Gurpreet S. Bal has spent 16 years watching where legal judgment actually lives — and where it doesn't. His view is unambiguous: "If your lawyer tells you they use end-to-end agentic AI on your deal, find a different lawyer. The judgment is what you're paying for." The question of whether an AI system is practicing law is no longer theoretical. It is an active regulatory and ethical problem that the profession is failing to address at the speed the technology demands.
Bal advises technology companies, investors, and acquirers on some of the most complex transactions in the market — and sees firsthand how AI is reshaping what legal work looks like, and where the lines should be drawn.
Unauthorized practice of law (UPL) statutes exist in every U.S. state. They prohibit non-lawyers from providing legal advice — but almost none of them define "legal advice" with precision. Courts have generally distinguished between legal information (permissible for anyone to provide) and legal advice (reserved for licensed attorneys). The line, in practice, turns on whether the provider is applying legal principles to the specific facts of a particular person's situation and recommending a course of action. For decades, this line was applied to humans. LegalZoom and its competitors pushed the boundary by automating document assembly, and the courts largely tolerated it on the theory that the software was providing forms, not advice. That theory is now under significant pressure from large language models, which do not produce forms — they produce tailored analysis that applies legal principles to specific facts with a fluency that is genuinely difficult to distinguish from licensed advice.
The document assembly services that survived early UPL challenges did so partly because their outputs were templates — structured, predictable, and clearly bounded. An LLM answering the question "should I sign this non-compete?" is doing something categorically different. It is reasoning about the specific language of a specific agreement, the jurisdiction, the client's circumstances, and the applicable case law — and producing a recommendation. Whether that constitutes the practice of law depends on the jurisdiction and the framing, but the honest answer is: it often looks a lot like it. Bar associations are beginning to catch up. Several state bars have issued guidance warning that AI-generated legal analysis provided directly to consumers without lawyer supervision may constitute UPL. The companies building consumer legal AI products are navigating this carefully — but the frontier is genuinely unsettled, and users of these products often have no idea they are in legally ambiguous territory.
The UPL concern is most acute with agentic AI systems — tools that do not merely answer questions but take sequences of actions toward a goal. An AI agent tasked with "handling my contract dispute" or "closing my seed round" is not providing information. It is making a series of decisions: what to file, when to respond, what terms to accept, what leverage to apply. Each of those decisions is a legal judgment. The chain of autonomous actions is precisely what UPL doctrine targets, and it is exactly what the current generation of goal-directed legal AI agents is designed to perform. A human lawyer remains nominally in the loop in most current deployments, but "nominally" is doing real work in that sentence. If the lawyer is not actually reviewing and exercising judgment over each consequential step, the supervision is a formality — and the AI is making the calls.
Goal-based agents present a distinctive risk that point-in-time tools do not: path risk. When you instruct an AI agent to achieve an outcome, it selects the path. The lawyer who sets the goal and approves the final product may never see — and may not have meaningfully supervised — the intermediate steps. Those steps can include concessions made in correspondence, representations about client authority, characterizations of the transaction, or procedural choices that have substantive consequences. If a client suffers harm from one of those unsupervised intermediate steps, the question of who is responsible is genuinely difficult. The lawyer may argue the AI acted outside its instructions. The client will argue the lawyer was responsible for the AI. Neither position is clean. This is not a hypothetical risk. It is a structural feature of how goal-directed agents work — and it is a reason why end-to-end agentic AI on active client matters is not, today, a responsible deployment.
Clients interacting directly with consumer legal AI products should understand that the advice they receive, however fluent and detailed, may not carry the accountability or protection of advice from a licensed attorney. There is no malpractice insurance, no bar discipline, and no fiduciary duty attached to an AI output. For low-stakes, high-volume legal tasks — understanding a lease, reviewing a standard NDA, checking a contractor agreement — the risk-benefit calculus may favor convenience. For anything with meaningful stakes, the absence of an accountable human lawyer is a serious gap. Clients working with law firms that use AI should ask whether a licensed attorney has reviewed and exercised judgment over every consequential output. If the answer is effectively no, that is the answer.
Gurpreet S. Bal is a corporate partner with 16 years advising on private equity, merger transactions, and public offerings for companies and investors at three of the world's top law firms. He has represented clients in hundreds of transactions with aggregate deal value exceeding $60 billion across AI, semiconductors, fintech, and emerging technology. For more information and to get in touch, visit gurpreetbal.com.