MMOOZZGG.com

AI for lawyers

Will AI replace lawyers: which tasks go to the machine and which do not

5 min read6 sources

Will AI replace lawyers is no longer only a theoretical question: on 6 May 2025 the solicitors’ regulator in England approved Garfield.Law, the first firm that gives regulated legal services only through AI. Even there named solicitors answer for the work, and courts around the world fine parties for case law invented by a neural network. Below is which tasks of a lawyer go to AI and which stay with a person.

On this page · 5 sections

What the research shows

In March 2023 Goldman Sachs estimated that generative AI could expose the equivalent of 300 million full-time jobs to automation. In the US about a quarter of work tasks can be automated: 44% of tasks in legal occupations and 46% in administrative ones.

An important detail: 44% is a share of tasks, not of jobs. It is about the part of a lawyer’s work that can go to a machine, not about the end of the profession.

Administrative tasks, where the share is even higher, exist in every law firm too: intake of requests, booking consultations, document flow, deadline reminders. Firms usually start with them, because the result is easy to count in hours.

An exam is not practice either. OpenAI reported GPT-4’s result on the US bar exam at about the 90th percentile; a recalculation by Eric Martinez in the journal Artificial Intelligence and Law (2024) gave below the 69th percentile, and about the 48th on essays.

The practical lesson for a firm: AI already takes on search, drafts and cross-checks, while written argument and the position on a case still need a lawyer. More numbers are collected in the statistics on AI in law.

A firm run on AI: the Garfield.Law experience

Garfield.Law was approved by the Solicitors Regulation Authority (SRA) of England on 6 May 2025. Its service is narrow: debt recovery up to GBP 10,000.

The firm does not offer case law research itself, and named solicitors answer for the work. The regulator let AI into the process but did not take responsibility off the people.

The regulator did not assess whether AI can write documents, but who answers to the client and how control works. A client of a Russian firm will ask the same question on learning that AI worked on the case.

For a Russian firm the lesson is the same: a narrow repeating process, clear rules and a lawyer who answers for the result. Firms usually start a pilot with such a process.

Fines for AI errors

The weak point of AI shows in courts. Damien Charlotin’s AI Hallucination Cases database counts 2,190 cases as of 15 October 2026: the US 1,492, Canada 232, Australia 112, the UK 76, Israel 58, Russia 1.

In May 2026 the Arbitration Court of the West Siberian District fined a company $600 for citing, in a cassation appeal, court rulings that did not exist and were created by a neural network. Media reports call it the first such fine in Russia.

Russia has only one case in the database so far, but the first fine is already there. The excuse “the neural network said so” does not work in court: the court judges the document, not the tool used to prepare it.

In all these cases the party that answered was not the service but whoever signed and filed the document. So a lawyer checks every citation from an AI answer against the primary source before filing, and this rule belongs in the firm’s policy.

What stays with the lawyer

The Supreme Court of Russia drew the same line for judges. Its order No. 59-P of 3 September 2026 approved the Concept for introducing AI technologies into court proceedings: recording hearings, assigning cases and preparing draft rulings. AI stays the judge’s tool, and a person makes the decision.

The Concept gives a firm a ready model for splitting roles: AI prepares the draft and the record, and a person checks, decides and signs. Responsibility for the result does not pass to the program, neither in court nor in a firm.

For a lawyer the line runs in the same place:

1. The position on the case and the choice of strategy.

2. The signature under the document and responsibility for each citation.

3. Negotiations, the court hearing and the conversation with the client.

4. Risk assessment that cannot be reduced to a template.

Drafts, search, cross-checks, timelines and the first review of documents go to AI. That is hours of work, not the profession.

How to rebuild the team’s work

Firms that get hours from AI rather than fines change the process, not only the subscription. Five steps:

1. Choose one repeating process and write down how many hours it takes now.

2. Give AI the draft and the search, and leave the check and the signature with the lawyer. Write the citation check rule into the AI use policy.

3. Work with client documents within the firm’s own environment or with a model on servers in Russia, not in a public service.

4. Reassign the freed hours: junior lawyers get more checking and review tasks, not only drafts.

5. After a month, compare the hours and the number of corrections with the starting figure. If the process gave a result, move to the next one.

Partners need to decide separately how the firm tells a client that AI took part in the work and how this shows in the fee. Write the rule down before the first client question, not after it.

The Garfield.Law model works the same way: the machine runs a narrow process, and a named lawyer answers for the result.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Legal Cheek, 06.05.2025 (legalcheek.com)
  2. MIT DSpace, 30.03.2024 (dspace.mit.edu)
  3. Goldman Sachs Global Investment Research, 27.03.2023 (gspublishing.com)
  4. AI Hallucination Cases Database, 15.10.2026 (damiencharlotin.com)
  5. Peterburg2, citing Rossiyskaya Gazeta, 19 May 2026 (peterburg2.ru)
  6. ppt.ru; Garant, September 2026 (ppt.ru)

Found a mistake? Tell us

Publications

Contact us