Client Intake Without UPL: A 2026 Compliance Playbook for Firms

For a while the open question around legal AI was whether to touch it at all. That question is settled. In 2024 the American Bar Association issued its first formal ethics opinion on generative AI, and the message was not "avoid it." The message was that your existing duties, competence, confidentiality, communication, supervision, candor, reasonable fees, already cover it. By 2026 the relevant question for a firm’s website is no longer whether the assistant is allowed. It is whether you can show you are supervising it.

This guide is the companion to the threat side of that story. The threat piece walks through what goes wrong when an unguarded chatbot sits on a law firm site. This one is the standard: what a compliant deployment looks like, and the specific lines the system has to hold.

The Duty Did Not Change. The Tool Did.

Formal Opinion 512 is worth reading the way it was written, as a checklist of duties you already owe. A lawyer has to understand the tool well enough to use it competently. A lawyer has to protect client and prospective-client information from going somewhere it should not. And a lawyer remains responsible for the work product and conduct that flows from the tool, the same as for any nonlawyer assistance under Model Rule 5.3.

That last point is the one firms underestimate. The rule does not let you treat the vendor as the responsible party. If the assistant speaks to a visitor under your firm’s name, you are the supervising lawyer, and "the software said it" is not a defense the bar recognizes.

The License Is Not in the Software

The deeper line is older than any AI rule. Applying law to a person’s facts and telling them what to do is the practice of law, reserved to licensed attorneys, and Model Rule 5.5 makes a lawyer answerable for unauthorized practice that runs through them.

So the dangerous answers are the ones that feel like ordinary help. A visitor describes a dispute and asks "do I have a case?" The honest, compliant response is that the question needs a lawyer, not a paragraph of analysis. General legal information is fine, your practice areas, how a consultation works, what a statute of limitations is in the abstract. The moment the assistant applies that law to the visitor’s situation, it has done the licensed thing, and your name is on it.

The 2026 Compliance Standard, Line by Line

A compliant firm assistant is defined by what it is built to refuse. Treat the list below as the floor.

• No analysis of a visitor’s facts. "Based on what you’ve told me" is the phrase that should never appear; the assistant gathers, it does not opine.
• It declines to predict outcomes or estimate the value of a matter. Whether a case is strong, what it might be worth, how long it will take, all of that routes to an attorney.
• Confidential intake stays out of unvetted systems. A prospective client’s story is protected under Model Rule 1.6, so the deployment cannot pipe it through a vendor your firm never audited.
• Conflicts come first. A tool that takes detailed matter information from anyone, including an opposing party, before a conflicts check is a problem; the assistant collects only what is safe and hands the rest to intake.
• Claims about results are off-limits. Model Rule 7.1 bars false or misleading statements about a lawyer’s services, which means no hinting at how a case will go.
• And it keeps a record. Every exchange is logged and attributable, which is what supervision looks like when someone asks you to prove it.

The pattern underneath the list is simple. The assistant answers what requires no license and routes every judgment that does to a person. That division is the entire standard.

Supervision Is the Part That Gets Skipped

Opinion 512 does not only care about the words the visitor sees. It cares whether you can describe and defend the system producing them. Meeting the standard means knowing what the tool can and cannot say, where the conversation data goes, what the vendor is contractually accountable for, and how you would show a bar investigator the boundary that keeps the assistant out of legal advice. A governance story that amounts to a vendor’s brochure is not supervision. It is the gap a complaint walks through.

Why an Instruction Cannot Meet the Standard

The usual shortcut is to write the rules into the assistant’s prompt. Tell it never to give legal advice, never to analyze a visitor’s facts, and call the boundary set.

It is not set, because of how the technology actually behaves. A language model obeys an instruction when the incoming question resembles the wording it was warned about. Reword the question and the guard slips. You tell it never to give legal advice. The visitor never asks for "legal advice." They write, "is a verbal agreement enforceable in my state?" The model reads a general question and answers with a tidy analysis that applies law to the asker. The instruction was loaded the whole time. It simply did not recognize the sentence that crossed the line.

That is the difference between an instruction and a standard. An instruction asks the model to behave; it does not stop the model from speaking. A real boundary is built into the system and decides what the assistant may say before it answers, so an analysis of a visitor’s matter never reaches them no matter how the question is framed. "Will not" is a suggestion. "Cannot" is an architecture.

What a Compliant Deployment Looks Like

Meeting the 2026 standard does not mean taking the assistant off your site. It means running one that was built to hold the line your license depends on, and that produces the record a regulator will eventually ask to see.

Fred is built that way. It answers from your firm’s own content, books the consultation, captures the lead, and routes anything that asks for analysis, a prediction, or a recommendation to a licensed attorney. It runs more than 50 industry guardrail packs, and the legal pack is built around unauthorized practice, confidential intake, and claims about results. Fred does not weigh a visitor’s case or guess at their odds. It cannot. It answers what it should, logs every exchange, and hands the lawyering to the people licensed to do it.

That is the difference between hoping the assistant stays in its lane and being able to show the bar why it cannot leave it.

Frequently asked questions