BRENNEN

Industries / Finance and business services / Legal and professional services / LegalTech

Help for people without lawyers

Books, software, services and rule changes that help people handle a legal problem without a lawyer.

7

Milestones achieved, from 1971 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1971

    Self-help law books

    Two former legal aid lawyers in California, Ralph Warner and Charles Sherman, started Nolo Press in 1971 with two books, "How to Do Your Own Divorce in California" and "The California Tenants' Handbook." Nolo later sold WillMaker, software for writing a will. In January 1999 a federal judge in Texas ruled that selling the Quicken Family Lawyer software was the unlicensed practice of law; the Texas legislature then changed the law, and on 29 June 1999 the Fifth Circuit lifted the ban.

    LessonTexas law now leaves out of the practice of law any software that "clearly and conspicuously" states it is not a substitute for a lawyer's advice.

    encyclopedia.comca5.uscourts.gov

    Credit

    Ralph Warner, Charles Sherman

  2. 2005

    Guided interviews fill forms

    A2J Author lets court staff and legal aid lawyers build guided interviews: web pages that ask plain questions and fill in court forms from the answers. CALI, the Center for Computer-Assisted Legal Instruction, which makes it with IIT Chicago-Kent College of Law, says more than 1,000 such interviews run on the national LawHelp Interactive server and that people without lawyers have used them more than 7 million times since 2005.

    cali.org

    Credit

    CALI; IIT Chicago-Kent's Center for Access to Justice & Technology

  3. 2017

    A Dutch platform shrinks

    Rechtwijzer, built by the Dutch Legal Aid Board, the Hague institute HiiL and the company Modria, guided separating couples to an agreement online, with mediators and a decision-maker if needed. In 2017 the partners split up "for commercial reasons"; it handled about 700 cases a year, against roughly 35,000 Dutch divorces. HiiL wrote, "We have not been able to make it financially viable in the way we had hoped."

    LessonA tool that works still fails if too few people use it and no one keeps paying for it.

    legalfutures.co.uk

    Credit

    Dutch Legal Aid Board, HiiL, Modria

  4. 2020

    Rules let non-lawyers help

    On 14 August 2020 Utah's Supreme Court opened a regulatory sandbox, a supervised trial in which companies and non-lawyers may offer legal services that the usual rules forbid. Arizona's Supreme Court ended its ban on non-lawyers owning law firms the same month, began licensing these "alternative business structures" on 1 January 2021, and had 114 active licenses at the end of 2024.

    LessonChange was smaller than hoped: 47 of Arizona's 51 new licenses in 2024 went to traditional law firms, and in September 2024 Utah barred for-profit immigration services from its sandbox.

    iaals.du.eduazcourts.gov

    Credit

    Utah and Arizona Supreme Courts; IAALS, a University of Denver institute, for its evaluation of Utah's sandbox

  5. 2020

    Court forms on phones

    When Massachusetts courts closed to walk-in visitors early in the pandemic, Suffolk University Law School's Legal Innovation and Technology Lab and volunteers built phone-friendly guided interviews for court forms, using Docassemble, free open-source software written by Jonathan Pyle. By August 2022, by the school's count, more than 20,000 forms had been downloaded, including 3,500 fee waivers, and the project won the College of Law Practice Management's InnovAction Award that year.

    suffolk.edubostonbar.org

    Credit

    Suffolk LIT Lab; Jonathan Pyle (Docassemble)

  6. 2025

    Robot lawyer claims stopped

    DoNotPay sold a subscription it advertised as "the world's first robot lawyer." On 11 February 2025 the US Federal Trade Commission finalized an order that bars it from claiming, without evidence, that its service works like a lawyer, and requires it to pay $193,000 and to notify subscribers from 2021 to 2023. The FTC found that the company "did not test whether its 'AI lawyer' operated to the level of a human lawyer."

    LessonA claim that software matches a lawyer needs a test behind it.

    ftc.gov

    Credit

    US Federal Trade Commission

  7. 2025

    Non-lawyer advice case lost

    Upsolve, a nonprofit, trained non-lawyer "Justice Advocates" to give free, limited advice to New Yorkers sued over debts, and in 2022 won a preliminary injunction against New York's ban on unlicensed legal practice. The Second Circuit vacated it on 9 September 2025; Judge Lewis Kaplan then ruled for New York on 6 March 2026, and the Supreme Court denied review on 30 March 2026.

