For 92% of their civil legal problems, low-income Americans get no real help.
We are writing down what an AI may now do about that — and what stays with a licensed human.

Five things stay with the lawyer, in every design we will ever propose.

Appearing before a court. Giving final legal advice. Signing pleadings. Setting and taking legal fees. Agreeing to settle. That is the reserved list, and it is short on purpose: everything outside it — research, drafting, forms, timelines, translating legalese into plain English — is preparatory work a licensed lawyer may supervise a machine in doing, the way lawyers have always supervised the people and tools that help them.

This is not our invention. Utah's own reform committee, redesigning what comes after its sandbox, is studying the same shape: the UK's "reserved activities" model, which — in the committee's words — flips the unauthorized-practice question on its head by defining which activities are reserved for fully licensed lawyers.

Courts have answered this question before — for people.

In 1961 a federal appeals court held that an accountant working at a lawyer's direction does not break the privilege, because he is the lawyer's agent, not a stranger in the room (United States v. Kovel). Minnesota's privilege statute already covers the lawyer's employees and interpreters. In 2013 the Minnesota Supreme Court placed supervised law students inside the privilege by rule, and in 2020 it did the same for legal paraprofessionals.

In February 2026 the first federal court to face the AI version left the door open in exact words: had counsel directed the defendant's use of the AI, it "might arguably be said to have functioned … as a lawyer's agent within the protection of the attorney-client privilege." The defendant had directed it himself, with no lawyer anywhere — so he lost. The condition that failed is the condition our whole design begins with.

The machine is the lawyer's instrument, and the record proves it.

Before any machine writes a word that is about your situation, a licensed lawyer is engaged — yours, named, responsible. The AI then works the way the bar's own guidance says a lawyer's tools must work: under supervision, with the lawyer reviewing and adopting what it produces before you rely on it. Every direction is logged — which lawyer, which task, which matter, when — because that log is what a court would read. The words themselves are not in the log; they stay in the lawyer's file, where client words belong.

And the system does not learn from your matter. Not as a promise in a policy — as an architecture: nothing identifiable leaves the matter, nothing enters the model. A promise can be broken quietly. An architecture can be inspected.

Minnesota has twice let a certified non-lawyer inside the attorney-client privilege.

Law students in 2013. Legal paraprofessionals in 2020 — as a pilot that measured itself and then became permanent. Both times the conditions were the same: certified by the Board of Law Examiners, supervised by a named lawyer who carries personal professional responsibility, identified to the client and accepted by them, bound to confidentiality. Both times the sky stayed up. And Minnesota's bar association has already recommended that its supreme court consider a sandbox for access-to-justice AI.

The list, in one paragraph.

The 25 Minnesota Rules of Professional Conduct and 4 statutes we read, from revisor.mn.gov. The Minnesota Supreme Court's orders — the student-observation amendment (file ADM09-8002, order of August 6, 2013, retrieved from the State Law Library's archive) and the paraprofessional pilot (ADM19-8002). The bar's national guidance on lawyers using generative AI (ABA Formal Opinion 512, 2024). The measurement of the gap (LSC, The Justice Gap, 2022). Utah's published sandbox data — roughly one complaint per 5,869 services through April 2025 — from its Office of Legal Services Innovation. One order in the 2013 file, from March, we have not yet retrieved; until we have read it, nothing on this site cites it.