Israel Securities Authority draft AI directive: what to do now

What the draft requires from investment advisors, investment marketers and portfolio managers, why it also reaches AI vendors without a license, and what to prepare now.

On 4 November 2026 the Israel Securities Authority (ISA) published a draft directive for public comment on the use of technological means and artificial intelligence by licensees (ISA website). It is a draft, not binding text. It is based on section 28(b) of the Regulation of Investment Advice, Investment Marketing and Portfolio Management Law, 5755-1995 (the "Investment Advice Law"). The ISA accepts comments on it until 6 December 2026.

A licensee is a business the ISA has licensed to give investment advice, market investments or manage portfolios. If you are a fintech company or an investment house, the draft sets what you will need to prove about every AI tool you run. If you are an AI vendor selling to these companies, know that they will require you to address these topics with them.

What was published, and what is its status?

The draft replaces the 2016 directive on providing services through technological means (the "existing directive"). It is framed as a "new version". It builds on an inter-ministerial report called "Artificial Intelligence in the Financial Sector", published in December 2025. The team behind it was set up at the end of 2022 with representatives of the Ministry of Justice, the Ministry of Finance, the Bank of Israel, the ISA, the Capital Market Authority and the Competition Authority.

Two services stay outside the draft for now: social trading and advice on self-directed trading. The existing rules for them continue to apply until a separate directive regulates them. Other laws, such as privacy law and anti-money-laundering law, continue to apply in parallel.

The draft is not binding, and the final text may change.

Who does the directive apply to, and what about AI vendors without a license?

The draft applies to licensees only. That means anyone holding a license under the Investment Advice Law, even when the service is delivered through an app, an algorithm or an AI system. If your company itself gives investment advice or manages portfolios, that requires a license. This is a separate question from the draft.

An AI vendor without a license is not bound directly. But the vendor's customer, the licensee, must regulate the engagement under section 34. In practice, if you are an AI vendor, you will get contract requirements, due diligence questionnaires and requests for documentation. We set out the details below.

What counts as an "AI system" under the draft?

Under section 1, an AI system is a computer-based system that operates at varying levels of autonomy. It is meant to produce outputs such as content, predictions, recommendations or decisions. And it may affect investors, clients or the licensee's operations.

The definition is broad in two ways. It is technology-neutral, so it does not depend on the type of model. And it covers internal systems with no client-facing interface, because they also affect the service indirectly.

There are two limits. Basic, common tools, such as email systems, internal messaging, knowledge-management tools and office software, are excluded. And a licensee may decide that a specific system is not AI by its nature. The ISA stresses this is not a blanket exclusion. The decision must rest on a review of the system, the purposes of its use and the risks.

Three tiers: which use of technology carries which duties?

The draft scales the duties to the risk:

  • Online-only service. Communicating with the client by electronic means instead of a face-to-face meeting, where the tool does not make or support decisions. Only a short list of sections applies (section 5).
  • Use of a technological means. An algorithm that makes decisions, or an AI system. All the duties of the directive apply.
  • Heightened use. Use that the licensee has classified as high risk and high materiality. Additional corporate governance duties apply. Only a corporation may provide this service, not an individual licensee (section 19).

The licensee does the classification itself, at its own discretion. The ISA is asking whether to turn it into fixed criteria. More on that in the section on open questions.

What will a licensee be required to do under the draft?

We grouped the main duties by topic, with the section numbers in the draft.

  • Risk and materiality assessment before use (sections 6-8). A materiality assessment checks how much the service depends on the tool. At least five parameters are weighed: reliance on the tool (8.1), scale of use (8.2), impact of a failure (8.3), human involvement (8.4), and manner of use and purpose (8.5).
  • Revalidation (sections 9-10). The assessment is revisited at least once a year, or sooner if the tool or its use changes materially. All of this is documented.
  • Written policy and procedures (sections 13-14). To manage the risks of using the tool. They are documented, and earlier versions are kept too, for 7 years.
  • Staff training (section 15). Staff must understand the system and its risks, and recognize abnormal outputs or faults.
  • Technology officer (sections 16-17). A dedicated person responsible for the technological aspects of the tool. They need not be a licensee, and may be external. According to the explanatory notes, final responsibility stays with the licensee.
  • Control and monitoring (sections 23-24). Checking the quality of data and outputs, looking for bias, discrimination and conflicts of interest, and handling abnormal outputs. It must be possible to stop or limit the tool, and there must be an alert on drift in model performance. One risk the ISA names is "hallucinations": a system that presents wrong information as fact.
  • Registry of tools (section 25). An up-to-date list of all technological means used to provide the services.
  • Retrievable documentation (section 26). Documentation of the service process through the tool, for 7 years from the date the service was provided.

