Access to justice, serving the American people

A renter gets a notice for unpaid rent with fourteen days to respond. Most of the help that renter needs is words: what the notice means, which defenses might apply, what to file and by when. Software can already prepare much of it. What the renter usually can’t do is pay for a licensed lawyer’s time.

That gap is measured. In its 2022 study, The Justice Gap, the Legal Services Corporation found that low-income Americans got no legal help, or not enough, for 92% of their civil legal problems.

Where the machine helps, and where it stops

The question is no longer whether AI can help. It is which parts of legal help a machine may do, and which stay with a licensed human. The Org.AI Foundation, a non-profit, has written that line down at law.org.ai, and published it before filing any petition, so that regulators can read it first.

Its catalog names five reserved acts, and beside each, the work a lawyer may supervise a machine in doing:

Everything not on the list (the research, the drafting, the forms, the timelines, the translation of legalese) is preparatory work. The machine carries a decision. It never makes one.

The lawyer stays

The design rule at law.org.ai is short: a licensed lawyer is engaged first, then the machine, never the reverse. The machine works as the lawyer’s supervised instrument, the lawyer reviews its work before a client relies on it, and every direction is logged while the client’s own words stay in the lawyer’s file.

The order matters because of privilege. In February 2026, in United States v. Heppner, Judge Rakoff of the Southern District of New York held that a defendant’s chats with a consumer AI tool were not privileged. He had used it on his own, not at his lawyer’s direction. The court left open whether use directed by counsel could come out differently.

That opening points back to United States v. Kovel, the 1961 Second Circuit decision that an accountant working at a lawyer’s direction can fall inside the attorney-client privilege. Supervision is the condition that does the work.

Why Minnesota first

The Foundation starts with Minnesota, and makes the case at law.org.ai/mn: its supreme court regulates its own bar, and has already extended privilege to supervised help twice, to certified law students in 2013 and to legal paraprofessionals in 2020, a pilot the court made permanent in 2024. Both came with the same conditions:

• certification by the Board of Law Examiners;
• a named supervising lawyer who carries personal responsibility;
• the client told, and accepting;
• a duty of confidentiality.

The Foundation’s proposal asks for the same shape for a supervised machine, to be tried first in a free, supervised law-school clinic. That clinic is the next step. The design also learns from Utah, whose regulatory sandbox reported about one complaint for every 5,869 services through April 2025, and then ran into problems the Minnesota design is written to avoid.

Where Marlow comes in

Marlow is working with lawyers on this initiative, to bring supervised legal help to the people who need it most and can least afford to ask.

The fit is the line itself. On Marlow, the reserved act stays with the lawyer: a licensed professional performs it under their own license, in their own judgment. Software prepares the work, and the lawyer decides what is theirs to decide.

That unbundling is also how legal help gets cheaper. When AI prepares and lawyers judge, a lawyer’s time concentrates where it always mattered, and someone who could never pay for hours of drafting may be able to pay for minutes of judgment.

The renter with fourteen days doesn’t need a machine that pretends to be a lawyer. They need the machine to do the preparatory work, and a licensed human to do the rest. Read the reserved acts and the Minnesota case in full.