Israel Bar Association Bars AI Agents From Making Legal Decisions

Israel’s Bar Association has published ethics guidance that draws a hard line around AI agents in legal work: they may assist lawyers, but they cannot make substantive legal decisions, give independent advice or move a case forward without close human supervision and approval. The position paper leaves the human lawyer personally responsible for work done through AI.

The National Ethics Committee’s updated position paper responds to AI systems that can plan workflows, select tools and interact with external systems with some autonomy. Its central concern is not simply whether a lawyer uses AI, but whether the system is allowed to replace professional judgment at a consequential point in representation.

A line between assistance and representation

The guidance prohibits systems from deciding matters of representation, acting independently toward a client, court or third party, filing pleadings or binding documents, or advancing proceedings without sufficient oversight and approval. That restriction applies even when a product is described as an AI agent rather than a conventional chatbot.

The paper does not impose a blanket ban on automation. It allows technical, limited, documented and reversible actions when they follow a prior human decision, operate within clear parameters and remain subject to controls and oversight. The distinction makes the lawyer’s earlier decision—and the ability to review or undo an action—the gate for acceptable use.

What remains acceptable under the paper

• Technical tasks with a limited purpose and defined parameters.
• Actions taken after a human has made the relevant decision.
• Work that is documented, reversible and subject to human controls.

Client data is not made safe by a label

The paper tells lawyers not to place restricted client information into open or unsecured AI systems when they cannot be confident the data will be protected to the standard required by confidentiality and privacy duties. The restriction reaches beyond entire case files: partial facts, email drafts, medical information and distinctive commercial details can also expose a client’s matter.

Removing a name is not necessarily enough. The committee warns that cross-referencing can re-identify a person or connect information back to a client’s case. And using a closed platform does not end the inquiry: lawyers must still assess storage, access permissions, logs, encryption, automation and links to other systems.

Consent and verification become operational work

Client disclosure is not required for every routine or ancillary use of technology. But transparency generally becomes necessary when AI materially changes how representation is carried out, is expected to materially influence professional decisions, affects fees or expenses, or is used in circumstances involving restricted information on an open platform. Where anonymization does not sufficiently prevent identification or linkage, the paper requires prior, explicit and informed consent.

Firms using outside providers must reasonably check their contractual, operational and technical protections for client data, including limits on sharing and training use. The paper also says access controls matter inside a firm, where an AI system must not expose a case to employees or teams who should not see it.

The committee’s warning on hallucinations gives the rules their sharpest practical edge. It says fabricated rulings, citations and legal sources have already appeared in documents submitted to courts, and blind reliance can lead to sanctions, reputational damage and disciplinary measures. Compliance with the paper can indicate ethical and reasonable conduct, but it does not remove the duty to exercise judgment in the circumstances of each case; noncompliance may be prima facie evidence of an ethical failure.