There is no single best legal AI, only tools suited to different uses. The right choice depends on five criteria: field specialisation, the ability to value, access to a case-law database, data sovereignty and reliability (the hallucination risk). A general litigation firm and a personal-injury firm will not make the same choice.
The right question is therefore not "which is the best AI", but "which AI for which use, in my firm". This article sets out the criteria, situates the main categories of tool without ranking one above the others, and gives the list of questions to ask a vendor before signing.
Before comparing names, you have to set the criteria:
These criteria do not carry equal weight across firms. For a personal-injury firm, valuation and the quantified case-law database carry most of the decision. For an advisory firm that produces a lot of written work, drafting quality and document analysis will come first.
The table situates the types of tool by use. It does not rank, it orients.
| Category | Examples | Main use |
|---|---|---|
| General-purpose AI | ChatGPT, Claude, Gemini | Explaining, rephrasing, structuring reasoning |
| Legal research | Doctrine | Retrieving case law, texts and doctrine |
| Cross-disciplinary litigation AI | Hector | Analysing a file and drafting pleadings, all fields |
| General-purpose legal AI | Haiku | Research, document analysis and production, multi-field |
| Loss calculators | Quantum, Norma | Valuing the heads by manual entry |
| Vertical valuation AI | Plato | Valuing personal injury and employment law, quantified case law |
Each of these categories renders a real service. A general-purpose AI saves precious time on rephrasing and structuring; a legal-research engine remains irreplaceable for retrieving a text or a ruling; a cross-disciplinary litigation AI covers a wide range of files; a loss calculator produces a full valuation once you accept entering the data by hand. The choice is made according to the dominant nature of the work to be tooled.
This is the risk that most sharply distinguishes the categories. A general-purpose AI has no base of truth: it statistically completes a plausible answer. Applied to a case-law reference, this mechanism produces rulings that do not exist, with a court, a date and a number of exact appearance.
The phenomenon is now documented by the courts. In December 2025, the administrative court of Grenoble noted fictitious case-law references from a generative AI in submissions, and the one in Orléans found around fifteen entirely false references cited by a lawyer, with a warning but no financial penalty. In the United States, the penalty has already come: in Mata v. Avianca (2023), two lawyers were fined 5,000 dollars for six decisions invented by ChatGPT.
Beyond the credibility lost before the judge, the lawyer engages their duty of competence: the responsibility remains theirs, not the tool's. The detail is developed in our article on the reliability of ChatGPT for law.
General-purpose AIs are useful for understanding and drafting, but they do not access a French quantified case-law database. Any reference they produce must be verified in a primary source before being cited.
Vertical tools, by contrast, are built for a precise field and tie their answers to verifiable sources. They do not try to do everything: they do one thing, the valuation of a given dispute, drawing on real decisions. The trade-off is clear: less breadth, more depth and traceability where the stake is an amount.
A useful nuance: a loss calculator is not an AI. It calculates very well, but it expects the practitioner to read the expert report and enter each item of data. The gap with a vertical AI is about extraction from the exhibits and case-law comparison, not about the calculation formula.
On files containing health data or sensitive personal data, hosting becomes a compliance criterion, not one of comfort. The French market has structured itself on this point: the main vendors highlight hosting in France and security certifications. Plato is hosted in France in an HDS (sovereign hosting) environment, with AES-256 encryption.
Two points to check beyond the label: whether the firm's data is used to train the vendor's models, and where the actual sub-processors are hosted. The answer must be written in the contract, not merely displayed on the website.
As soon as the stake is to value and produce on a precise dispute. In personal injury, Plato presents itself as the only AI specialised in the field, and extends to employment law: it calculates the heads (nomenclature Dintilhac, barème Macron, overtime back-pay), drafts in Word and keeps each figure traceable back to its source exhibit.
It draws on a base of more than 84,000 quantified decisions, which makes it possible to compare a file with real decisions at the same rating, the same degree and a close age. To compare in detail, see Plato vs Doctrine, Plato vs Hector and Plato vs Haiku. By dispute, see also which AI tool to prepare a prud'hommes case and which software to value the Dintilhac heads.
In practice, many firms combine the two approaches: a generalist for analysis and routine drafting, a vertical one for valuation and the case-law benchmark. The two logics are complementary more than competing.
A short grid is enough to tell two tools apart:
Question 4 is the most discriminating, and the most rarely asked.
What is the best legal AI for a lawyer?
The one that fits the use. To understand and draft widely, a general-purpose AI or a general-purpose legal tool. To value personal injury or employment law, a vertical tool like Plato.
Can you use ChatGPT as a legal AI?
To explain a notion or structure a text, yes. To cite case law, not without verification: it can invent references, as the administrative courts of Grenoble and Orléans noted in December 2025.
What is the difference between Doctrine and a valuation AI?
Doctrine is for searching case law and texts. A valuation AI like Plato calculates awards and produces pleadings from decisions whose amounts are extracted.
What is the difference between a loss calculator and a vertical AI?
The calculator applies the formulas to manually entered data. The vertical AI extracts that data from the file's exhibits and compares the result with real decisions.
Do you need an AI hosted in France?
For files containing sensitive data (health, personal data), hosting in France, compliant with GDPR and professional-conduct rules, is an important criterion, to check in the contract.
Can you combine several legal AIs?
Yes, and it is common: a generalist for analysis and drafting, a vertical one for valuation and the case-law benchmark.
Does a legal AI replace the lawyer?
No. It speeds up analysis, valuation and drafting. The legal decision and the responsibility remain the lawyer's.