Vice President, Services Operations UK, Presidio

In 2022, I wrote a paper exploring the opportunities, challenges and implications of introducing AI into a hospital environment that was, hypothetically, at the very beginning of its AI journey. Using Gartner’s AI maturity model, I placed most healthcare organisations at Level 1 – Awareness: they knew AI existed and were already using elements of it (such as facial recognition on security cameras, next-generation antivirus and other embedded technologies), but lacked a coherent strategy to bring those capabilities together. In many cases, organisations didn’t even realise AI was already part of their technology estate.

Four years later, I couldn’t have predicted just how quickly AI would evolve.
Generative AI barely featured in most conversations in 2022. Today, it’s arguably the biggest driver of AI adoption across healthcare. While the technology has advanced at extraordinary speed, what has surprised me most is how well many of the strategic principles from that original paper have stood the test of time.

If I were writing it today, here’s what I would say differently.

The maturity picture has moved, but unevenly


Healthcare is no longer quietly waiting at the back of the AI adoption queue. Most hospitals now use AI in at least one clinical or operational function. NHS trusts are actively exploring AI-supported triage, medical imaging, virtual assistants and administrative automation. Momentum has unquestionably shifted.
But using AI somewhere and having an AI strategy remain two very different things.
Much of the industry commentary surrounding 2026 has focused on governance, and for good reason. Leadership teams are increasingly playing catch-up with clinicians who have already embraced generative AI tools, often without formal oversight. It’s a familiar pattern. Technology is once again moving faster than organisational strategy, with “shadow AI” rapidly replacing the “shadow IT” challenge of a decade ago.


The lesson from 2022 still holds.


Small, well-governed AI initiatives consistently outperform ambitious programmes launched without the right foundations. Organisations still early in their maturity journey will achieve more by proving value through focused, measurable use cases than by pursuing transformation before governance, skills and processes are ready.

The biggest change wasn’t diagnosis, it was documentation


Four years ago, the most exciting AI conversations centred on diagnosis: medical imaging, symptom checkers and early cancer detection. Those use cases continue to evolve, but they haven’t become the defining story of AI adoption. Instead, the breakout success has been ambient clinical documentation.


AI scribes that listen to clinician-patient conversations and automatically generate structured clinical notes have moved from innovation to mainstream deployment at remarkable speed. Large health systems now report that most physicians use ambient documentation daily. Across the United States, roughly two-thirds of hospitals using major electronic health record platforms had deployed ambient documentation tools by 2025. In the UK, around 40% of GP practices and an increasing number of secondary care organisations are already using similar technologies, with adoption expected to accelerate further as NHS England expands support.


The evidence is more nuanced than some headlines suggest. Some large studies demonstrate significant documentation time savings, while others show more modest improvements. But the overall direction is unmistakable. Ambient documentation is rapidly becoming an expected part of clinical practice, particularly in primary care, mental health and emergency medicine.
Looking back, this has changed my thinking about one of the questions I raised in 2022: would patients trust AI in healthcare?


The answer has turned out to be simpler than I expected.


The AI patients trust most is often the AI they barely notice.


When technology quietly removes administrative burden and allows clinicians to spend more time making eye contact, listening and engaging with patients, trust increases naturally. That’s a far easier proposition than asking patients to replace human interaction with chatbots or automated decision-making. Ultimately, patients want technology that enables better care – not technology that becomes the care.


The next phase is already emerging.


Industry discussions throughout 2026 increasingly focus on agentic AI: systems capable of summarising information and acting on it by coordinating referrals, following up results and managing elements of clinical workflows. That represents a far greater leap in both capability and risk than anything I discussed four years ago, making governance more important than ever.


Business services: the principles remain, the tools have matured


In my original paper, I described HR, Finance and IT as the operational foundation for enterprise AI adoption.
That framework still holds.
The difference is that today’s tools are considerably more capable.

HR continues to benefit from reducing repetitive recruitment and onboarding tasks, allowing teams to focus on people rather than paperwork – arguably even more important given the continuing workforce pressures across healthcare.

Finance has expanded beyond robotic process automation into intelligent reconciliation, anomaly detection and forecasting.
Meanwhile, IT service desks have evolved from rules-based chatbots into conversational assistants capable of understanding context, resolving more complex requests and reducing support backlogs without sacrificing the empathy healthcare professionals need when systems fail.


Perhaps the biggest lesson remains unchanged.


Technology alone rarely delivers transformation.


The organisations seeing the greatest value from AI are investing just as heavily in change management, governance and user adoption as they are in software licences. People, process and technology remain in that order for a reason.


Clinical AI: better technology doesn’t remove ethical responsibility


Back in 2022, I referenced IBM Watson for Oncology as the defining example of clinical AI gone wrong.


It’s still a valuable lesson, but no longer the most relevant one.


A well-documented example is the Epic Sepsis Model, a sepsis prediction tool used in many US hospitals. In 2021, researchers at Michigan Medicine published an external validation in JAMA Internal Medicine. It reported that, at the threshold they tested, the model identified about a third of patients who developed sepsis. It also generated a high volume of alerts, with roughly 109 patients needing evaluation for each true case found.


Epic has noted that results depend on how each organisation configures and uses the tool, and it has since released an updated version with greater emphasis on local validation. That is a sensible response, and it points to the real lesson.


The issue was not that AI doesn’t work. It was that models behave differently outside the settings in which they were built, and that performance can shift as patients, clinical practice and data change over time (data drift). Alongside that sits automation bias, where clinicians can place too much confidence in an algorithm’s output simply because it exists. Both are shared challenges for vendors and healthcare providers alike, and both are best managed through local validation, ongoing monitoring and clear clinical governance.
Liability also continues to shape adoption.


