Five Principles for Bringing AI into Israel’s Public Sector

Israel is a global supplier of AI solutions, yet when it comes to applying them internally, it lags behind. We know what to do, we just don’t have the machinery to do it.

The obstacles are familiar. Ministries take a “territorial” approach to information and are reluctant to share it with one another. The Mandatory Tenders Law imposes procurement timelines that cannot keep pace with AI. And the National Digital Agency lacks the authority to compel ministries to act. Without parallel organizational reform, investment in technology may yield only limited benefit.

The five principles below are not theoretical ideas. Each is based on a proven model that has been tested in one or more countries, and together they set a practical order of priorities for the national AI directorate, established in September 2025 within the Prime Minister’s Office following the Nagel Committee’s recommendations.

1. Build a shared data infrastructure, with clear obligations to participate

The component that every expert points to as a necessary condition for adopting AI in the public sector is a unified data infrastructure that allows secure information exchange between ministries.

Estonia’s X-Road offers a useful model. Rather than creating a single central database, Estonia chose a federated model: each ministry retains ownership of its data while allowing other ministries encrypted access. Launched in 2001, X-Road is relatively inexpensive to implement and has been found to save the Estonian public sector thousands of working hours a year. It now supports more than 3,000 public services in Estonia alone and has been adopted by more than twenty countries.

For citizens, the practical significance is straightforward. An Estonian citizen applying for a benefit does not need to collect documents from other ministries: the system brings the data together and makes it available.

Israel does not need to reinvent the wheel. In 2019 it established a comparable infrastructure, the Government Information Highway. Yet a November 2024 State Comptroller report found that only 9% of government services had actually been connected to it. The Comptroller stated explicitly that the root of the failure is not technological but institutional: the National Digital Agency has no authority to compel ministries to connect, and they have not done so.

The required work therefore has both a technical and an institutional dimension. A dedicated data-cleaning team within the national directorate should map existing datasets, identify problems, remove duplication and establish common structures. The most critical component, however, is primary legislation that obligates the ministries that hold information “territorially”, such as Health, the National Insurance Institute and the Tax Authority, to share it, balanced against the Privacy Protection Law, and backs that obligation with an enforcement mechanism.

Without a binding obligation to share, the failure of the Information Highway will simply repeat itself in the age of AI.

2. Adapt procurement to the pace of AI development

Under the Mandatory Tenders Law, a typical government tender takes eight to eighteen months, and that is only the first stage of a long procurement process. This is a structural barrier: by the time a tender for a language model is completed and a contract signed, the model on offer will be obsolete.

The solution starts with amending the Mandatory Tenders Law to allow open framework agreements. Any supplier that meets basic criteria, including information security, financial stability and proven experience, would join a pool of approved suppliers within a few weeks, without a separate tender for each order. Ministries could then buy from the pool through a short “mini-tender” lasting only weeks. This model has been widely adopted in the British and American public sectors: Britain’s G-Cloud framework has registered more than 5,000 suppliers since 2012.

Usage-based purchasing would complement this approach. Instead of committing in advance to an annual usage ceiling, a ministry would pay according to actual consumption, such as the number of data-sharing requests, data volume or computing hours. This suits the early stages of adoption, when the scale of use is not known in advance.

Alongside these purchasing routes, Israel needs regulatory sandboxes. These are controlled environments in which the state allows companies to test a new technology for a limited period under close supervision. They let both companies and regulators experiment with the technology, which can lead to lighter and more effective regulation. Israel has already started down this path: in 2025 the Innovation Authority and the Ministry of Education opened a NIS 10 million sandbox for personalized learning, and in 2026 the Ministry of Health followed with one for medical AI. The next step is to extend the model to the National Insurance Institute, local authorities and the Tax Authority.

3. Give public employees consistent access to secure AI tools

Singapore offers the model: building an organizational infrastructure that allows public servants themselves to lead AI adoption in their work. Its government technology agency, GovTech, developed and distributed three main tools across the public service: Pair, a general AI assistant for writing, summarizing, translation, document processing and research; SmartCompose, a writing tool; and AIBots, a platform for building chatbots. By the second quarter of 2025, Pair had some 64,000 monthly active users, about 42% of Singapore’s civil service.

Britain completes the picture. In February 2025 it published an AI Playbook, a detailed technical document that sets out rules for the safe and effective use of AI by public sector employees. A trial among 20,000 British civil servants found that each saved close to two working weeks a year.

Information security is a precondition. Three common computing architectures balance AI capabilities with data protection and should become standard in Israel’s public sector: open-source models, such as Llama or Mistral, running directly on closed government servers; a private government cloud, in which data is stored and processed only within a defined jurisdiction and is not used to train generic models; and public cloud computing, with encryption and masking applied before data is transferred.

