Roy Wills, global head of healthcare business and partnerships at Intellias, argues that healthcare’s AI problem is not innovation, it’s implementation.
Artificial intelligence (AI) is not in short supply in healthcare. Health systems are inundated with technology promising to revolutionise patient care, employee productivity and administration. From ambient scribes to sepsis prediction models, the pipeline of innovation is rich and expanding at pace.
And yet, most healthcare providers are not reaping the rewards.
Despite significant investment and enthusiasm, most AI tools never make it beyond pilot programmes. They’re tested, validated in controlled settings and then quietly put aside. And that’s not because they don’t work, it’s because they don’t work in practice.
AI pilot paralysis
The AI pilot conundrum across health systems is no secret. The problem isn’t a lack of appetite for new technology, especially for tools that remove friction from the layers of work that keep clinicians from actually practising medicine.
In fact, hospitals and healthcare organisations are often running dozens of small-scale trials simultaneously. And engagement among employees is also high: 80% of NHS employees surveyed supported the use of AI for patient care, according to a survey by The Health Foundation.
The problem is that very few of these scattered initiatives translate into organisation-wide deployment. What many healthcare leaders have learned the hard way is that success in a pilot does not equal success in production.
In controlled environments, AI tools are tested under ideal conditions: complete data, controlled workflows, and dedicated support. It’s no surprise they show strong results. But real, clinical environments are far less predictable. Patient data can be incomplete or delayed, clinicians are juggling competing priorities, time pressures are constant, and there’s little room to learn new systems or adjust workflows.
Why clinicians abandon new technologies
Healthcare professionals are already operating at or beyond capacity. Burnout is high, and the administrative burden continues to grow, weighing heavily on already stretched primary care teams.
In this environment, even the most promising AI tools will fail if they add friction to already complex workflows. Too often, clinicians are expected to access separate platforms, navigate additional processes, interpret outputs outside their core clinical systems, or manually reconcile insights with electronic patient records (EPRs).
Each additional step is a tax on time and attention, and it’s hardly surprising that many clinicians disengage.
It becomes almost irrelevant if a tool is clinically sound. If it’s perceived as disruptive rather than supportive, and if a solution does not seamlessly integrate into the way care is delivered, it’s unlikely to survive in high-pressure environments.
Healthcare’s unique complexity problem
Hospitals operate within the most complex environments imaginable: fragmented I.T. networks with legacy infrastructure, strict regulatory and compliance requirements, financial constraints and razor-thin margins, and high-stakes decision-making where errors have life-changing consequences.
Compounding this is a less visible, but equally critical issue: data. Clinical data is often fragmented, inconsistent and poorly aligned, spread across EPRs that struggle to communicate, alongside legacy systems and documentation practices shaped as much by billing as by delivery of care. While interoperability is being addressed, high-quality data integration remains some way off.
These deeply entrenched systems can make introducing AI technology incredibly challenging. This goes beyond a technical exercise, requiring reimagined workflows, not simply layering new tools onto existing tasks, and a willingness to reflect on whether the way things have always been done is actually the best way.
The most effective solutions – the ones where organisations can measure a clear ROI – will feel less like a tool, and more like an extension of the clinical environment.
We’re already seeing successful examples of this shift: Ambient documentation tools that integrate into clinical workflows without requiring additional input, decision support systems that present recommendations within existing order sets, advanced analytics solutions for cancer treatment therapy.
These solutions succeed not because they are more advanced, but because they are more aligned with how healthcare actually works.
High-impact AI
Successful health systems using AI focus on scaling a small number of high-impact AI use cases rather than spreading efforts across pilots. They rigorously tie every initiative to measurable financial and clinical outcomes, while embedding AI directly into core systems like EPRs to ensure seamless adoption.
Just as importantly, they establish clear governance and ownership, and align incentives across clinical, operational, and financial stakeholders to drive real, sustained impact.
If the industry can master implementation, AI adoption will look very different to how it does in today’s fragmented landscape.
We’ll see fewer pilots, but more scalable deployments, AI embedded within core systems, not just layered on top, clinician experiences that are simplified, not expanded, measurable outcomes tied to operational and financial performance.
Most importantly, AI will begin to fade into the background. Just as clinicians no longer think about the underlying infrastructure of EPRs during patient care, the same will become true for AI. Its presence will be defined by outcomes, and not interfaces.
AI technology must be designed with real-world constraints in mind, prioritising integration over novelty. Health systems must become more disciplined in how they evaluate and scale technologies. And we all must recognise that success is not measured by what AI can do, but by what it actually changes in practice.
Until then, the industry risks continuing to mistake activity for progress; running more pilots, testing more tools, and solving the wrong problem. Healthcare doesn’t have an AI innovation gap, it has an execution, economics, and alignment gap.
