani Nuredini, Jasmine Balloch, Andrew M. Taylor, Neil Sebire, Robert Robinson, Catherine Peters, Shankar Sridharan, and Dominic Pimenta. 2024. ”Impact of Electronic Health Record Use on Cognitive Load and Burnout Among Clinicians: Narrative Review.” JMIR Medical Informatics 12: e55499. https://medinform.jmir.org/2024/1/e55499

Yap, Greg, Derek Xiao, Johnny Hu, J.P. Sanday, and Croom Beatty. 2025. “2025: The State of AI in Healthcare.” Menlo Ventures. https://menlovc.com/perspective/2025-the-state-of-ai-in-healthcare/

Moy, Amanda J., Mollie Hobensack, Kyle Marshall, David K. Vawdrey, Eugene Y. Kim, Kenrick D. Cato, and Sarah C. Rossetti. 2023. “Understanding the Perceived Role of Electronic Health Records and Workflow Fragmentation on Clinician Documentation Burden in Emergency Departments.” Journal of the American Medical Informatics Association 30 (5): 797–808. https://academic.oup.com/jamia/article/30/5/797/7076268

National Institute of Standards and Technology. 2002. The Economic Impacts of Inadequate Infrastructure for Software Testing. (Planning Report 02-3). Washington, DC: U.S. Department of Commerce. https://www.nist.gov/document/report02-3pdf

Nguyen, Thuy D., Christopher M. Whaley, Kosali Simon, Neil Mehta, Hao Yu, Ryan K. McBain, Ateev Mehrotra, and Jonathan H. Cantor. 2025. “Adoption of Artificial Intelligence in the Health Care Sector.” JAMA Health Forum 6 (11): e255029. https://pmc.ncbi.nlm.nih.gov/articles/PMC12639477/

Park, Eunsoo H., Hannah I. Watson, Felicity V. Mehendale, and Alison Q. O’Neil. 2022. “Evaluating the Impact on Clinical Task Efficiency of a Natural Language Processing Algorithm for Searching Medical Documents: Prospective Crossover Study.” JMIR Medical Informatics 10 (10): e39616. https://medinform.jmir.org/2022/10/e39616

Sage Growth Partners. 2025. The Healthcare C-Suite’s Take on AI? https://sage-growth.com/market-report/healthcare-csuite-ai-trust/

Sutton, Robert I., and Huggy Rao. 2024. The Friction Project: How Smart Leaders Make the Right Things Easier and the Wrong Things Harder. New York: St. Martin’s Press. https://us.macmillan.com/books/9781250284426/thefrictionproject/