Healthcare artificial intelligence (AI) has moved well beyond the experimental. It is already accelerating research, improving patient journeys and supporting diagnostics. The next wave, driven by generative AI, promises more still: intelligent triage, automated clinical administration, and earlier intervention for patients whose symptoms suggest urgent need.
The scale of the opportunity is huge. EU-commissioned research from earlier this year suggests AI could free up the equivalent of €252 billion currently lost to inefficiency in clinical decision-making alone. Yet the European Commission’s own assessment, published in August 2025, found that while the AI pipeline is strong, real-world deployment is lagging – held back by fragmented, non-standardised data, and outdated digital infrastructure.
The problem, in other words, is not the AI model. It is the data feeding it and – critically – whether that data is current, connected and governed well enough to be trusted.
That distinction matters enormously. An out-of-context AI recommendation in a retail setting might mean a customer receives the wrong offer. In healthcare, it could mean a patient receives the wrong care at the wrong time.
The biggest barrier is not AI capability. It is fragmented data.
Many healthcare systems still operate on fragmented technology foundations. A patient may interact with a hospital, insurer, pharmacy, and remote monitoring provider, each generating useful information. But if that data is locked in separate systems, the promise of AI becomes much harder to realise.
From my conversations with customers, it’s clear that this challenge is not unique to any one organisation or country. Across Europe and beyond, health systems are grappling with the difficulty of connecting systems that were never designed to talk to each other – a problem that is becoming even more acute as the European Health Data Space pushes towards cross-border data sharing.
NHS England’s Federated Data Platform offers one model of how to approach this, providing connective infrastructure across hospital trusts for use cases such as waiting list management and discharge planning.
But the lesson extends beyond any single national initiative. Infrastructure that connects data is necessary but not sufficient. The data flowing through it must also be real-time. An AI model cannot safely support triage, prioritisation, or personalised care if the data behind it is incomplete, late, inaccessible, or poorly governed.
Trust is the fundamental adoption challenge
Healthcare has a unique relationship with trust. People may be uncomfortable if a coffee shop app knows where they are and sends them an offer. Yet they may welcome a health alert from a wearable device if it warns them about an irregular heart rhythm. The difference is perceived value, transparency, and control.
Insurance offers a useful example. Vitality’s Active Rewards model – built on Confluent’s real-time data streaming platform – links activity tracking and personalised incentives to individual health goals, with wearable technology recording activity and heart-rate workouts as they happen.
Done well, these models can benefit both the individual and the insurer: people are encouraged to become healthier, while insurers reduce risk. The real-time element is not incidental to the model, it is what makes it credible and commercially viable. A batch-updated version of the same concept would not deliver the responsiveness that members expect or that accurate risk modelling requires.
However, the Vitality example also shows why trust is so important. The same data that can help someone improve their health could create serious discomfort if it is shared too widely or applied in ways the individual does not expect.
This is especially true as generative AI becomes more prevalent. If an AI system prioritises a referral or drafts clinical advice, healthcare leaders need to know what data informed that output. They need lineage and auditability. They also need controls around access, encryption, retention, and consent. Most importantly, they need clear human oversight where decisions could affect patient care.
What does good look like?
A strong AI-driven healthcare system is not one where every process is automated. It is one where the right data reaches the right person, system, or model at the right time – with the right permissions attached, and a clear record of how it got there.
That means healthcare leaders should focus on five foundations.
First, they need interoperability. Data must be able to move securely between systems, providers, and regions where possible without forcing clinicians or patients to bridge the gaps manually.
Second, they need real-time context. Remote monitoring, hospital capacity, test results, patient deterioration, and discharge planning all become more valuable – and reliable – when information is current.
Third, they need governance by design. Privacy, consent, access control, lineage, and auditability cannot be added after the fact. They must be built into the data architecture from the start as engineering requirements.
Fourth, they need explainability and accountability. If AI is used to support decisions, organisations must be able to show what data was used, what rules applied, and where human oversight sits.
Finally, they need inclusion. Digital channels can improve access for many people, but they can also create barriers for those who are less digitally confident, older, vulnerable, or unable to navigate online systems. A healthcare service that is efficient for some but inaccessible to others has not truly transformed.
Real-time infrastructure is the strategy
What the European Commission’s own research highlights is the urgent need for health systems to have the data infrastructure to use AI responsibly, at scale, and in a way that patients and clinicians can actually trust.
That means connecting fragmented systems and moving data in real time, not in yesterday’s batches. Governing it, proving its lineage, protecting sensitive information while still making it useful – none of this is straightforward. But in my view, this is where the real work is.
AI may prove to be one of the most consequential technologies healthcare has ever adopted. But the models alone won’t get it there. The data feeding them needs to be current, governed, and trustworthy. Because in healthcare, when the data fails, it is not a system that suffers, it is a person.
By Peter Pugh-Jones, Chief Field Data Officer, Confluent