AI and Digital Health Regulation and the Israeli Regulatory Sandbox

1. Key takeaways

Israel has no dedicated legislation governing medical artificial intelligence. AI and machine-learning driven clinical tools are managed under the general medical device regulations, applying reference-country reliance and the three registration routes. The Ministry of Health issued guiding principles for the development of digital health AI technologies in April 2023, which are non-binding in nature. In 2025, the Ministry of Health together with the Israel Innovation Authority jointly launched a request for information on a regulatory sandbox, seeking views from companies, start-ups, research institutions and healthcare service providers, with the submission period closing in April 2026. If the sandbox mechanism takes shape, the route to approval for autonomous medical AI could undergo a structural change. Until then, companies should plan on the basis of the existing device framework, treating any sandbox-driven changes as upside rather than as the planning baseline.

2. Legal basis

The table below lists the current non-binding instruments, the applicable mandatory framework, and the parallel privacy obligations relevant to AI products.

3. Expert insights

Israel’s orientation in medical AI regulation is to observe first and legislate later. This orientation produces three characteristics:

Characteristic one: the existing system is sufficient to cover the field. Because AI devices are managed as general devices, and the Israeli general device system is built on reference-country reliance, review conclusions on the algorithm by the FDA or a notified body can be accepted directly. This means Israel does not need to build a separate review capability for AI, and it also means the quality of the reference-country file largely determines the Israeli outcome. The practical corollary is that an algorithm change approved abroad should be carried into the Israeli file with the same discipline as any hardware change, since the reliance mechanism transmits both the approval and its scope.

Characteristic two: privacy and data protection are the substantive constraints. The Protection of Privacy Law and the 2017 data security regulations apply across the entire lifecycle from model training to deployment, and the Israeli Privacy Protection Authority takes an active enforcement position on health data. For AI companies, compliance pressure on this line is usually greater than the pressure of device registration itself, because training data provenance and consent records are difficult to retrofit. A dataset assembled before the compliance framework existed cannot be repaired by documentation written afterwards, so the privacy work has to begin at the point the data is collected, not at the point the product is submitted for registration.

Characteristic three: the sandbox is the future variable. The problem the sandbox is meant to solve is that autonomous medical AI (such as autonomous diagnosis or closed-loop control) struggles to demonstrate safety with static evidence under the traditional review framework. If it takes shape, an approval route of deployment first within a controlled environment, plus independent audit, plus full process recording, may emerge. GRZAN notes that if the sandbox is legislated, its scope of application, its duration and the rules connecting it to formal registration will determine its actual value, so the drafting details matter more than the announcement. Companies preparing for that possibility should keep their algorithm governance records in the format regulators already expect, since whatever route the sandbox eventually creates will still begin from an evidence base, and an evidence base assembled in advance is the only one available on day one.

4. Practical recommendations

• Use reference-country reliance as the main path: complete evidence generation including algorithm governance with the FDA or in Europe first, then enter Israel.
• Run privacy compliance in parallel: assess the lawful basis for training data, informed consent, transparency notices and data scraping restrictions item by item.
• Follow sandbox legislation closely: list sandbox legislation during 2026-2027 as a quarterly tracking item.
• Document algorithm governance: version control, change history, statements on training data representativeness and bias assessments belong in the technical file, retrievable on demand.

5. Verification checklist

• □ The device classification of the AI/ML product has been determined (whether the intended use falls within diagnosis, prevention, monitoring or treatment)
• □ Reference-country approval has been obtained and covers the current algorithm version
• □ The privacy impact assessment has been completed under the Protection of Privacy Law, 5741-1981, and the 2017 data security regulations
• □ Algorithm version control and bias assessment documentation are in the technical file
• □ Progress on sandbox legislation is on the quarterly tracking list
• □ Claims resting on a single source have been flagged and are not used as a basis for decisions (practical estimate)

Regulatory information verified as of 29 October 2026. The regulation names, numbers, publication dates and article numbers cited in this article were checked against official Israeli legal databases and authoritative public sources on that date. AMAR procedural documents are updated frequently, so before any formal submission please rely on the current version published on the website of the Israeli Ministry of Health.

How GRZAN can help. GRZAN provides medical device and IVD registration, quality system, authorised representative and local testing support for the Israeli market. Our team manages adverse event reporting and recall coordination with the MOH, and supports post-market vigilance obligations. Business enquiries: info@grzanglobal.com