A personal health agent wants to read your laboratory results, check previous prescriptions and arrange a follow-up. A hospital capacity agent wants to forecast beds and coordinate patient movement. A research system wants to learn from records across several institutions.
All three scenarios depend on access to healthcare information. But they do not need the same data, permissions or level of independence.
At Web Summit 2026, Portuguese state-reform minister Gonçalo Matias connected the government's AI ambitions to cloud infrastructure and sovereignty. He argued that public services should know where data resides, which laws protect it and how essential services continue operating if something goes wrong. He also outlined a future in which state agents could communicate with one another and perform authorised tasks for citizens.
The speech was about public services broadly. The healthcare implications are substantial, especially as vendors describe systems that can observe information, make recommendations and increasingly take action on a patient's or organisation's behalf.
Location is only one part of control
It is easy to reduce data sovereignty to a map. If information sits on a server inside a country, the thinking goes, it must be safe and controlled locally. Geography matters, but it is not the whole picture.
Who administers the system? What permissions can support staff exercise? Can information be transferred to sub-processors? Which jurisdictions have legal authority over the provider? What happens if a contract ends, the provider suffers an outage or the health service wants to migrate away? Can the organisation retrieve complete records and audit logs in a usable format?
Those questions determine whether a hospital has practical control over its data. A local server with poor access management is not automatically safer than a carefully governed service elsewhere. Conversely, an attractive cloud product does not excuse vague answers about cross-border processing and access to sensitive records.
Healthcare organisations need a coherent answer covering technical controls, contractual obligations, governance and operational resilience.
Health information has a different risk profile
A consumer app can learn someone prefers window seats and vegetarian food. A healthcare agent may learn intimate details about symptoms, medication, family history, genetic findings and diagnostic uncertainty. Some data can change a person's employment, insurance or social circumstances if mishandled.
Under the GDPR, health information belongs to a special category of personal data, subject to additional protections and conditions for processing. The European Data Protection Board emphasises the sensitivity of health and research data.
That does not mean healthcare information can never be used by an AI system. It means the legal basis, purpose, safeguards and responsibilities must be defined carefully. Consent is not a universal shortcut that makes every possible future use acceptable, and a friendly chatbot interface does not replace proper information governance.
An agent that arranges a consultation, for example, may need a limited set of scheduling details. It should not automatically receive the patient's complete clinical record. A different system helping a clinician review treatment history may require broader access, but only within an appropriately authorised workflow.
The principle is straightforward: permission should follow the job.
The complication of agents acting for people
Matias outlined a future in which a citizen would no longer need to move the same information from one public office to another. With permission, agents could coordinate services behind the scenes. The prospect is attractive. It could reduce duplicated paperwork and spare people from repeating information already held within the state.
In healthcare, however, autonomy requires careful boundaries. Suppose an agent notices that a test was completed and schedules a follow-up. Was the appointment clinically appropriate? Did the patient authorise the action? Was the receiving service permitted to view the result? What happens when the patient changes their mind, or when the test is later corrected?
Delegation is not a single yes-or-no switch. A patient might allow an agent to retrieve a laboratory report but not to share it with an employer or book a diagnostic procedure. A clinician might approve drafting a referral but require a final review before submission. The system should make these distinctions visible and enforceable.
It also needs an audit trail. A person should be able to find out which system accessed a record, when it happened, what information was used and which action followed. If an error occurs, responsibility needs to be traceable.
Europe's data infrastructure is still being built
The European Health Data Space is intended to improve people's access to their own electronic health information, support safer exchange for care and enable regulated secondary use for research and policy. It entered into force on 26 March 2025, but implementation is phased rather than immediate.
The European Commission describes 2029 as the planned milestone for the first group of cross-border priority data categories, including patient summaries and electronic prescriptions. Medical imaging, laboratory results and hospital discharge reports are planned to follow in 2031.
These dates matter because an agent cannot provide truly longitudinal support if essential records remain inaccessible or inconsistent. Interoperability is not simply the ability to display a PDF. A useful system needs to identify the correct person, understand the provenance and timing of records, and distinguish current information from historical data.
Standardisation also has to allow for clinical nuance. A medication may be listed as prescribed but no longer taken. A discharge summary may include a provisional diagnosis. A result may have been corrected after the original report. Data quality, context and reconciliation remain essential even when the systems can technically exchange information.
The problem of vendor dependence
One of the less glamorous risks is lock-in. A hospital may build useful workflows on a proprietary platform and later discover that exporting configurations, data or audit records is difficult. For an agentic system, dependence may extend to model providers, tool integrations, identity infrastructure and orchestration software.
A sensible procurement process should therefore assess portability before the product becomes critical. Can another supplier take over? Are records stored in documented formats? Can the hospital disable one integration without breaking the entire service? What happens if the provider changes its data-processing terms or pricing?
The answers influence patient safety as well as finances. An unavailable system is particularly serious when staff have come to rely on it for time-sensitive coordination.
Data sovereignty without exit options is a fragile kind of control.
Trust has to be demonstrated in routine use
Policies and legal declarations are necessary, but patients and clinicians judge trust through experience. Can they see who accessed information? Can they correct an error? Can a patient refuse an optional secondary use without losing essential care? Does the system explain when it lacks enough information to act?
The WHO's guidance on AI for health stresses governance, reliability and stakeholder involvement. Those are not peripheral ethical questions. They influence whether people will use a system as intended and whether institutions can deploy it responsibly.
Founders should make data governance a product feature rather than a late-stage compliance appendix. Clear permission controls, provenance, logging, exportability and incident response should shape the architecture early. They become harder and more expensive to bolt on when the product already has customers and complex integrations.
The question behind the infrastructure
Matias described sovereignty as a combination of geography and values. That framing is especially useful for healthcare because legal protection, patient autonomy and reliable service matter more than any single cloud location.
A future in which agents coordinate medical information could spare patients time, reduce repeated administration and improve continuity. But the system must be able to answer a basic question at every step: who allowed this action, using which information, for whose benefit?
If HealthTech companies can give clear and verifiable answers, intelligent services may strengthen trust. If they cannot, adding more autonomous features will only increase the distance between what the technology can do and what patients feel safe letting it do.
Key Signals
Data sovereignty is not only about where records are stored; it is about who can use them, for what purpose, and with whose authority, and AI agents make all three questions urgent at once. The European Health Data Space creates the rails for sharing, but consent and purpose limitation still decide what may flow across them. Products that cannot answer the authority question clearly will not survive procurement in regulated health systems.
