Harvey, Legora, Claude: what the legal AI race changes for a legal department

Three trends marked the year: funding rounds of a size never seen before in the sector, the entry of a general-purpose model provider into legal work, and a stronger presence of these players in France. The main milestones, with dates:

• February 2026. Anthropic launches a legal plugin for Claude Cowork: contract review, triage of non-disclosure agreements, compliance tracking. Legal IT Insider dates the announcement to February 3.
• March 2026. Harvey is valued at $11 billion, according to TechCrunch.
• April 2026. Legora reaches a $5.6 billion valuation, with a $50 million extension to its round, according to TechCrunch.
• May 2026. Harvey announces a library of more than 500 agents and opens an office in Paris. On May 12, Anthropic adds more than 20 connectors and 12 plugins by practice area, according to LawSites.
• September 2026. Harvey announces $550 million raised at a $15.5 billion valuation, according to Legal IT Insider.
• October 2026. Legora announces $550 million raised at an $8.8 billion valuation, according to Le Monde du Droit.
• November 2026. On November 4, Harvey and Jus Mundi announce the integration of the French agent Jus AI into Harvey.

General-purpose models are not designed for law alone, but lawyers already use them. According to the 2025 barometer published by Oxygen+ with the AFJE and the Cercle Montesquieu, 87% of the 387 professionals surveyed had already used a generative AI tool such as ChatGPT. This figure measures at least occasional use, not a daily practice or a tool approved by the company.

The providers of these models set limits on this use. Anthropic’s public repository presents each output of its plugins as a draft meant for review by a lawyer, not as legal advice. OpenAI’s usage policy of October 2025 requires the involvement of a licensed professional for tailored legal advice. Both positions point the same way: the provider does not replace the lawyer.

The layers are not always mutually exclusive. Among the Claude connectors announced in May 2026, LawSites mentions Harvey, Thomson Reuters’ CoCounsel, Ironclad and DocuSign. The same team can therefore combine a model, a legal database and a contract tool. It must then track what each link in the chain receives and keeps.

The answer depends on the task, and independent measurements remain rare. The Vals AI study on legal research published in October 2025 gives an idea. It compared three legal AI tools, ChatGPT and lawyers on 210 questions, according to the LawSites report.

• Accuracy. The tools scored 79% to 81%, ChatGPT 80%, and the lawyers 71%.
• Source reliability. ChatGPT scored 70%, against an average of 76% for the legal AI tools. Vals AI links this gap to access to proprietary databases.
• Limits. Several major providers did not take part, and the study does not cover drafting. The report does not mention French law either.

These figures therefore do not support a conclusion for a French contract. They only show that the quality of an answer must be distinguished from the quality of its sources. For a lawyer, the second matters as much as the first, because it is what makes verification possible.

The Vals AI report of February 2025, focused on specialized tools, points the same way. The best tool outperformed the lawyer baseline on summarization and document questions, but the lawyers kept the best score on redlining, at 79.7%. These results relate to an earlier version of the products.

The French market is now directly part of this competition. According to Maddyness, Harvey announced its Paris office on May 12, 2026 and lists August Debouzy, Bredin Prat, CMS Francis Lefebvre and the Adecco group among its references. In June 2026, Legora announced the opening of a Paris office in the third quarter, and says it has clients in France without naming them in that press release.

The agreement between Harvey and Jus Mundi illustrates another trend: a global platform that relies on a specialized database. Its scope is narrow. It covers international arbitration and cross-border disputes, and it targets the two companies’ shared clients, by the end of the year according to the announcement. It is therefore not general coverage of French contract law.

For a company of 100 to 500 people, these announcements have a practical effect: the offering is broadening and the messages sound alike. Many public references also involve large law firms or large groups. It remains to be checked what they say about your own context.

A few simple questions make it possible to compare very different offerings on useful criteria, without stopping at how well known they are:

• What work do you want to give the tool? Reviewing an incoming contract, researching the law and tracking deadlines are three separate needs.
• What law is covered? Ask which French sources are queried, and whether the cited references can be opened.
• Where do your documents go? Identify the subprocessors, the processing locations and the terms for data reuse, offering by offering.
• Who checks, and when? The product must make it possible to check an answer against the passage or source it relies on.
• What happens to the information after the answer? A useful analysis should be able to join the contract, its version and its deadlines, or your team will have to re-enter it.
• What proof does the provider offer? Distinguish independent studies, the provider’s own figures and a trial on your own contracts.

A trial on a few real contracts whose difficulties you already know is often more telling than a demo. It requires settling the confidentiality question beforehand.

Responsibility stays with the company and its lawyers. In the presentations we read, no provider presents its output as a decision to follow without review. Review therefore still has to be organized, even as the tool becomes faster.

Document disorder does not disappear either. An AI cannot work with a contract that no one can find, or guess which of three versions was signed. Before changing tools, it can help to know where your contracts are and who tracks their deadlines.

Finally, the pace of the market calls for caution in commitments. The features and terms of these offerings change within a few months. A dated evaluation, reviewed at regular intervals, is better than a fixed ranking.

Docform is a contract lifecycle management platform, with AI built into that lifecycle. The assistant summarizes a contract, flags points of attention based on your review playbook, cites the source passage and queries Légifrance and Judilibre. The lawyer checks and decides. Contracts are hosted in France, and the Trust page lists the subprocessors, including Azure OpenAI for the assistant.