Israel can choose between three routes: mass licensing of tools from companies such as Microsoft and Google; developing an Israeli model with the private sector, as the Nagel Committee recommended; or combining international systems with an Israeli adaptation layer. The decision depends on cost and information security requirements.

4. Make training and measurement part of management

Classic management literature points to the need for structured, quantitative targets. Without them, even excellent recommendations remain declarations.

Each ministry should have measurable training targets. For example: 15% of employees trained in the first quarter, 30% by the end of the first year and 70% by the end of the second. Each ministry’s progress should be published on the Civil Service Commission’s website as part of its existing Digital Index. Basic AI literacy should include an understanding of models’ capabilities and limitations, practical instruction in formulating requests, methods for identifying errors and information security requirements.

Each division should also have an “AI champion”, similar to the safety and privacy officers ministries already have. This employee, not necessarily a technologist, would help colleagues formulate prompts, identify tasks suited to AI and find the relevant tools. Champions would receive 40 to 60 hours of in-depth training a year and form a cross-ministry knowledge network. Similar models operate in Singapore and Britain, and the network could be strengthened by a regular forum modeled on Singapore’s AI Wednesdays.

Gradually, basic AI literacy should also become a requirement in internal tenders for managerial and senior professional roles. This would turn training from a voluntary bonus into part of the standard career path, and signal to employees that investing in learning pays off. Careful implementation is needed to avoid excluding experienced employees with extensive professional knowledge.

Finally, ministries should measure actual use. Monitoring systems for AI tools can track the share of active users, frequency of use and types of use, and this information should be part of the management data delivered to each ministry’s head and to the national directorate. The British experience is instructive: the active usage rate was about 72%, and employees who built the tool into their daily work saved three times as much time as those who used it occasionally. The purpose of measurement is not coercive supervision, but to help build an organizational culture in which learning AI is a standard part of public sector work.

5. Make oversight concrete and preserve access to human review

The “human in the loop” principle, and the need to guard against algorithmic biases, including political ones, require a concrete oversight framework. The Ministry of Innovation, Science and Technology’s 2023 policy document already embraces these principles, listing human involvement, with ongoing oversight and the right to appeal to a person, alongside transparency, explainability and accountability. Three tools, each already proven abroad, can put this framework into practice.

The first is an algorithmic impact assessment before deployment, such as the one adopted in Canada. The Canadian directive requires every federal department to conduct a structured assessment before deploying an automated system for administrative decisions. The assessment classifies the system into one of four risk levels, each with its own requirements for oversight, transparency and appeal. The Canadian tool consists of 106 questions that translate into a numerical score, and it is open to public scrutiny.

The second is a public model card: a document of one to three pages accompanying every government AI system. It describes the model and its version, the data it was trained on, its intended uses and uses that are not recommended, expected error rates for different population groups, and who is responsible for inquiries. The concept was proposed by Google researchers in 2019 and has since been adopted by many companies and public bodies.

The third is a right to human review. Any citizen affected by a decision made by a government AI system should be entitled to have it examined by an official who is not bound by the system’s outcome. This right is anchored in Article 22 of the GDPR and reinforced by the EU AI Act, and it is the practical translation of the human-in-the-loop principle into a civil right.

In Israel, all three could be anchored in legislation or in an amendment to Government Resolution 3375. Far from slowing adoption, they would enable it by building the public trust without which any broad initiative will meet resistance, and they would give ministries a clear framework in place of the legal uncertainty that holds initiatives back.

Unlocking Israeli innovation

This is where the gap is sharpest: there is no clear route for an Israeli startup to become a supplier to a government ministry. Israel already has some of the pieces, including pilot programs and regulatory sandboxes, but they lack coordination, move too slowly for startups and offer no clear path from pilot to procurement.

Three international models show how to close this gap. The American SBIR program funds companies in stages, from feasibility through prototype, to a final stage in which the federal agency itself becomes a customer of the technology; a large percentage of DARPA’s tenders go to companies that came through it. G-Commerce, an American platform built on Glass’s system and launched in 2026, lets public bodies buy AI tools through contracts that have already been reviewed and approved in a joint tender, instead of each body running its own. And Singapore’s AI Trailblazers program, run with Google, has supported 84 organizations since its launch, each of which builds a prototype in roughly 100 days.

