On 14 August 2025, Amendment 13 to the Privacy Protection Law came into force and changed one rule that reaches you personally: when your company's AI system mishandles personal data, the exposure is no longer the company's alone. It is yours, as an officer. In more than twenty years of running companies I cannot remember another regulatory change that exposes so many CEOs while so few have actually read it.
What Exactly Took Effect in August 2025
Three things. Any one of them would justify a management meeting on its own.
First, fines of up to NIS 3 million, which the Privacy Protection Authority can impose directly through an administrative track. No lawsuit, no court.
Second, personal liability for officers. The law requires directors and CEOs to actively supervise the organization's compliance with it. "I did not know what the system was doing" has stopped being an answer.
Third, a duty to appoint a Data Protection Officer (DPO), which now applies to a large share of organizations in the economy.
So why is this an article about AI rather than about privacy? Because you do not need a dedicated AI law to be exposed. The chatbot answering your customers collects personal data. The model summarizing call-center conversations processes personal data. The employee who pastes a customer table into a free tool has transferred personal data to a third party. Every one of those already sits inside the Privacy Protection Law, in its new and sharper form.
If You Are a Supervised Entity, the Bar Is Higher Still
In December 2025, Israeli financial regulation set out an AI governance obligation for supervised entities. Two of its principles are worth reading twice. The first: the supervised entity carries one hundred percent of the responsibility for what the AI system produces, even when an outside vendor built the model. "The vendor got it wrong" is not a defense. The second: a duty to tell the customer that AI is in use at all.
Boards abroad are already staring at the same picture. An analysis of S&P 500 company filings found that roughly 83% now define AI as a formal risk factor, against only 12% in 2023. On the other side of the ledger, only about 2.7% of directors hold real AI expertise. That gap, between the risk companies declare and their ability to supervise it, is exactly the exposure the new law lands on officers. The response abroad has already begun: the share of companies with a dedicated board committee for AI jumped from 6% to 20% inside a single year. In most Israeli boardrooms the subject has not reached the agenda yet.
The Chatbot Fails Differently From What You Are Used To
An AI system does not break the way ordinary software breaks. It does not throw an error screen. It answers wrong things with complete confidence. It promises a customer a discount that does not exist. It lets one customer's details slip into a conversation with another. It invents a clause in your own policy. Small wonder that industry surveys heading into 2026 rank privacy and governance as the number one barrier to AI adoption in organizations, at 61.3% of respondents, ahead of model accuracy and ahead of integration difficulty.
Underneath that sits a quieter layer: your own employees. Israel leads the world in personal adoption of AI tools, which means that even if your organization never formally brought AI in, it is already inside, through every employee's browser, with no controls, and with nobody in management knowing what data has already left the building.
From running systems like these against live customers, I have learned that the protection is in the architecture, not in a vendor's promises. In one system I built, the filtering layer keeps working even when the AI model itself is unavailable: hard-coded rules go on blocking whatever must not get through. The model is a talented employee, but it signs off on nothing. The signature belongs to the controls.
Three Documents That Remove Most of the Exposure
The good news is that most of the risk comes off the table without a year-long project. It takes three defined deliverables.
An AI governance document. A handful of pages setting out which AI uses are approved in the organization, which data may be fed into which tools, who approves a new use, and who owns each system. Once the document exists, supervision exists; once supervision is documented, the officer has done what the law asks.
Controls in practice. Logs for every interaction between an AI system and a customer, permissions that govern who reaches what, a blocking layer over prohibited content, and a periodic check of the system's answers against reality. A control nobody documented may as well not exist.
Documentation of decisions. When the inquiry from the Authority arrives, the first question is not "did you get it wrong" but "what did you do to prevent it". Minutes of a management discussion, a risk assessment before a system goes live, a record of fixes: those are the assets that decide whether an incident ends in a remark or in a fine.
All three can be stood up in weeks, not years. I have done it, including in environments carrying heavy regulatory sensitivity, and the difference between an exposed organization and a protected one came down mostly to a management decision to take the subject seriously before the regulator calls.
Three steps for the coming week. Ask for a map of every AI tool already in use in the organization, including the ones employees adopted on their own. Make sure there is a single address responsible for privacy protection, even before a formal DPO appointment. Put the subject on the agenda of the next board meeting, with minutes. The three cost nothing, and they change your legal starting point from end to end.
If you want to see what a governance document like that looks like in practice, or a live demo of working controls on a real system, send me a message. I would be glad to show you.
