Health systems are pouring billions into AI – but are we measuring the right things?

In 2025, global investment in healthcare AI exceeded $18 billion. Yet most health systems still cannot answer a deceptively simple question: what are we actually trying to achieve with this investment?

A new white paper published by the World Economic Forum and the AI and Digital Unit at LSE Health argues that the defining challenge of health AI is not technological. It is a governance problem, and the consequences of getting it wrong are already visible.

“Health systems are under considerable fiscal pressures, even without the expectation to embrace AI technologies. While the technology is undoubtedly powerful and has clear use cases in healthcare, not having a clear view of where to use it and for what purpose risks an inefficient use of resources – further jeopardising an already struggling system.”

The value inversion

The paper introduces a concept at the centre of the conundrums AI deployment in health brings: the "value inversion." The outcomes most convincingly demonstrated so far – documentation time saved, billing codes captured, appointment volumes increased – are those closest to the technology itself. Whether any of this translates into better patient health, reduced inequity or a more sustainable workforce remains largely unmeasured. This is because those making procurement decisions are not the people bearing the consequences of getting them wrong. The higher the governance layer at which an AI adoption decision is made, the lower the evidentiary bar applied.

This mirrors a pattern playing out across industries. A 2026 National Bureau of Economic Research study of nearly 6,000 senior executives found that while 70% of firms actively use AI, more than 90% reported no measurable impact on productivity. In healthcare, the stakes are not quarterly returns: they are human lives.

The risks hiding in plain sight

The real-world evidence emerging from clinical AI deployments is more complicated than the headlines suggest:

• A multicentre study in The Lancet Gastroenterology and Hepatology found that endoscopists regularly exposed to AI-assisted colonoscopy showed a measurable decline in unassisted diagnostic performance: the first clinical evidence of AI-driven deskilling in medicine
• A randomised trial found that even clinicians who completed a structured AI literacy programme remained susceptible to over-reliance on AI outputs
• Ambient AI scribes have reduced documentation burden in real health systems, but a recent JAMA editorial noted that healthcare has become adept at measuring time saved, and far less equipped to measure what happens to health equity and patient experience as these tools become the default

Four principles for getting it right

Rather than calling for a slowdown in innovation and adoption, the paper argues for redirection. It sets out four principles:

1. Meaningful outcomes must be identified bottom-up, from patients and frontline professionals, then pursued top-down through procurement and accountability frameworks
2. Public-private collaboration must move towards co-creating outcome definitions that are clinically meaningful, commercially viable and governance-ready
3. Technological sophistication matters, but is secondary to pursuing the right outcomes
4. Evaluation frameworks must work for contexts of scarcity as well as abundance, recognising that in many parts of the world, the alternative to AI is not imperfect care, but no care at all

In parts of Indonesia, roughly 300 million people are served by approximately 1,000 psychiatrists. The evidentiary bar appropriate for a well-staffed European hospital is not the right bar everywhere. Getting this wrong does not just slow progress, it entrenches existing inequities.

The race to deploy AI in healthcare is moving faster than our ability to define what success actually looks like. Today, we measure what is easiest to count, like documentation time saved or visits added, while the outcomes that matter most to patients and the professionals who care for them remain largely unmeasured. The healthcare systems that lead the AI era will be those that bring patients, clinicians, developers, payers, investors, and policymakers together to define the outcomes that truly matter, and to build the accountability to achieve them. A multisectoral approach is paramount in making impactful progress and to scale use of AI in healthcare.

Who defines value, and for whom

The paper is ultimately a challenge to a comfortable assumption: that deploying more sophisticated AI will, by itself, improve healthcare. What it asks of health system leaders is more demanding: to pause not the technology, but the assumptions underneath it. This work was carried out by Robin van Kessel, Co-Director of LSE Health's AI and Digital Unit, as part of his Hoffmann Fellowship on the Future of Health Systems at the World Economic Forum, with contributors from Kaiser Permanente, Google DeepMind, Roche Diagnostics, Royal Philips and Duke Margolis Centre for Health Policy, among others.

Meaningful Outcomes Determine the Winners of the Health AI Race is available in full via the World Economic Forum.