AI in Portuguese Healthcare: What Should Portugal Want It to Do?

The debate over AI in Portuguese healthcare usually starts with what the technology can do. The sixth episode of Saúde em Perspetiva (Health in Perspective) asks the harder question instead: what should Portugal want it to do inside its National Health Service?

The podcast, produced by the Nova SBE Health Economics & Management Knowledge Center, released the episode on 9 September 2026 as part of its “Debates claros sobre decisões difíceis” (Clear debates on difficult decisions) series. The project sits under the BPI | Fundação “la Caixa” Chair in Health Economics, an Initiative for Social Equity created in partnership between Fundación “la Caixa”, Banco BPI and Nova School of Business and Economics.

Hélia Bernardo, science communication lead at Nova SBE, moderated the conversation between José Miguel Diniz, a physician at ULS São José, RISE-Health researcher and doctoral student in health data science at the Faculty of Medicine of the University of Porto who coordinates the IA@SNS group, and Teresa Soares, a public sector and health consultant at NTT DATA Portugal and a member of the same group. Quotations below are translated from Portuguese.

Ad

What artificial intelligence is, and what it is not

Diniz opened with a definition he described as more legal than technical. Artificial intelligence, he said, is a machine-based system that converts inputs into specific results, holds a stated objective, can act with varying degrees of autonomy, and often learns as it interacts with its users. The consequences are real, whether they land online or in the physical world.

“We talk about it in a very broad way, and this is a definition that is more legal than technical: it is a machine-based system that essentially converts input data, be it sensor data or anything else, into specific results.”

He drew a sharp line between AI and the rule-based tools already common in health. Software and electronic forms that follow instructions from a human operator are not AI, he stressed. Real AI tools in health include support for triage, analysis of medical images, and help for clinicians making decisions.

The distinction matters because AI systems are trained on data and adapt over time, whereas rule-based tools stay fixed to the rules they were given.

AI in Portuguese healthcare: still more conversation than impact

For all the talk, Diniz said the technology has not yet changed much in daily practice.

“In my particular work there have been relatively few changes. But if we extend this to the other health professionals, the truth is that at the moment it is more a question of conversation than of impact.”

He pointed to the national strategy presented by SPMS, the shared services arm of Portugal’s Ministry of Health, which he said lists fewer than a dozen initiatives under exploration for the SNS. They include support for triage on the SNS24 contact line and help with diagnostic coding to build administrative data. Individual use of large language models such as ChatGPT is growing, he added, but it has not yet been integrated into care delivery.

Both guests agreed on where AI should start: the non-clinical and administrative work that currently absorbs clinicians’ time. Diniz offered the example of reading a CD of an exam performed outside the hospital, a step that often fails and forces manual rework, and said automating such tasks could reduce the variability these systems introduce.

Interoperability comes first

For Soares, integrating AI well depends first on interoperability. Hospitals and other SNS institutions run many applications, each with its own task, but the systems often do not talk to each other.

“We have different systems with different tasks, but they do not communicate with each other. It is not beneficial to have several systems that allow us to do several things if we do not guarantee communication between them.”

Diniz added a caution that applies to any new device, from an X-ray machine to a laboratory test. Artificial intelligence should not be dropped into every stage of a care pathway at once. Different professionals use different languages, and each step has to keep its own quality standard while preserving continuity of care.

When not to use it

The pair also spent time on the limits. Diniz argued that health systems need to name the errors they will not accept, the red lines. Most people, he said, would agree on goals such as minimizing hospital deaths and diagnostic errors. The complication is that AI output is not deterministic in the way a simple sum is.

“We do not have a determinism of what is going to happen. There is a volatility in what it produces, and then we have to compare that with what we do today and with what we are willing to accept.”

If a given outcome must never happen, he said, the system has to spend resources auditing how these tools behave, which can itself become counterproductive.

Soares placed the root of the question earlier still. Before AI is applied, expectations and needs have to be aligned, and the existing process has to be mapped. The SNS is not just patients and clinicians, she noted; it also includes managers and administrative staff, and all of them have to be heard before a tool replaces part of a circuit.

Value, equity and the question of replacement

On value, Diniz said AI should be judged the way any other intervention is judged: whether the investment is justified against the system’s priorities. The objective matters. If the goal is budget containment, the tools might be pointed at fraud detection and cost avoidance. If it is prevention, the answer looks different.

On equity, both guests returned to the same foundations: interoperability, the involvement of every stakeholder, and the recognition that regions differ. Populations carry different chronic disease burdens, and needs change over time, so a tool fitted to one area will not transfer unchanged to another. Regular monitoring is not optional.

On the question everyone asks, whether AI will replace health professionals, Diniz gave a measured answer. The evidence, he said, is fairly consistent that the tools can outperform medical students on the multiple-choice exams used to assess them, and even some professionals.

“What we find is that, for multiple-choice questions with four or five options, the tools perform quite well, and quite a bit better than most students and perhaps some health professionals.”

But he warned against reducing the complexity of care work to exam scores. The immediate opportunity is to take away work done inadequately or unnecessarily, including administrative tasks. He cited a UK tool that triages photographs of skin lesions with high sensitivity and specificity, referring only the patients most likely to need care, and noted that Portugal has related research projects. For now, he said, the idea that AI will leave clinicians without work runs into a simpler fact: there are not enough health professionals to meet the population’s needs, and those professionals are needed to train and validate the systems.

The real dangers

Diniz named the biggest danger as forgetting what these systems really are. In health, he said, AI tools are usually digital medical devices, and that means a regulatory approval path requiring a clear demonstration of benefit, much like a medicine or a prosthesis.

“There is a great temptation to forget that these are tools, and in the case of health in particular we are talking about medical devices, normally digital medical devices, and that has a regulatory approval process that requires the demonstration of results.”

He also flagged more insidious risks: a growing dependence on the tools, a quiet shift of responsibility away from their users, and the unknown unknowns of what information the systems are given and what consequences they are allowed to trigger.

A risk of avoiding the harder reforms

The episode closed on a question that matters for a public system under pressure: whether AI could be used to avoid deeper reforms. Diniz was direct.

“Absolutely. We already have a certain culture of adding layers to solve the problem.”

If the system is willing to spend large sums and staff on AI, he said, it might gain more from investing in care quality, prevention and health promotion instead. Soares offered a different reading: because implementing AI forces institutions to listen to patients and characterize their populations, it could become a way to surface needs that are otherwise ignored, provided the systems are always adapted to the people they serve.

The episode ends where it began. AI can support clinical decisions, improve processes and help a public system that serves millions of people, but it only makes sense in the service of a clear purpose. The real question, the hosts concluded, is not what artificial intelligence can do. It is what Portugal wants it to do inside its National Health Service.

Source: Saúde em Perspetiva, Nova SBE Health Economics & Management Knowledge Center, episode six, 9 September 2026.