For heightened use, further duties fall on the board: to approve the usage strategy, the assessments, the tool itself, the procedures and the appointment of the technology officer (section 21), and to receive periodic training (section 22). There is also a duty to report to the ISA on the start or end of heightened use and on any significant change to it (section 20).

How does the draft treat explainability and client disclosure?

Explainability is the ability to explain how a system reached an output. The draft distinguishes two kinds:

  • General explainability (section 27). Always applies. The licensee must understand and explain the tool's logic, its limits, the data it uses and the considerations it applies.
  • Specific explainability (section 28). Applies to heightened use. The licensee must explain why a particular output was produced and not another. If that is not possible, compensating measures that significantly reduce the risk are allowed, such as stronger monitoring, human approval or reduced reliance. According to the explanatory notes, if even these do not reduce the risk, the tool may not be used.

On disclosure, the draft sets four rules:

  • Disclosure before the service (section 29). Tell the client how the service differs from a service without technology, and what the risks are.
  • Notice of AI use (section 30). Clear, prominent and effective. It is not required where it is already obvious that AI is used.
  • No "AI washing" (section 31). No misleading statements about the tool's capabilities, meaning presenting it as more advanced or accurate than it is.
  • No liability disclaimers toward clients (sections 3 and 31). A licensee cannot exempt itself from liability in advertising or in the agreement, even when the tool came from an external vendor.

Three more topics are worth knowing. On information security and privacy (section 33), the ISA mentions cyber risks such as prompt injection, an attempt to slip instructions into a system's input so it departs from the guidelines it was given, and data poisoning. On human assistance (section 37), a licensee may decide whether to offer clients human help and to what extent, as long as it states this clearly to the client before the engagement. And on client suitability (section 38), it must make sure the service suits clients with limited digital literacy or limited access to online services.

Outsourcing: what does the draft require of the contract with the AI vendor?

Outsourcing is using a third party's system, model or service. The definition in the draft is broad. It covers a product built entirely on AI, an ordinary product with AI features embedded in it, and use of a product built on another vendor's model.

The licensee stays responsible even when the tool is not its own (section 34.2). Before engaging, it must assess risks and run due diligence on the vendor (section 34.3). According to the explanatory notes, this covers the vendor's reputation, knowledge and resources, concentration and single-vendor dependency risks, transparency in managing risks, and security or privacy incidents at the vendor.

Section 34.5 is the core of the chapter. It says the agreement must address eight topics:

  1. Duties to report incidents and failures.
  2. Protection of the information of the licensee and its clients.
  3. Use of sub-vendors and other parties.
  4. Material changes to the tool or to the AI model it is built on.
  5. Retention of data and documentation.
  6. Return and deletion of information.
  7. Provisions on ending the engagement.
  8. A service level agreement (SLA) for handling faults.

Three more requirements touch vendors. The licensee must hold the knowledge, information, permissions and tools to control and monitor the tool (section 34.4). So a vendor will be asked for access to logs, documentation and performance data. That is our interpretation. The licensee must also set a business continuity plan for the end of the engagement or a failure of the tool (section 34.6). And finally, using a vendor cannot impair the ISA's ability to obtain information and exercise its powers (section 34.8).

Note the consequence: if a licensee depends on a vendor so heavily that ending the engagement would stop it from continuing to serve clients, it must tell its clients before engaging them (section 34.6). For an AI vendor, that is a reason to plan an orderly exit: data export, a transition period and documentation.

Which open questions has the ISA raised?

At the end of the document the ISA lists four topics where it has not yet decided. These are directions that may change in the final text, not duties.

