AI & the future of healthcare

Image recognition has been used in healthcare for years. A model learns from thousands of scans, skin photos or retinal images what an abnormality looks like and points out suspicious areas. It looks at the hundredth image as sharply as the first, which matters in a screening programme where nearly every image is normal. The FDA has cleared software that assesses retinal photos on its own for diabetic eye damage that needs referral to an eye specialist. In a large breast cancer screening study, AI decided which mammograms went to one radiologist and which to two: more tumours were found, with almost half as many readings. Whether every extra finding needs treatment is still for the doctor to weigh.

An AI scribe listens in during a consultation and turns the conversation into draft notes: the symptoms, the examination, what was agreed, and if needed a referral letter and a summary for the patient. That saves typing, and the doctor can look at the patient again instead of the screen. In a trial in an emergency department, it saved doctors more than three quarters of an hour per shift. The notes remain a draft, which the doctor reads before they go into the record. Because such notes read smoothly, the doctor checks what really counts: is the allergy the patient mentioned in there, and has the dose been copied correctly?

Triage means deciding who needs care first and where that patient should go. AI adds a prediction: how likely is this patient to be admitted, and how busy will the emergency department be tonight? With that answer, a hospital can arrange a bed sooner and spread its capacity better, so patients wait less. In one emergency department, a model recalculates that admission probability every five minutes; above fifty per cent, the ward can start preparing a bed. That has already saved time. How much depends on clear agreements about when doctors request an admission on the strength of the prediction.

Early warning means spotting signs that a patient is getting worse, so clinicians can judge sooner whether to step in. Nurses have done this for years with scoring charts for heart rate, breathing and blood pressure. AI goes further: it follows those values continuously, together with lab results, and looks for a pattern that none of the separate signals reveals. With sepsis, the body's runaway response to an infection, every hour counts. In five hospitals, the death rate among sepsis patients was almost a fifth lower when doctors confirmed such a system's alert within three hours; that is an association, not yet proof. Performance varies widely between systems, which is why a hospital first tests a model on its own patients.

An AI agent is AI that takes the steps needed to finish a task on its own. Where a chatbot gives an answer, an agent puts the appointment in the diary, sends the message or calls the patient back, working inside the practice's own systems. That suits agents to the predictable work now clogging the phone lines, such as the eight o'clock rush at the GP surgery: asking about symptoms, booking an appointment, calling a patient after a procedure. In hospitals, AI already drafts replies to patients' questions, which a clinician checks before they are sent. Rules set out what the agent may do at each step and what makes a person take over.

A language model is AI trained on vast amounts of text. It answers questions in plain language, summarises records and reasons about what it reads. Models that have also absorbed medical literature help a doctor with a difficult diagnosis, like a reference book that talks back. In a test with complex cases from a leading medical journal, an AI system that mimicked a panel of doctors reached the right diagnosis more often than experienced doctors working without colleagues or reference material. Patients use language models too. Some have already shown their results to a chatbot and arrive with a list of specific questions. The conversation starts at a different level, and the doctor checks what the model claimed.

Precision medicine fits a treatment to the characteristics of one patient rather than the average of a group. AI finds patterns in tissue images, lab results and clinical data that predict how a disease will progress and who will benefit from which treatment. Two men with the same prostate cancer, the same stage, the same age: one gains from hormone therapy alongside radiotherapy, the other mostly gets the side effects. Treatment guidelines now include an AI test that uses biopsy images and clinical data to estimate which group a patient belongs to, developed with data from thousands of patients in randomised trials. What the doctor gets back is a probability, and with it a better conversation about how much benefit is worth the burden.

In mental health care, AI can help in two ways. It supports therapists with initial assessments, notes and tracking symptoms, leaving more time for the conversation itself. And it talks to clients: a therapy chatbot repeats exercises from the treatment and is available when the therapist is not, in the evening or between two sessions. In a randomised study, depression symptoms among participants who used such a chatbot fell by about half on average. Researchers monitored the conversations and could step in when needed. As an extension of treatment, a chatbot like this needs clear agreements: who follows the conversations, and when does a therapist take over?

