The unauthorized practice of law doctrine prohibits non-lawyers from practicing law, but every jurisdiction defines "practicing law" differently, and none of the definitions translated cleanly into the digital age. The general framework distinguishes between legal information (permissible for anyone to provide) and legal advice (reserved for licensed attorneys). Legal advice typically means applying law to specific facts in a way that creates or implies a professional relationship, influences a legal decision, or gives rise to reliance. Gurpreet Bal observes that this distinction was already strained by document assembly tools, self-help legal forms, and legal research platforms before large language models arrived. AI makes it more acute, because a sufficiently capable model can analyze a specific situation, identify applicable statutes and case law, and recommend a course of action in a way that is functionally indistinguishable from what a lawyer does — except no licensed attorney was involved.
The LegalZoom and Rocket Lawyer cases, litigated across multiple states in the 2010s, established a rough precedent: automated document assembly that uses standardized forms and does not involve human judgment applied to specific facts generally does not constitute the practice of law. Those companies survived UPL challenges primarily by arguing that their tools did not provide advice — they helped users complete forms that users chose based on their own understanding of their situation. Gurpreet S. Bal notes that this framing is under severe pressure from modern AI systems. An LLM that takes a user's description of a dispute and generates a demand letter tailored to the specific facts and legal theory is doing something qualitatively different from a fill-in-the-blank form — it is applying law to facts. Whether that triggers UPL liability depends on jurisdiction and judicial temperament, but the argument that it does not constitute legal advice is significantly weaker for an AI that reasons about a specific situation than it was for a static template system.
The frontier of AI legal tools is not document generation — it is agentic AI that takes multi-step autonomous actions: drafting and filing court documents, sending demand letters on behalf of a party, conducting negotiations by email, submitting regulatory filings. Each of these actions is one that a licensed attorney would traditionally perform. Gurpreet Bal considers agentic legal tools the highest-risk category in the AI-law space for precisely this reason: a human-in-the-loop model, where an attorney reviews AI-generated output before it is sent or filed, provides at least a defensible argument that the attorney is the practicing lawyer and the AI is a sophisticated tool. A fully autonomous agent that files a motion, sends a settlement offer, or submits a regulatory response without contemporaneous attorney review removes that argument entirely. The liability question when an autonomous agent takes an unauthorized legal action — and causes harm — has not been resolved by any court, and the company that deploys that agent is likely to find out the answer the hard way.
For corporate legal departments considering AI tools that touch legal workflows, Gurpreet S. Bal recommends a framework that distinguishes between three zones. The first is clearly safe: AI tools that assist attorneys in doing legal work more efficiently — research, first-pass document review, clause identification, summarization — where a licensed attorney reviews and approves all output before it is acted upon. The second requires careful design: tools that generate legal documents or analysis for non-lawyer employees, such as contract review tools used by procurement teams or compliance tools used by business units, which are permissible if properly scoped, limited to information not advice, and accompanied by clear disclosures directing users to consult counsel for consequential decisions. The third is a zone that requires legal counsel before deployment: any agentic tool that takes actions with legal effect — sending communications, making commitments, filing documents — on behalf of the company or its employees without contemporaneous attorney review. Deploying tools in that third zone without a clear analysis of UPL exposure and a liability framework is, in Gurpreet Bal's view, one of the most underappreciated legal risks facing technology companies in 2026.
There is a version of this conversation that requires no regulatory analysis. If your lawyer tells you they use end-to-end agentic AI — meaning AI handles the transaction workflow without a human attorney reviewing the substance before it goes out — on a deal significant enough to justify hiring a major law firm, you should find a different lawyer. Not because AI tools are bad. Because the entire value proposition of experienced legal counsel is judgment: knowing which issues to fight, which to concede, what the other side actually cares about, and what the words in a document will mean when something goes wrong two years after closing. No agentic AI system in the current state of the technology reliably provides that. The tools are genuinely useful as accelerants for experienced lawyers. They are not substitutes for experienced lawyers. A firm that has removed the human from the loop on consequential decisions has not found an efficiency — it has eliminated the thing you were paying for. Gurpreet S. Bal's view is direct: the appropriate role for AI in major transactions right now is to make good lawyers faster, not to replace the judgment that good lawyers bring. Clients who accept anything less are paying law firm rates for something they could get from a software subscription.
The deeper concern — the one that gets less attention — is not whether the AI reaches the goal. It is how it gets there. An agentic AI system given a goal-based instruction ("get me a better deal," "win this negotiation," "close this transaction") has no inherent constraint on the path it takes toward that goal. Guardrails help, but guardrails are not the same as judgment and they are not infallible. Without a human approving each material step, you have no real visibility into what representations were made, what was conceded, what communications went out, or what the AI concluded was an acceptable tradeoff on your behalf. In a legal context, those steps have consequences — binding commitments, waived rights, statements that become admissions. An AI agent optimizing toward a goal can take actions that individually appear reasonable and collectively produce an outcome no lawyer and no client would have approved if asked at each juncture. The goal-based framing is precisely what makes fully autonomous legal agents dangerous: it outsources not just execution but the entire sequence of judgment calls that experienced counsel is actually being paid to make. Gurpreet S. Bal advises clients that specific human approval at each material step is not inefficiency — it is the safeguard. Removing it in the name of speed is a tradeoff that should terrify anyone who has seen a deal go wrong because of a single sentence that nobody caught.
Gurpreet S. Bal is a Partner at Foley and Lardner LLP in Silicon Valley, where he advises technology companies, founders, and investors on corporate transactions and the evolving intersection of law and artificial intelligence. He has represented clients in hundreds of transactions with aggregate deal value exceeding $60 billion across AI, semiconductors, fintech, and emerging technology.