Healthcare AI sandbox points to the next regulatory test: Autonomous medicine

Artificial intelligence in healthcare is moving from decision support towards something more consequential: regulated systems that can perform defined clinical tasks with a degree of autonomy. This shift is now being tested in Israel, where the Israel Innovation Authority and the Ministry of Health have launched a regulatory sandbox for artificial intelligence in healthcare.

The first cohort comprises three Israeli health technology companies developing AI applications for pregnancy care, heart failure management and foetal ultrasound assessment. The programme is designed to test highly autonomous medical AI systems in real clinical settings while regulators work out how such technologies should be assessed, controlled and authorised.

The development matters because much of the discussion around medical AI has so far centred on tools that assist healthcare professionals: software that flags suspicious radiology images, highlights risk factors, supports triage, or helps clinicians interpret data. The Israeli sandbox is directed at the next stage: systems that may independently perform parts of a clinical workflow, make recommendations under predefined protocols, or reduce the need for every routine output to be reviewed by a physician. This is where the regulatory questions become sharper. Who is accountable when an algorithm acts? How much human oversight is sufficient? What evidence is needed to demonstrate safety across different patient groups and care settings?

A regulatory sandbox is not a relaxation of standards. Properly designed, it is a structured environment in which companies, hospitals and regulators can test technologies that do not sit neatly within existing rules. Israel’s version aims to identify regulatory barriers, evaluate safety and effectiveness, generate clinical evidence and establish pathways for adoption in both domestic and international markets. For a small but highly digitised healthcare system with a strong healthtech sector, this is a strategically logical move. It allows regulators to observe technologies as they are used in practice rather than attempting to write rules in the abstract.

The three selected companies illustrate where autonomous or semi-autonomous medical AI is heading. Pulsenmore will work with Rabin Medical Center’s Beilinson Hospital on an AI system for analysing at-home ultrasound examinations performed by pregnant women. The longer-term regulatory question is whether such a system could make autonomous clinical determinations for routine scans without requiring physician interpretation of every examination. Pulsenmore has also stated that its home ultrasound platform has already supported more than 250,000 home ultrasound scans, creating a substantial real-world dataset for developing and validating AI applications.

Cordio Medical will pilot an AI-enabled heart failure management system with Tel Aviv Sourasky Medical Center. The system combines voice analysis, questionnaires, smartphone data, smartwatch data and electronic medical record information to detect early clinical deterioration and recommend medication adjustments according to a protocol defined by the treating physician. This moves AI beyond passive monitoring into therapeutic intervention support, raising questions about explainability, escalation, physician control and the boundary between clinical advice and automated care.

The third company, Simahook, will work with Hadassah Ein Kerem and Hadassah Mount Scopus medical centres on an AI-guided fetal weight assessment system. Here, the regulatory issue is partly one of task-shifting. If intelligent guidance allows a broader group of healthcare professionals to perform ultrasound assessments safely and accurately, this could help address workforce shortages. However, any expansion of who performs a clinical task must be supported by evidence around training, competency, quality assurance and diagnostic reliability.

Healthcare systems everywhere are under pressure from ageing populations, clinician shortages, rising costs and growing patient expectations. AI tools that can automate routine workflows or support home-based care are attractive because they promise scale. Yet healthcare is not a sector where innovation can be judged only by speed. Regulators must consider bias in training data, cybersecurity, model drift, clinical validation, human factors, adverse event reporting and the possibility that automation changes how clinicians behave. The World Health Organization has warned that AI in health must be governed with ethics, human rights, transparency and accountability at its core.

Canadian healthcare

Canada provides an important parallel. Health Canada has issued guidance for machine learning-enabled medical devices, setting out expectations for design, risk management, data selection, development and training, testing and evaluation, clinical validation, transparency and post-market monitoring. The Canadian framework also incorporates the idea of a predetermined change control plan, a mechanism intended to manage planned future changes to machine learning systems while maintaining regulatory oversight. This reflects a common challenge: AI-enabled medical devices may need to evolve after approval, but uncontrolled evolution is incompatible with patient safety.

The Canadian approach is especially relevant because it points to the need for lifecycle regulation rather than a one-off approval model. A traditional medical device may be relatively stable once marketed. AI software can change through retraining, updates, expanded datasets or adaptation to new clinical environments. This means the evidence package must extend beyond initial performance claims. Regulators need to know how performance will be monitored, how drift will be detected, what changes are permitted, when new review is required and how users will be informed.

In Europe, the regulatory picture is becoming more layered. The EU AI Act entered into force in 2024 and applies a risk-based framework to artificial intelligence. AI systems used as medical devices or as safety components of regulated medical devices are generally treated as high-risk, creating obligations around data governance, technical documentation, human oversight, transparency and post-market monitoring in addition to existing medical device requirements. For companies seeking global markets, this means that clinical evidence, quality systems and AI governance will need to be designed with multiple regulatory regimes in mind from the outset.

Israel’s sandbox should therefore be seen less as a local experiment and more as part of a global regulatory learning process. Its value will depend on whether the pilots generate transferable evidence: how autonomy is defined, how risks are controlled, how clinicians interact with AI outputs, how patients understand automated care, and how responsibility is allocated when decisions are shared between software and professionals.