A medicine usually works because a molecule binds to a protein in the body and switches something on or off. To design that molecule, you need to know the protein's shape, and working out that three-dimensional structure could once take years of lab work. AI now predicts it in minutes; the people behind the method received the 2024 Nobel Prize in Chemistry. That speeds up the start of drug development: finding the target and designing a molecule that fits it. A candidate drug for lung fibrosis, whose target and molecule were both found with AI, reached phase two of clinical trials. Whether such a drug works and is safe still has to be shown in people.

In an AI-first healthcare organisation, AI and new technology take over the repetitive and predictive work: the work clinicians need not do themselves. Technology can handle recording, scheduling, triage questions and follow-up calls, and forecast pressure and admissions, so professionals keep time for work where human contact and professional judgement count: having a difficult conversation, weighing a borderline case, noticing that someone is more frightened than they say. Many organisations still use AI on a small scale, as an aid to the old way of working. An AI-first organisation looks department by department at which tasks technology can take over and how much time that frees up for personal contact.

When AI takes over part of the work, the work that remains changes, and work appears that did not exist before. Broadly, the shift touches three kinds of role. Clinicians who check what AI produces: they search less, verify more and decide when to depart from its advice. People who keep watch over the technology: they check whether a model still performs after an update, investigate errors and decide when a system has to be switched off for a while. And specialists in fields that grow with the technology, such as genetics, robotic surgery and remote care. Many of these roles are not yet reflected in job descriptions, training programmes or time set aside in the rota.

On a hospital ward, people and AI increasingly work as one team, with a clear division of tasks. AI turns the conversation at the bedside into notes, helps plan beds and shifts and analyses readings and records, so a deteriorating patient stands out sooner. A next step is an AI agent that prepares the treatment: pulling together results, medication and earlier agreements and scheduling that day's tests. The nurse checks, steps in and sees what no record shows: a patient who is quieter than yesterday, a family that can no longer cope. The team agrees who reviews the AI's work and who answers for it when something goes wrong.

Leaders in healthcare take the initiative on AI and new technology themselves. By setting the example, they get clinicians building the future with them. It starts with direction: a picture of the care their organisation will provide ten years from now, concrete enough that a ward manager knows what it means for the team. Then initiative: starting themselves, small and visible, even before the technology is mature or every possibility is clear. And it means dealing with uncertainty. Nobody knows exactly what healthcare will look like in ten years. A good leader says so out loud, still sets a course and adjusts it when the facts change.

Buying software is quick; changing how a ward works is not. Change in healthcare is rarely about technology. It is about whether clinicians and leaders are willing to explore new possibilities and then genuinely work differently: dropping a habit, handing over a task, reporting it when the tool gets something wrong. In a survey of hospitals, they name the sticking points themselves: too little knowledge among staff and difficulty with change, alongside money. That willingness grows where people get room to try things, in working time and with real cases, and where leaders are the first to work differently. Where it is missing, even the best AI stays a pilot.

The employee of the future needs knowledge that training courses do not always provide in time: how to work with AI, when to trust a system and when not to, how to keep your own judgement sharp. So organisations are taking charge themselves, in their own future academy. Existing staff build their skills there, and new colleagues are prepared for working with AI and new technology. Early adopters bring their colleagues along.

An ecosystem is the network of parties an organisation builds solutions with: partners, start-ups, scientists and governments. An ageing population, staff shortages and AI are too big for one organisation, region or country. Partners elsewhere sometimes already have experience with a problem your organisation is still trying to solve. Working with such partners lets you learn from their experience, can help you avoid mistakes others have already made and brings an innovation into practice faster. An organisation's success depends more and more on the quality of its ecosystem.