Artificial Intelligence in Medicine: Where Are We Heading?

Without a doubt, everything looks promising: a surge in investments in digitalization and Artificial Intelligence (AI), thousands of MedTech and HealthTech startups, millions of users of health & wellness apps, dedicated congress panels and events focused on healthcare innovation, and almost daily news of new breakthroughs.

However, the vast array of existing and emerging AI products in medicine is not yet a market. It is a layered and diverse landscape – with varying oxygen levels, depths, and completely different ecosystems: some well-developed, others scarcely inhabited. Viable models are already emerging in some areas, while elsewhere development is underway in completely unsuitable spots. Some solutions are fully operational, others are at the starting line, and some are stuck in demo mode or unlikely to succeed in the short or medium term.

The first true breakthrough has likely occurred in medical imaging – radiology, dermatology, ophthalmology, pathology. These are no longer prototypes but FDA-approved algorithms, CPT codes for insurance reimbursement, integrations with medical information systems, and cases of commercial deployment. There have been major exits and IPOs. AI-assisted imaging took the lead due to a simple fact: medical images are the most structured and standardized segment of the vast world of medical data and thus highly trainable for machines. Descriptions of medical images are not diagnoses by definition, meaning this segment more easily clears regulatory and ethical implementation hurdles.

The next layer includes clinical decision support systems. It’s important to distinguish between two subtypes: first, systems based on a single type of medical data; second, systems that consolidate various data types. Mono-systems might operate only on symptoms, lab values, or auscultation data. These are already in use and bring clinical benefits to early adopters. Their limitation is the lack of full context – so far, these remain assistant tools for human doctors. There is no stable reimbursement model yet, and no clear answer to who pays: the patient, the doctor, or the hospital (spoiler: no one wants to pay). Hence, monetization remains uncertain.

The second type – multimodal systems – combine various sources: symptoms, medical images, lab results, patient history, and medical records. Many see these systems as having the greatest potential, as they mimic a physician’s real-world clinical reasoning. They are seen as a possible path toward fully “replacing the doctor”– at least in scenarios that don’t involve physical diagnostics or procedures (e.g., surgery). But these systems are much more complex technologically. One of the main barriers is the deep fragmentation of healthcare systems – even within a single country, let alone across borders. Medical data is stored in incompatible systems by various providers, often still manually. Gathering a quality dataset is a major challenge in itself.

Implementation is even harder. Responsibility and decision-making centers are spread across various institutions, and change is slow. Budgets are fragmented across providers, insurers, and governments–no one can agree, but all expect quick returns. These innovative solutions face the typical “value-capture mismatch”: the benefit is recognized and obvious, yet no one wants (or is able) to pay for it.

A notable example of an AI-assisted clinical decision support system is the “digital family doctor”– a chatbot-like tool to replace primary care. It’s a highly attractive target for developers and simultaneously one of the most controversial areas.

In demos, the chatbot-doctor looks great. Many prototypes offer impressive UX, rapid and accurate responses, and vast knowledge bases. They shine at trade shows and investment pitches. But in real life, adoption is nearly non-existent. There’s no stable track toward broad implementation yet.

Let me explain: the issue is not the technology. It’s that it tries to replace not just anyone–but a human doctor and the patient’s personal choice. Today, a “doctor” isn’t necessarily someone in a nearby clinic. It could be a trusted contact via messaging, an online consultation from a reputable clinic, or a familiar specialist contacted directly from anywhere in the world.

And here’s the insight: not just the wealthy, but also much of the middle class–those with internet access and some ability to pay for basic care – will often prefer a human over a bot. It doesn’t have to be face-to-face – often it’s online. The chatbot might supplement the human doctor, possibly using open-source language models. At worst, patients will pay out-of-pocket for offline consultations.

Massive adoption of such chatbot systems is technologically ready – but anthropologically, it is not. That is the real barrier.

Also, discussions about large-scale digitalization often overlook (or deliberately ignore) the fact that vast parts of the world still lack digital access. Millions walk barefoot on dirt roads. Many still lack consistent internet access. Millions are illiterate. Еven within Ukraine, life in cities and depressed rural areas reflects two entirely different anthropological realities. One lives with mobile apps and online shopping, the other with button phones and habits unchanged for 50 years.

One more problem: a medical chatbot (like a family doctor) is useful only if there is a clear next step – specialist diagnostics, treatment, intensive care, or surgery. Even prevention today involves physical procedures – more specialized and increasingly costly. A symptom tracker only makes sense if early detection (even pre-symptomatic) leads to timely, affordable treatment before the illness progresses. And I emphasize: this often requires not just “lifestyle changes” but specific and costly treatment. If the system has no such path, or wait times stretch into years, the digital doctor will exist in a vacuum.

Moreover, healthcare architectures vary so widely that building a “universal” chatbot is fantasy. Identical symptoms require different next steps in different systems. And finally: trying to digitize something that doesn’t exist – namely, adequate available services–is more parody than innovation. That’s the real challenge.

Yet this transformation doesn’t have to be evolutionary.

Pre-2020, people used Skype and held conference calls – but worked and met mostly in person. “A New Year corporate Zoom party? Don’t make me laugh,” said anyone in the private sector. Then lockdowns caused a social shock and forced mass adaptation. In weeks, behaviours once considered immutable changed permanently. Post-pandemic, many social groups became more mobile – remote work skills are here to stay.

Something similar could happen with digital family doctors. The trigger won’t be better code, safer protocols, a new startup, or “institutional readiness.” All that exists already. The last remaining factor is the human – the end-user. Most likely, the quantum leap will follow a major shock–not necessarily epidemiological. It could be political or economic.

The ground is ready. Primary care is cracking worldwide: aging doctors, burnout, workforce shortages, and an exodus from rural practice. Patient numbers are rising – while physicians dwindle. Many countries are already close to failing to ensure basic access to care – and it’s only getting worse. Add another major war or a recession that wipes out public budgets, and you have a bifurcation point. This won’t be a “digital upgrade”– it’ll be a rupture. That’s when the digital doctor will take its place, irreversibly – at least in countries with widespread internet. For those unable to make or execute their own healthcare decisions, it will be the only replacement for today’s primary care. What happens after this shift is a separate story.

Let’s move on to wellness and self-monitoring. Apple Watch, Oura, Levels, CGM – this is no longer futurism, but routine for millions of users. Clarification: only for those not walking barefoot on dirt roads, but driving personal cars. Still, the spectrum is wide. There are simple apps for monitoring blood pressure, pain, sleep, or pulse. And then there are “medical Lamborghinis”– expensive tools for genetic risk assessment, hi-tech biosensor data interpretation, and premium “track-everything” solutions.

This is a separate world in the healthcare innovation ecosystem, and it is rapidly expanding. At the top are products aimed at niche audiences who pay for status, convenience, aesthetics, and the right to be at the forefront. These are where luxury AI-based products thrive – like 7-star hotels: not for the masses, but commercially viable and profitable in their monetization model. Those who stay in the game will reap the benefits – financially included.

But don’t confuse this picture with most people’s reality. Like in life – some go barefoot, some take public transport, some own economy cars, and a few drive Maseratis. The healthcare innovation market will be structured the same way. This perspective is unpopular in Europe’s current center-left ideology and public health institutions, where every quality label must have an “equal access” sticker. But in reality, some digital health tools will “build new roads” for some, improve “public transit” for others, and offer futuristic “premium vehicles” for the few.

Just like life itself.

Developers who understand their niche, audience, payer, and required resources will be far more likely to survive the selection process and reach the finish line – as winners.

To request a correction, please send an email.