Every week brings another announcement: AI tools that can interpret scans, predict deterioration, automate clinical notes, identify patients at risk, support diagnosis or help people manage long-term conditions.

The technology is moving extraordinarily quickly.

But I increasingly think we are asking the wrong question.

The question is not simply: what can AI do?

It is: what do health systems need AI to do — and how do we bridge the enormous gap between technical possibility and adoption at scale?

I have spent much of my career working at the intersection of government, the NHS, science and industry. And one lesson has been remarkably consistent: brilliant technology is not enough.

When I joined Genomics, I encountered some of the most brilliant science I had ever worked with. The possibilities seemed almost endless.

And that was partly the problem.

The challenge wasn’t to explain everything the science could do. It was to turn the telescope around and ask: what problem does the NHS need us to solve?

That meant understanding where the evidence was strongest, where the technology could make the greatest difference and, critically, how it aligned with what the NHS and government were trying to achieve.

It eventually led us towards a much simpler proposition around population health and prevention and ultimately to the concept of a Genomic Population Health Service.

I think exactly the same discipline now needs to be applied to AI.

Start with the health system

Too often the conversation begins with the technology:

We have developed an extraordinary algorithm. Where can we use it?

The better question is:

What problem are we trying to solve?

Workforce capacity? Waiting lists? Earlier diagnosis? Avoidable hospital admissions? Long-term conditions? Mental health? Prevention?

Only then should we ask how AI might help.

It sounds like a subtle distinction. In practice, it changes everything: the evidence you generate, the people you need to convince, the economic case and, importantly, the story you tell.

From point solutions to system solutions

One risk with the current AI boom is that we create hundreds of highly effective individual products without solving the problem of how they work together.

A diagnostic algorithm may be excellent. A wearable may generate extraordinarily useful information. A genomic risk tool may identify disease risk years before symptoms appear.

But patients don’t experience healthcare as a collection of technologies. They experience a journey.

The opportunity is therefore not simply to produce more AI tools. It is to connect data, technology and clinical services around the needs of patients and populations.

And AI should ultimately be judged not by the cleverness of the algorithm, but by what happens next.

Does it improve outcomes? Release clinical capacity? Prevent an admission? Diagnose disease earlier? Give clinicians more time with patients?

From reactive medicine to predictive health

For me, this is where AI becomes really exciting.

Healthcare remains largely reactive. We become ill, develop symptoms and seek treatment.

But we now possess increasingly rich sources of data: electronic health records, imaging, genomics, laboratory results and wearables.

AI gives us a way of interpreting those data at pace and scale.

Combine AI with genomics, longitudinal health data and other biomarkers and we begin to move towards something much more powerful: identifying risk earlier, stratifying populations and intervening before disease develops or deteriorates.

That was what excited me so much about genomics. The important question wasn’t simply what someone’s genetic information could tell us. It was what could we do differently because we knew it?

AI potentially takes that idea much further and could be instrumental in delivering the NHS ambition to make the ‘left shift’ described in the NHS 10 Year Plan and move from sickness to prevention.

Adoption is the hard part

None of this happens simply because the technology exists.

Who pays? Who changes the pathway? Who integrates it? What evidence does the medical director need? What does the finance director need? And why should an already overstretched clinician change the way they work?

Then there is trust.

Patients need confidence about how their data are being used. Clinicians need confidence in the technology. We need clarity around accountability, bias, regulation and governance.

These aren’t peripheral issues to be addressed once the clever work is finished. They are fundamental to adoption.

I’ve come to believe that adoption is itself a discipline.

Successful health technology companies need to understand not just their product, but policy, clinical practice, economics, procurement and human behaviour. And they need to do much more of the heavy lifting required to make adoption easy for health systems.

There is no doubt that AI will become an increasingly important part of healthcare.

But “more AI” isn’t the objective.

Better health and a more sustainable health system are.

So perhaps the question we should be asking is:

Are we using AI simply to make today’s healthcare system more efficient or can we use it to create a fundamentally different model of healthcare: more predictive, more preventative and more personal?

For me, that’s where the real opportunity lies.