Recent UK research found that clinicians who avoid AI most commonly cite professional and medico-legal concerns. Interestingly, clinicians already using AI report many of the same anxieties.


That aligns closely with one of the points I made in 2022. GDPR’s protections against decisions based solely on automated processing continue to place meaningful limits on how much clinical autonomy AI systems should be given.


Radiology remains one of the clearest examples of why representative training data matters. Historical biases embedded within imaging datasets don’t simply disappear because models become more sophisticated. If anything, regulators are now treating bias mitigation as a legal requirement rather than a professional aspiration.


Governance has moved from best practice to legal obligation


If there is one structural shift that defines the past four years, this is it.


When I wrote my original paper, I argued that healthcare organisations should extend existing governance structures, such as Architecture Review Boards and Institutional Review Boards, to oversee AI.


Today, that is no longer simply good practice.


It is rapidly becoming a regulatory expectation.


The EU AI Act classifies many healthcare AI applications – including diagnosis, clinical decision support, patient monitoring and treatment recommendations – as high-risk systems. Organisations deploying them must demonstrate technical documentation, human oversight, conformity assessments and ongoing monitoring. While implementation timelines continue to evolve, the direction is unmistakable: healthcare AI is increasingly being regulated like a medical device rather than conventional software.
The United States presents a more fragmented picture, with individual states introducing healthcare-specific AI legislation while federal agencies strengthen guidance around algorithmic fairness and discrimination.


Regardless of geography, the practical implication is the same.
Healthcare organisations need formal governance structures capable of evaluating AI use cases, assessing clinical risk, understanding data provenance and ensuring appropriate human oversight. Waiting for regulatory certainty is no longer a viable strategy.


Governance needs AI literacy
Governance alone is not enough.


As AI becomes embedded in everyday clinical workflows, organisations must also invest in AI literacy across both clinical and operational teams.
Healthcare professionals don’t need to understand how large language models are engineered, but they do need to understand where they are likely to fail, how to identify hallucinations, recognise bias and know when clinical judgement should override algorithmic recommendations.


The most mature organisations increasingly view AI education as a core professional capability rather than optional digital training.
Technology changes quickly. Human judgement remains irreplaceable.


Patient experience remains the north star


The central principle underpinning my original paper has not changed.
Healthcare AI should ultimately improve patient experience and patient outcomes.
What’s different today is the breadth of tools available to achieve that goal. Success is no longer measured by how intelligent an AI assistant appears, but by how effectively it removes friction, reduces administrative burden and allows clinicians to spend more meaningful time with patients.

Choice is equally important.


Patients should be able to opt into AI-supported interactions where appropriate, not feel compelled to use them when they would rather speak with another person.
Technology should create more human healthcare, not less.

Looking back


Revisiting my 2022 paper has reinforced something I didn’t fully appreciate at the time.


The technology has changed dramatically.


The strategic fundamentals have not.


Start small. Govern early. Keep people, process and technology in that order. Never lose sight of the patient experience.

What I underestimated was how quickly one category of technology, ambient documentation, would become standard clinical practice, and how rapidly governance would shift from a recommendation to a legal expectation.

The organisations that will lead the next phase of healthcare AI won’t necessarily be those adopting the newest technologies first.

They’ll be the organisations that build trust, governance and capability just as quickly as they build AI itself. Learn how Presidio’s AI blueprint can help you create value in your AI investments.

References

  1. Panetta, K. et al. (2019), ‘The CIO’s Guide to Artificial Intelligence’, Gartner.
  2. NVIDIA (2025/2026), ‘State of AI in Healthcare and Life Sciences’ – trend analysis.
  3. Wolters Kluwer (2025), ‘2026 Healthcare AI Trends: Insights from Experts’.
  4. Skills for Health (2026), ‘AI in Healthcare: What Clinicians and NHS Leaders Need to Know in 2026’.
  5. Healthcare Dive (2026), ‘Top Healthcare AI Trends in 2026’.
  6. AJMC (2026), ‘Ambient AI Tool Adoption in US Hospitals and Associated Factors’.
  7. Medical Economics (2026), ‘Take Note: The AI Scribe Era Is Here’.
  8. STAT News (2026), ‘Large AI Scribe Study Finds Modest Time Savings, Inconsistent Use’.
  9. Annals of Emergency Medicine (2026), ‘Ambient AI Scribe Adoption and Documentation Time in the Emergency Department’.
  10. Tandem Health (2026), ‘EU AI Act Explained: What Healthcare Organisations Need to Know’.
  11. MedDeviceGuide (2026), ‘EU AI Act for Medical Devices: Compliance Guide’.
  12. Wong, A. et al. (2021), ‘External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients’, JAMA Internal Medicine, 181(8), 1065–1070. https://doi.org/10.1001/jamainternmed.2021.2626.
  13. Censinet (2026), ‘AI Liability Risks in Healthcare: What to Know’.
  14. Information Commissioner’s Office, ‘Guide to the GDPR – Article 22(1), Rights Related to Automated Decision Making’.
  15. Care Quality Commission, ‘What Can You Expect From a Good Hospital?’.
  16. Geis, J.R. et al. (2019), ‘Ethics of Artificial Intelligence in Radiology’, Canadian Association of Radiologists Journal.

Related Posts

View All
November 20, 2026

An Appraisal of the Opportunities, Challenges, and Implications of Implementing Artificial Intelligence in Healthcare

Learn more
November 16, 2026

What 2026 Told Me About AI in Broadcast Workflows and Where Media is Headed

Learn more
November 13, 2026

Don’t Fall Into the AI Governance Trap

Learn more