An AI agent that answers customers, approves refunds, sends quotes, or fixes code and pushes it to production is no longer an experiment. It’s part of how many Israeli startups actually operate. The question most founders haven’t asked yet: what happens when it gets something wrong? Who holds the outcome, legally, when the agent you deployed takes an action you never would have taken yourself?
Israel has no AI law, so where does liability come from?
The short answer: from the ordinary rules. In Israel, as in most of the world, there is currently no dedicated legislation determining who is liable when an AI agent causes harm. That means a claim against a company that deployed an AI agent will rest on the same legal tools that existed long before AI: the Torts Ordinance (negligence, negligent misrepresentation), contract law (breach, non-performance), and the law of agency (when an act by an "agent," in our case, the AI agent, binds the "principal," meaning the company). Even in the US, where the public debate is further along, there is no dedicated federal statute; a handful of states, California among them, have already enacted provisions that explicitly bar companies from arguing that "the AI acted autonomously" as a defense. Israel hasn’t seen a similar move yet, but that doesn’t mean there’s no liability, it means it’s assessed under old rules applied to a new reality.
Four ways an AI agent gets into trouble, and who’s left holding it
Not every agent mistake is legally identical. It’s worth distinguishing four scenarios, each with a different analysis:
- Faithfully executing a bad instruction. The agent did exactly what you asked, but the instruction itself was negligent or too broad (e.g., "auto-approve any refund request under ₪500," with no check for repeat requests from the same customer). Liability here is relatively straightforward, it’s ordinary negligence in process design, exactly as if a human employee had followed a bad procedure.
- Inference that goes beyond the instruction. The agent retains context and makes decisions it was never explicitly asked to make, for example offering an unapproved discount because it "sensed" the customer was about to churn. This raises the question of apparent authority toward a third party: under agency law, if a customer reasonably believed the agent had authority to act this way on the company’s behalf, the company may be bound by the act, even if it never explicitly approved it.
- An agent that "goes rogue" against its instructions. The agent ignores a constraint it was given in order to achieve a goal, a known phenomenon in autonomous models (reward hacking). Here the legal question returns to negligence in design and monitoring: did you have a way to detect and stop this behavior before it caused harm, and if not, why not.
- False information, or a "hallucination." The agent simply invents a fact, policy, or promise that doesn’t exist, and a customer relies on it. This is a classic case of negligent misrepresentation: legally, what the agent said is what the company said.
The precedent already exists: when a company paid for what its chatbot promised
This isn’t theoretical anymore. In February 2024, a Canadian tribunal (Moffatt v. Air Canada, 2024 BCCRT 149) ruled that the airline had to honor a discount its customer-service chatbot mistakenly promised, even though the discount contradicted the official policy published on its own website. The tribunal’s reasoning was simple: there’s no legal difference between information on a static web page and information delivered by that same company’s chatbot. If the chatbot promised it, the company is bound. There’s no equivalent Israeli precedent yet, but Israeli courts would likely adopt similar reasoning, especially under consumer protection law.
Six things a founder should lock down before the agent goes to work
Most of the legal risk here is avoided in advance, not fixed after the fact. Six things worth closing before an AI agent gets access to real customers:
- Write down what the agent can and can’t do. A short usage policy separating "safe" actions (answering questions, routing to a human) from actions with financial, legal, or irreversible consequences.
- Keep a human in the loop for every irreversible action. Refunds, data deletion, sending a binding document, all of these need human approval, at least until you have enough operating history to trust the agent.
- Test the agent under attack, not just for success. Before production, run deliberate edge cases: what happens when a customer tries to "talk the agent into" breaking policy? Can it recognize that attempt?
- Log every action. A record of every decision the agent made and why, so you can prove after the fact, to a customer or a regulator, exactly what happened.
- Check your contract with the underlying AI model vendor. Who bears the risk if the mistake originates in the base model (e.g., an LLM hallucination)? Is there indemnification? A liability cap that shifts the entire risk onto you?
- Update your customer terms of use accordingly. Terms clarifying the status of the agent’s responses help, but know their limit: they don’t protect you against someone who never saw them, and they don’t override protections consumer law treats as mandatory.
An AI agent isn’t "autonomous" in any legal sense. Someone bears responsibility for its actions, and it’s always whoever deployed it.
Bottom line: the agent’s "autonomy" is a marketing point, not a legal defense. To a court, the difference between an AI agent and a new employee who made a mistake is much smaller than it looks. The founders who come out ahead are the ones who build the controls before the mistake happens, not the ones who discover them in a statement of defense.
Frequently asked questions
If the agent acted in direct contradiction to its instructions, is the company still liable?
In most cases, yes. Arguing that "the agent acted autonomously" isn’t a recognized legal defense, any more than a company could disclaim an employee’s mistake because the employee "did it on their own." The real legal question is whether the company took reasonable precautions to prevent the outcome.
Do terms of use saying "the bot doesn’t bind us" protect us?
They help, but not absolutely. Against a customer who never saw or read the terms, and never against a third party, the protection is limited. And in any case, consumer protection law restricts overly broad exemption clauses toward consumers.
Is there a specific AI liability law in Israel?
Not as of today. Liability is assessed under the general frameworks, tort, contract, and agency law, exactly as before the AI era. That may change with future regulation, but that’s the current state.
What’s the first thing a startup should do this week?
Map every AI agent currently facing customers or making independent decisions, and flag which of its actions are irreversible (money, data, legal commitment). Those are exactly the points that need immediate human control.
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