· 4 min read
AI In The Legal Field: Where The Lines Of Supervision Should Be Drawn
The future should be hybrid. Legal services should be AI-powered and attorney-supervised. This model is commercially durable because it combines AI’s speed with the accountability, contextual knowledge and licensed judgment that only a human lawyer can provide.
From the ZheroTax article collection
Imagine a small business owner who’s just starting out. They need to set up an LLC, register intellectual property, create contracts for clients and take care of other common legal matters. They don’t have the budget to hire a lawyer, so they turn to an AI tool for guidance. The LLM stresses that it’s not a lawyer but gives the small business owner advice. The small business owner applies that advice. They launch their business, and things are going well—until a dispute with a client. At the center of the dispute is a clause in the contract. The small business owner copied and pasted that clause from the LLM’s output, not realizing the legal can of worms it could open.
I fear these experiences may become increasingly common. Many people, including business owners, are turning to AI for legal help. As are many lawyers. That raises the question: Where do the lines of supervision get drawn?
AI can now handle its share of legal work. The core tension is balancing AI’s power with the need for responsible AI in legal services. Defining responsible AI requires identifying who is liable for errors AI tools make when used for legal advice. In the U.S., the legal system has traditionally tied responsibility for providing legal services to licensed human professionals. However, that’s starting to change through both legislation and the courts.
On the legislation side, I’ve previously written that New York State Bill S7263, which is still under consideration, attempts to answer that by stating that if it would be illegal for an unlicensed person to provide legal advice, then it would be illegal for an AI chatbot to do so, too (and the company running the chatbot would be liable). That’s just one example. According to JD Supra, California’s SB 574 became law on September 30, 2026. JD Supra notes that the new law states, among other provisions, that lawyers must “disclose the use of generative artificial intelligence in all documents submitted to the court.”
Courts are also holding lawyers accountable for AI-assisted work. NPR reported that 2025 “saw a rapid increase in court sanctions against attorneys for filing briefs containing errors generated by artificial intelligence tools.”
Why Purpose-Built Legal AI Is Not A Safe Harbor, Either Purpose-built legal AI tools might seem like safe options, but they aren’t. They too have inherent limitations, such as the possibility of hallucinations and context drift.
A Stanford University study published in 2025 demonstrates the extent of this problem. The study’s authors examined certain AI legal research tools and found that “each hallucinate between 17% and 33% of the time.”
Attorney responsibility is non-delegable. Lawyers are ethically responsible for all AI outputs they use or share. Without strict verification—a “trust, but verify” approach—they can jeopardize client outcomes and harm their careers.
That’s not to say that there is no room for AI in the legal field. There’s no going back. The future should be hybrid. Legal services should be AI-powered and attorney-supervised. This model is commercially durable because it combines AI’s speed with the accountability, contextual knowledge and licensed judgment that only a human lawyer can provide.
As a practicing attorney and founder of a legal technology company that combines AI agents with attorney supervision, I’ve seen both sides. AI legal platforms need grounding, confidence-scoring and guardrails in their design, with attorney review and oversight embedded throughout the user experience. That’s the approach my company has taken.
Grounding means giving AI the relevant context it needs to produce more accurate outputs. In our case, that involves parsing legal agreements clause by clause and enriching each clause with plain-English explanations, risk scores and preapproved alternatives. This gives AI a narrower set of instructions and context to work from, helping reduce truncation, summarization and drift.
Confidence scoring can then gate answers, preventing AI tools from responding when they’re not sufficiently confident. It mirrors the rule many legal interns are given: “Never answer a question that you don’t know the answer to.”
These capabilities also support guardrails around how AI handles legal work. For example, preapproved alternative clauses let clients use AI to automate certain redline changes without having to return to an attorney for every revision. The alternatives are developed from the legal expertise built into the platform, giving the AI defined options to work with rather than requiring it to generate a response from scratch.
Attorneys using any AI tool in their work should follow best practices. First, they should remain responsible for all AI-generated outputs and supervise whatever AI tools produce. They should never take an AI tool’s output at face value. Their modus operandi should be “trust, but verify.” They should also prioritize becoming competent in AI, understanding how the tools work and the privacy implications. Finally, they should keep client confidentiality top of mind. Safeguarding confidential client information means being mindful of what information they enter into AI tools.
Ultimately, the legal industry is being disrupted. The durable model for the future should be an industry that is AI-powered and attorney-supervised, not one or the other.