Building on these models, the national AI directorate can run a four-stage bridge between Israeli startups and the public service. First, needs: together with the main ministries, the directorate publishes an annual list of specific organizational needs, framed as open descriptions of the problem rather than closed tender specifications. Second, a sandbox: startups propose solutions, and the directorate approves five to ten companies for a six-to-twelve-month pilot in a sandbox covering privacy, information security and legal liability, with an initial allocation of NIS 1 to 3 million per company. Third, evaluation: an independent body measures each pilot against predefined criteria, such as productivity gains, cost savings and service quality; results are published, and the best-performing companies are selected to provide the services. Fourth, export: companies that have gained experience in Israel’s public service receive government support to export their solutions, creating both more efficient services in Israel and technology exports that grow Israeli high-tech.

There will be obstacles. The Mandatory Tenders Law makes it difficult to choose a specific supplier, even after a successful pilot; the solution is a legal amendment allowing direct procurement for AI initiatives that have completed a successful pilot, as the United States does through Other Transaction Authority. Israel’s strict privacy law limits the transfer of citizens’ data to external companies; as the Nagel Committee recommended, the law should be adapted to allow regulated use within a sandbox. This does not mean abolishing privacy protection, but creating a legal framework for supervised experiments. Public-sector wage and collective agreements may limit the ability to change work processes, which requires a firm stance that puts the public interest first. And each ministry will prefer its own pilots over a coordinated framework; the answer is clear political ownership of the national directorate by the prime minister, with a public commitment to measurable results.

From technological capability to public benefit

The five principles form a deliberate sequence. Shared data is the precondition for everything else. Agile procurement and sandboxes provide the legal infrastructure for effective action. Unified AI tools bring the technology into daily work. Training and measurement ensure that adoption does not remain with individuals but becomes a cross-government capability. Oversight ensures that public trust is not harmed along the way.

International comparison does not offer a single model to copy, but different elements from which Israel can draw. Israel operates in unique circumstances: a large public sector with strong unions, a fragile political structure and a leading AI industry whose achievements have not yet been translated into domestic applications. Moving from declarations to practice will require patience, political continuity and primary legislation that anchors the basic principles of government use of AI.

The AI revolution is not only technological. It is also a managerial and organizational revolution, and its successful adoption depends on redesigning organizational structures, overcoming territoriality between ministries and maintaining transparency toward citizens. Investment in technology is a necessary condition, but not a sufficient one. The most telling measure of the national directorate will not be how many billions are allocated to infrastructure, but how many Israeli startups manage to cross the bridge from the lab into the public service within its first two years. 

1.  Frequently Asked Questions:

What are the five principles for bringing AI into Israel’s public sector?

The five principles form a deliberate sequence. Shared data is the precondition for everything else. Agile procurement and sandboxes provide the legal infrastructure for effective action. Unified AI tools bring the technology into daily work. Training and measurement ensure that adoption does not remain with individuals but becomes a cross-government capability. Oversight ensures that public trust is not harmed along the way.

Why does Israel lag behind in using AI inside government?

The obstacles are familiar. Ministries take a “territorial” approach to information and are reluctant to share it with one another. The Mandatory Tenders Law imposes procurement timelines that cannot keep pace with AI. And the National Digital Agency lacks the authority to compel ministries to act.

What is Estonia’s X-Road?

Rather than creating a single central database, Estonia chose a federated model: each ministry retains ownership of its data while allowing other ministries encrypted access.

Why has Israel’s Government Information Highway fallen short?

The Comptroller stated explicitly that the root of the failure is not technological but institutional: the National Digital Agency has no authority to compel ministries to connect, and they have not done so.

How can government procurement keep pace with AI?

The solution starts with amending the Mandatory Tenders Law to allow open framework agreements. Any supplier that meets basic criteria, including information security, financial stability and proven experience, would join a pool of approved suppliers within a few weeks, without a separate tender for each order. Ministries could then buy from the pool through a short “mini-tender” lasting only weeks.

What is a regulatory sandbox?

These are controlled environments in which the state allows companies to test a new technology for a limited period under close supervision. They let both companies and regulators experiment with the technology, which can lead to lighter and more effective regulation.

What is an algorithmic impact assessment?

The Canadian directive requires every federal department to conduct a structured assessment before deploying an automated system for administrative decisions. The assessment classifies the system into one of four risk levels, each with its own requirements for oversight, transparency, and appeal.

What is a public model card?

The second is a public model card: a document of one to three pages accompanying every government AI system. It describes the model and its version, the data it was trained on, its intended uses and uses that are not recommended, expected error rates for different population groups, and who is responsible for inquiries.

How could Israeli startups become suppliers to government ministries?

Building on these models, the national AI directorate can run a four-stage bridge between Israeli startups and the public service.