    LessonThe ban covers advice "to a specific person about that person's individual legal problems", and leaves self-help materials free to publish.

    probonoinst.orgreason.comsupremecourt.gov

    Credit

    Upsolve and Rev. John Udo-Okon, the plaintiffs; the Second Circuit

Watch for

  1. Legal aid AI time savings

    Thomson Reuters gives legal aid groups its CoCounsel AI tool through an "AI for Justice" program begun in October 2024, and says users save "Approximately 250 hours" per lawyer a year. The figure comes from 44 survey answers from users at nine organizations, and LawNext reports groups "serving as many as 50% more clients daily."

    Achieved whenan independent evaluator compares clients served and hours spent against a baseline or a control group.

    thomsonreuters.comlawnext.com

    Claimed by

    Thomson Reuters, October 2025

  2. Cheap AI sorts intake

    Intake is the first step of legal help, when a request is read and sent to the right service. Quinten Steenhuis of Suffolk University Law School reports that FETCH, which combines low-cost AI models, keyword matching and Suffolk's Spot classifier, put 419 real requests to the Oregon State Bar's lawyer referral service in the right category, within its top two guesses, 97.37% of the time, and did better than GPT-5 alone.

    Achieved whensomeone else gets similar accuracy on held-out requests from another intake service.

    arxiv.org

    Claimed by

    Quinten Steenhuis, September 2025

Open

  1. 01

    A form-finding test

    300 invented problem stories for one state, each labeled with the correct official court form and its category in LIST, the Legal Issues Taxonomy from Suffolk's LIT Lab, and a measure of how often AI chat tools and court chatbots name the right form. The stories are invented, so no real person is advised.

    The brief
    Credit · unclaimed
  2. 02

    Form translations, scored

    Take the court forms that a state has already had translated by people, machine-translate the English originals with several tools, and score the errors against the court's own versions in five languages.

    The brief
    Credit · unclaimed

Open 01 · the brief

A form-finding test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

300 invented problem stories for one state, each labeled with the correct official court form and its category in LIST, the Legal Issues Taxonomy from Suffolk's LIT Lab, and a measure of how often AI chat tools and court chatbots name the right form. The stories are invented, so no real person is advised.

Why it is still open

Stanford's quality benchmark for AI legal help is still a draft. The published tests I found, LegalBench's issue-spotting tasks (2023) and FETCH (2025), sort stories into legal categories, not to a specific court form.

justiceinnovation.law.stanford.eduarxiv.orgarxiv.org

Built with what exists

one state's forms, such as Massachusetts' through Suffolk's Court Forms Online; the LIST taxonomy; the Spot classifier as a baseline; Claude, Gemini, GPT and an open-weight model; a scoring script.

  1. First step

    List the 30 most-used self-help forms in one state and write 10 stories for each.

  2. Done looks like

    300 stories with answer keys on which two reviewers agree at least 90% of the time; the share of stories where each tool names the right form first and within its top three, with 95% confidence intervals; and errors sorted by kind (wrong form, outdated form, another state's form), with the data released.

Open 02 · the brief

Form translations, scored

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Take the court forms that a state has already had translated by people, machine-translate the English originals with several tools, and score the errors against the court's own versions in five languages.

Why it is still open

The NCSC's 2025 guide to machine translation for courts, funded by the State Justice Institute, says "we do not have a lot of data about reliability in legal translation," and the accuracy figures it cites (94% for Spanish, 82% for Chinese) come from Google-translated medical discharge instructions, not legal text.

ncsc.org

Built with what exists

Maine's translated court forms, which come in nine languages including Arabic, Chinese, Somali, Spanish and Vietnamese; Google Translate, DeepL, Claude, Gemini, GPT and Meta's open NLLB translation model; automatic scores such as chrF and COMET; bilingual reviewers marking errors with MQM, a standard list of translation error types.

courts.maine.gov

  1. First step

    Download Maine's translated forms and their English originals and pair them by form number.

  2. Done looks like

    At least 20 documents in each of five languages, every machine output scored by a qualified bilingual reviewer, and serious errors per 1,000 words reported for each tool and language, with all data released.