In Israel, putting customer data into an AI tool counts as processing personal data under the Privacy Protection Law. If a vendor gets access to a database you hold, the Data Security Regulations require you to assess the risk before you sign and to fix what the vendor may do in the agreement. So decide which customer data the AI actually needs before you connect anything, and get the terms in writing.
This is general information from primary sources, not legal advice. Where a lawyer has to decide, the post says so.
Does the Privacy Protection Law cover an AI tool?
The Privacy Protection Authority (PPA) says so. Its draft guidance on applying the Privacy Protection Law to AI systems (the document is dated April 2025) states that the law applies to an AI model that stores or processes personal data, both in training and in use, and that a legal basis must be ensured for any processing of personal data that infringes privacy, at every stage of the system's life cycle. It is a draft: we found no final version, so treat it as the PPA's stated position, not as a settled rule.
The law itself changed on 14 August 2025, when Amendment 13 entered into force. The PPA confirms that date in its opinion of 14 July 2026. The PPA's guide to Amendment 13 (cover dated May 2025, before the amendment took effect) says that processing includes any action on personal data, such as storing it, viewing it, transferring it or giving access to it. Feeding a customer record into a tool is therefore processing, even if the tool only reads it.
The same guide widens the category of data with special sensitivity. Beyond health, genetic and biometric data, it includes data about a person's salary and financial activity, and data under a legal duty of confidentiality. A small business that handles invoices, payroll or client files may hold more sensitive data than it assumes.
What must the contract with an AI vendor say?
When you give an outside party access to a database, regulation 15 of the Data Security Regulations 5777-2017 applies. In plain terms it requires you to:
- Assess the risk before signing. Examine the information-security risks of the engagement before you engage the vendor.
- Put specific points in the agreement. The permitted data and purposes, the systems it may access, the type of processing, the duration, how the data is returned and destroyed when the engagement ends (with a report), how the vendor meets its security duties, confidentiality undertakings from its staff, the rules for any sub-contractor, and a duty for the vendor to report to you at least once a year on its compliance, and to notify you of a security incident.
- Supervise. Apply controls and oversight over the vendor's compliance, to the extent the risks require.
The regulation does not set a fixed audit schedule for you; the one fixed interval in its text is the vendor's annual report. The PPA publishes a practical guide to regulation 15 with optional questionnaires you can use before signing and during the relationship.
Which security level applies to your database, and whether a specific vendor contract satisfies these points, are questions for a lawyer.
What if the AI vendor processes data abroad?
Most AI services run outside Israel. Sending personal data to a database abroad has its own rules, the Privacy Protection (Transfer of Data Abroad) Regulations 5761-2001. The PPA has published two final opinions on them: one on regulation 3 (13 October 2024) and one on regulation 2(4) (13 April 2026). Two points a business owner can use. Regulation 3 requires the data owner to secure a written undertaking from the recipient abroad to take sufficient measures to protect privacy and not to pass the data on. The PPA's 13 October 2024 opinion allows onward transfer with the data owner's written consent, and says the undertaking need not match the full conditions of regulation 2(4). And the PPA says that when the transfer also involves outsourcing or giving an outside party access to your database, regulation 15 applies too.
The opinions do not mention cloud services or AI vendors specifically, and they do not say that a particular vendor's data processing terms are enough. That judgment is for counsel.
What do the AI vendors say about retention and training?
The vendors' own terms are the second half of your due diligence. For their API (business) products:
- Anthropic says in its privacy centre (page updated 1 July 2026) that API inputs and outputs are automatically deleted from its backend within 30 days, except when you use a service with longer retention under your control (for example the Files API), when you and Anthropic agreed otherwise (for example a zero data retention agreement), when needed to enforce the usage policy, or when the law requires it. Its commercial terms (effective 17 June 2025) say it may not train models on customer content from its services, and incorporate a data processing addendum (effective 24 February 2025). Whether that addendum satisfies Israeli requirements is for counsel.
- OpenAI says in its data controls guide (no date shown) that data sent to the API is not used to train its models unless you opt in, that abuse-monitoring logs are retained for up to 30 days by default, and that eligible customers can request zero data retention controls, which need OpenAI's prior approval and apply to eligible endpoints.
These are business-product terms. A free consumer chat app is governed by different terms, and staff pasting customer data into one is a separate risk worth a rule of its own.
What to do this week
- List which customer data each AI tool receives, and cut anything the task does not need. Check whether any of it is health, financial, payroll or confidential data.
- For each vendor, get the retention period, whether your data trains their models, where it is processed and whether a data processing addendum exists, in writing. Compare it with the regulation 15 points above.
- Tell people when they are talking to an AI system (the PPA draft, section 6.3.1, treats this as part of informed consent where it can materially affect their decision), and put a lawyer on the legal basis for the processing before you go live.
Sources
All pages below were read on 2026-10-07. On 2026-10-28 we checked again whether the PPA AI guidance had a final version, and found none.
- Draft guidance on applying the Privacy Protection Law to AI systems, Privacy Protection Authority, April 2025 draft. No final version found as of 2026-10-28.
- Opinion on the DPO duty under Amendment 13 (confirms entry into force on 14 August 2025), Privacy Protection Authority, 14 July 2026.
- Guide to Amendment 13, Privacy Protection Authority.
- Privacy Protection Regulations (Data Security) 5777-2017, regulation 15, Nevo.
- Practical guide to regulation 15, Privacy Protection Authority.
- Opinion on regulation 3, Transfer of Data Abroad Regulations, Privacy Protection Authority, 13 October 2024.
- Opinion on regulation 2(4), Transfer of Data Abroad Regulations, Privacy Protection Authority, 13 April 2026.
- How long do you store my organization's data, Anthropic Privacy Center, updated 1 July 2026.
- Commercial Terms, Anthropic, effective 17 June 2025.
- Data Processing Addendum, Anthropic, effective 24 February 2025.
- Data controls in the OpenAI Platform, OpenAI, no date shown.