AI is transforming cancer research and therapeutics — but not in the way most headlines suggest. A view from the clinical infrastructure front line.

I’ve spent a lot of time in hospital corridors I wasn’t supposed to be in — not as a clinician, but as a technologist trying to understand what oncologists actually do with their days. What you notice quickly is how much of their time isn’t spent on medicine. It’s documentation, chasing results, reviewing scans that arrived three days late, writing referral letters that could have been drafted in seconds. The cognitive load is extraordinary. And cancer doesn’t wait for admin to clear.

This is where I think the AI conversation goes wrong. We talk about AI defeating cancer as though it will be some singular algorithmic moment — a model that reads a biopsy and delivers a verdict with superhuman precision. That may come. But the transformation happening right now is quieter and arguably more impactful: AI is absorbing the analytical and administrative burden that consumes some of the most trained minds in medicine and giving that time back to the patient in front of them.

In drug discovery, the numbers are sobering. Developing a cancer therapeutic typically takes over a decade and costs north of a billion euros, with failure rates that would disqualify almost any other industry. Foundation models trained on protein structures and genomic data are now guiding which molecules are worth synthesising at all — compressing years of hypothesis-testing into weeks. Companies are running what are called self-driving labs, where AI proposes the experiment, robotics run it, and results feed back into the model within hours.

In pathology, AI analysis across whole-slide images is surfacing patterns that human review would almost certainly miss — not because pathologists aren’t skilled, but because the combinatorial space is simply too large. I think about a patient in a regional hospital where the nearest specialist is two hundred kilometres away. AI-assisted analysis gives her a faster, better-informed first read while the referral is in process. That latency reduction, measured in days or weeks, can be the difference between stage two and stage three.

“The breakthrough won’t be a single model that defeats cancer. It will be a thousand quiet improvements that compound — in the lab, in the clinic, in the ward round.”

And then there’s the therapeutic frontier. Personalised mRNA cancer vaccines — where a patient’s tumour is sequenced, the neoantigens identified, and a bespoke construct designed for that individual — are in clinical trials right now. The sequencing, the prioritisation, the manufacturing logistics: all AI-assisted. It’s a workflow that would have been operationally impossible five years ago. Biology foundation models are doing for proteins what large language models did for text — building generalisable reasoning about molecular structure that can be applied across disease types.

What I find most interesting, though, is the layer beneath all of this — the infrastructure. When I sit with hospital technology leaders across EMEA, they’re not asking me about frontier research. They’re asking whether these workloads can run without patient data leaving the building. They’re asking about governance, integration with legacy systems installed before some of their junior doctors were born, and accountability when a model gets something wrong. These are exactly the right questions. Clinical transformation doesn’t happen because a model performed brilliantly in a paper. It happens when the infrastructure is trustworthy enough, the governance clear enough, that a clinician can rely on it during a busy Tuesday afternoon clinic.

That’s unglamorous work. But it’s the work that determines whether any of this actually reaches patients. The machines aren’t here to replace oncologists. They’re here to give them their time back — and what a good doctor does with that time remains entirely, irreducibly human.