AI Agents at Work: What Israel’s Privacy Authority Now Expects

Israel’s Privacy Protection Authority published guidance on AI agents. Here’s what it means for your organization, and which approvals you need before an agent gets access to your systems.

On July 9, 2026, Israel’s Privacy Protection Authority published recommendations for the public’s personal use of AI agents. The recommendations are written for private end users, not for processing data in an organizational or business setting, which is addressed in a separate document. Still, the underlying principles, minimizing access, minimizing permissions, control and documentation, are exactly what regulators expect to see from organizations running AI agents too. Worth knowing before an agent gets access to your systems.

What an AI agent actually is

A “regular” AI tool generates text or an image in response to a one-off request. An agent is different: it is built from three components that turn it into something that acts independently and continuously, not just responds.

  • A “brain”, a large language model that interprets instructions, analyzes information, and builds a sequence of actions to reach a goal.
  • A “memory”, the ability to retain preferences and context over time, and tailor future actions accordingly.
  • “Hands”, an actual connection to external tools and interfaces (calendar, email, booking sites, operational systems), so the agent doesn’t just suggest an action, it carries it out.

The Authority frames it this way: when an agent replaces tasks a person used to do, reading email, managing a calendar, making purchases, it effectively has to act as a kind of “digital confidant.” Someone expected to understand context, exercise judgment, and act according to the user’s wishes (as it understands them) without the user weighing in on every step. That’s exactly where the risk lives, an agent doesn’t necessarily have the human intuition to know what should and shouldn’t be disclosed.

Three risks the Authority flags

An agent connected to your information systems, calendars, email, customer databases, exposes you to three main risks, illustrated with the Authority’s own examples:

  • Gathering more knowledge than needed. An agent given access to an information system tends to “analyze” far more than the task actually requires. In the Authority’s example, an agent meant to sort only the last two years of correspondence ends up scanning 15 years of email, including drafts and backups, accumulating a comprehensive profile of habits, relationships, and financial or health status, without anyone giving informed consent to that.
  • Acting outside the original context. An agent with memory may use a conclusion it drew earlier in a way the user never anticipated. The Authority’s example: a health-management agent infers from diet and activity data that the user has a particular medical sensitivity, then, when booking a restaurant for a business meeting, proactively notes that a special dietary accommodation is needed, without the user intending to disclose that.
  • Autonomous action that damages data. Agents don’t just read information, they edit, update, delete, and transfer it. In another of the Authority’s examples, an agent meant to “free up space” in an inbox permanently deletes old documents, which later turn out to include a lease agreement or a prescription needed for an insurance claim.

Five control points worth implementing

In practice, the recommendations translate into five working rules that anyone running, or considering running, an AI agent, organization or individual, should apply:

  1. Minimize data access, the agent gets access only to what the specific task actually requires (e.g., a dedicated mailbox, not the entire historical inbox), not a blanket connection to a whole database. Avoid “training” the agent on the data or letting it retain data long-term in memory, prefer temporary, controlled access, and delete working files once they’re no longer needed.
  2. Minimize outward exposure, make sure the agent’s external interfaces (booking a restaurant, arranging travel, etc.) don’t disclose more to third parties than necessary, e.g., meeting attendees’ names or a trip’s purpose, without explicit approval.
  3. Minimize permissions, the default should be view/read-only; edit, delete, send, or transfer permissions are granted only where clearly necessary, and after approval.
  4. Control and approval for every action, don’t let the agent perform an impactful action without explicit human approval, and consider requiring it to lay out its planned sequence of actions in advance and log every action it takes, so it can be audited after the fact.
  5. Back up in advance, before granting an agent broad permissions in any system, make sure the relevant data is backed up so it can be restored if the agent gets something wrong.

The step most organizations skip: prior approval

The recommendations themselves are written for personal use, but processing data in an organizational or business setting, belonging to the organization, employees, customers, or suppliers, is governed by a separate, stricter framework. So every connection of an AI tool or agent to an organization’s information systems should require prior approval from the relevant functions: IT, the person responsible for information security, and the Data Protection Officer (DPO). That holds even for a familiar, widely used tool, a use that looks one-off or minor, or data that seems non-sensitive, because that’s exactly where most incidents happen.

An AI agent doesn’t “ask permission” on its own. Someone in the organization needs to grant it in advance, and define its exact boundaries.

Bottom line: deploying AI agents is a powerful tool, but it’s also a material change to how your organization governs information. An organization that maps in advance which agents are running, which systems they’re connected to, and who approved that, walks into the Authority’s next piece of guidance already prepared, instead of fixing things retroactively under pressure.

Frequently asked questions

The recommendations are written for private end users, are they even relevant to my organization?

The recommendations themselves address personal use only; organizational and business data processing is governed separately. But the principles, minimizing access, minimizing permissions, approving every action, documentation and backups, are exactly what the Authority and other regulators expect from organizations deploying AI agents too.

What’s the practical difference between a regular AI tool and an AI agent, risk-wise?

A regular AI tool responds to a one-off request and produces an output. An AI agent retains context over time and is actually connected to your systems, so it can act, not just suggest, repeatedly, without human involvement at every step.

Which approvals are needed before connecting an AI agent to organizational systems?

Prior approval is recommended from three functions: IT, the person responsible for information security, and the Data Protection Officer (DPO), even for a common tool or a use that looks minor.

What’s the first step an organization should take this week?

Map which AI agents are already running in the organization (including ones teams rolled out on their own), which systems and data they’re connected to, and who approved that. This mapping exposes most of the risk in under an hour.

The above is general information only and does not constitute legal advice. For guidance tailored to your specific circumstances — talk to us.

Sources

Want to know what applies to your company?

Talk to us

Meet your legal department.

Leave your details and Tal will get back to you within one business day. We map what you need, and you leave knowing exactly what it costs.

  • No commitment, no hourly billing
  • A fixed price before any work starts
  • Every output signed off by a lawyer

Prefer to pick a time yourself? Book a call in the calendar

We use these details only to get back to you. Privacy Policy

Thanks, we got it.

Tal will get back to you within one business day.

Leave your details