MMOOZZGG.com

AI for lawyers

Harvey AI and AI at the world’s big law firms

7 min read10 sources

Harvey AI, an AI service for lawyers, raised 550 million dollars at a valuation of 15.5 billion dollars on 9 September 2026, and according to the company, 80% of the Am Law 100 firms use it. Below: how large firms and legal publishers adopt AI, and what from their experience applies to firms in Russia.

On this page · 5 sections

Harvey

Harvey began as a platform on a version of OpenAI’s models tuned for legal work. Its first major client was Allen & Overy in 2023; more on that below.

In September 2026 the company raised 550 million dollars at a valuation of 15.5 billion dollars. The round was led by Diffusion and Lightspeed Venture Partners, and earlier investors also took part.

Clients, according to the company: 80% of the Am Law 100, the ranking of the hundred largest US law firms by revenue, and the legal departments of five Fortune 10 companies.

Industry surveys confirm the demand for such services. According to the Clio Legal Trends 2025 report, 79% of lawyers use AI: 87% in large firms and 71% among solo practitioners. At the same time, 53% say their firm has no AI policy or they don’t know of one.

Two steps came before the round. Harvey presented its first own model, fine-tuned from an open-weight model, and the Harvey LAB benchmark for evaluating legal AI agents.

The company describes the goal of the round as helping firms and legal departments build and own their own intelligence. For the market, this is a signal: a service valued at 15.5 billion dollars itself invests in open-weight models that can be fine-tuned for a task.

Thomson Reuters and Casetext

On 17 August 2023 Thomson Reuters closed the purchase of Casetext for 650 million dollars in cash. Casetext’s main product, CoCounsel, was an AI assistant for lawyers built on GPT-4.

What CoCounsel could do at the time of the deal:

1. Document review.

2. Case law research memos.

3. Deposition preparation.

4. Contract analysis.

Casetext had more than 10,000 clients among law firms and legal departments. At the same time, Thomson Reuters announced it would invest more than 100 million dollars a year in AI for its legal databases and products.

The logic of the deal is clear to a lawyer. A model on its own doesn’t know current law; the value comes from a verified base of sources that it consults before answering. The legal publisher bought a ready AI team and joined it to its database.

A base of sources lowers the number of errors but does not remove them. In May 2024, Stanford RegLab and HAI found errors in more than 17% of answers from Lexis+ AI and Ask Practical Law AI and in more than 34% of answers from Westlaw AI-Assisted Research. In an earlier study, general chatbots were wrong on 58–82% of legal questions. So every citation is checked against the original source; how this ends in court is covered on the page about AI hallucinations in court.

The same principle works at the level of a firm. Its own base (contracts, opinions, case positions) is worth more to AI than the model’s general knowledge, and that is exactly what must not go to someone else’s service.

Allen & Overy

On 15 February 2023 Allen & Overy opened Harvey to more than 3,500 lawyers in 43 offices. Today the firm works under the name A&O Shearman.

A test from November 2022 came before the launch. It was run by the Markets Innovation Group, the firm’s team of lawyers and developers, and by the end of the test about 3,500 lawyers had asked about 40,000 questions in their daily client work.

The tasks the firm named at launch: contract analysis, due diligence before a deal, litigation and regulatory compliance.

The way A&O rolled it out works as a template for a small firm too:

1. A test by a small team of lawyers and developers on real tasks.

2. Extending the test to practicing lawyers and collecting questions from daily work.

3. Launch for the whole firm with a list of tasks for which the tool is allowed.

4. A rule that a lawyer checks the result, written down before launch, not after the first error.

The firm wrote the condition straight into the launch announcement: the output needs careful review by a lawyer of the firm. AI prepares a draft and finds information, and the lawyer checks and signs.

This approach matches what lawyers’ regulators require. The duties of a lawyer working with AI are covered in detail on the page about AI ethics for lawyers.

What is available in Russia

The first barrier is availability. Russia is not on the list of countries where the OpenAI API is officially available, and under Anthropic’s policy of 12 May 2026 its products and API are available only in the listed regions, with no Russia on the list (both lists checked on 16 October 2026).

At launch, Harvey ran on OpenAI models. For a Russian firm, this means foreign legal services on such models are not officially available.

The second barrier is legal. Case details and clients’ personal data sent to a foreign service need a separate check under 152-FZ, Russia’s personal data law, and the rules on professional secrecy. The procedure is covered on the page about cross-border data transfer.

The third factor is legislation. Since 1 September 2026, 243-FZ has been in force: data of sovereign and national models is processed in Russia, and the government decides when such models must be used.

So a subscription in the style of the Am Law 100 firms is not an option for a Russian firm. There are two working paths: an open-weight model on the firm’s own server, or a model in a cloud in Russia with a data processing agreement.

Your own model instead of a subscription

The experience of Harvey, Thomson Reuters and A&O gives three conclusions. AI works on a base of documents and sources, not only on the model’s general knowledge. A lawyer checks the result. The firm’s data and knowledge stay with the firm.

A closed setup for a Russian firm works like this:

1. The model runs on the firm’s server or in a cloud in Russia, and text doesn’t go abroad or to services without a data processing agreement.

2. The AI has its own account with rights only to its own process, for example to read the contract archive.

3. The AI’s actions go into a log; decisions, signatures and negotiations stay with the lawyers.

4. The model can be replaced, while the system, data and access stay with the firm.

You can estimate how many hours this returns to lawyers in the time savings calculator: it counts by the number of lawyers, the hourly rate and the share of routine work, and the data never leaves your device.

How to choose a model and server for your document volume is covered on the page about a local neural network for lawyers. It’s easier to start with one process: a pilot of one AI agent costs from $1,499 and takes 14 days.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Harvey, 09.09.2026 (harvey.ai)
  2. Thomson Reuters, 17.08.2023 (thomsonreuters.com)
  3. A&O Shearman, 15.02.2023 (aoshearman.com)
  4. ppt.ru, 26.07.2026 (ppt.ru)
  5. OpenAI, Supported countries and territories (platform.openai.com)
  6. Anthropic, Supported Regions Policy, 12.05.2026 (anthropic.com)
  7. Data security in AI adoption
  8. MMOOZZGG.com pricing and timelines
  9. 2Civility (review of the Clio report), 2025 (2civility.org)
  10. Stanford HAI, 23.05.2024 (hai.stanford.edu)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us