  • Fixed criteria for heightened use. Today the licensee classifies for itself. The ISA worries that licensees will avoid classifying use as heightened, and that application will vary between firms. It asks whether to set binding criteria, quantitative or qualitative.
  • Further requirements for the board and management. For example a minimum number of directors, expertise or dedicated committees, and a clear split of responsibility between the board and management.
  • AI agents. Systems that carry out complex tasks autonomously or semi-autonomously and use external tools and systems. The ISA asks whether they create risks the draft does not address, in control, permissions and information security. On liability for agents, see also When an AI agent gets it wrong, who pays?
  • Connecting general AI models to financial platforms. Clients connect their financial data to external generative AI tools, sometimes from the financial firm's own interface and with its encouragement. The ISA asks, among other things, whether this is an unlawful referral of clients to investment advice, contrary to section 3A of the Investment Advice Law, and what responsibility the connecting firm bears. The ISA stressed that raising the topic is not permission, express or implied, for such activity.

When does the directive take effect?

6 months after the final text is set and a notice about it is published in Reshumot (section 43). Reshumot is the official gazette. The notice has not been published yet, so there is no date. Anyone who receives a license between that publication and the effective date must meet all requirements from the moment of licensing (section 44).

Half a year sounds like a lot, but a tool registry, a policy, an officer appointment and contract updates take months. The safe assumption is to start before the final text is published.

What should you do now?

The text may still change, but the direction is clear. These steps do not depend on the final text. This is our recommendation, not something the draft says.

  • Map and classify every tool. List every AI tool and algorithm, including internal ones. Grade them by the five parameters in section 8. That tells you whether you have a use that could count as heightened.
  • Set up a registry of tools. For each tool: name, vendor, purpose, owner, date of last classification. It is also what a regulator or client will ask for first.
  • Write a policy and procedures. Decide when a person must approve an output before it reaches a client. Keep every version.
  • Appoint a technology officer. An employee or an external party, with real knowledge of the tool and its risks.
  • Go through vendor contracts. Check them against the eight topics in section 34.5. AI vendors: prepare a standard addendum that covers all of them, so you do not renegotiate with every client.
  • Check client disclosures. Make sure there is a clear notice of AI use, no liability disclaimer and no inflated claims about the tool.
  • Set up documentation for clients and the regulator. Keep logs and records of how the tool runs, in a retrievable form, for 7 years.

If you also sell in Europe, check your AI Act exposure in parallel with our AI Act checker.

Bottom line: the draft is not law yet, but it shows how the ISA wants AI to look at licensees: a known, classified, documented and supervised tool, with responsibility staying with the licensee. That is why the vendor contract becomes a regulatory document. Whoever starts on the registry and the contracts now will not be rushed when the directive is set.

FAQ

What is the ISA draft on artificial intelligence?

It is a draft directive the Israel Securities Authority published on 4 November 2026 for public comment, under section 28(b) of the Investment Advice Law. It will replace the 2016 technology directive and set rules for the use of technological means and artificial intelligence by investment advisors, investment marketers and portfolio managers. The ISA accepts comments until 6 December 2026. It is a draft, not binding text.

We are an AI vendor without a license. Does the directive apply to us?

Not directly, but yes through your customers. The directive applies to licensees. Section 34.5 requires them to cover eight topics in the contract with an AI vendor, including incident reporting, sub-vendors, model changes, deletion of data and ending the engagement. A vendor selling to licensees should be ready for these requirements.

What is "heightened use" of a technological means, and who may provide such a service?

It is use of a technological tool that the licensee itself has classified as high risk and high materiality. Only a corporation and its licensees may provide such a service, not an individual licensee (section 19). Extra duties apply: board approvals (section 21), specific explainability (section 28) and reporting to the ISA (section 20).

Can a licensee exempt itself from liability when the tool came from an external vendor?

No. Under section 3 of the draft, using a technological tool does not reduce the licensee's responsibility, and section 34.2 says using an external vendor does not reduce it either. Section 31 also bars any liability disclaimer in the disclosures or the agreement that does not comply with the directive.

Must clients be told that the service uses artificial intelligence?

Yes, in most cases. Section 30 requires a clear, prominent and effective notice that the service includes an AI system. The notice is not required only where it is already clear to the client that AI is used. Section 31 also bans misleading statements about the tool's capabilities, known as "AI washing".

When does the directive take effect?

Not before it is set in final form. Under section 43 it takes effect 6 months after the ISA publishes a notice of the final text in Reshumot, the official gazette. As of 17 November 2026 the notice has not been published, so there is no effective date.

The above is general information only and is not legal advice. For advice on your situation, talk to us.

Sources

Want to know what applies to your company?

Talk to us