Legal AI Tools in 2026: What They Do Well, and Where Mauritius Falls Outside Them

04 August 2026 · Jenny Legal AI

Harvey, CoCounsel and Lexis+ AI are genuinely good products. None of them are built on Mauritian law — and for a mixed jurisdiction, that is the whole problem.

The legal AI market has consolidated fast. If you have looked at tooling in the last year you will have seen the same four or five names, and most of the comparison articles you find are written for a firm in New York or London deciding how to spend a six-figure software budget.

This is not that article. The question here is narrower and more useful if you practise in Mauritius: what do these tools actually do, and what happens when you point them at Mauritian law?

The short answer is that the products are good and the coverage is not there. The longer answer is worth understanding, because it explains something structural about how legal AI works — and why the gap is unlikely to close on its own.

What the major tools are, as of August 2026

Vendor facts age quickly in this market. Everything below is current as at the date of this post; check it before relying on it.

Harvey is the best-known name and the one most often raised in conversation. It is aimed squarely at large firms — the Am Law 100 tier, with AI budgets in the six figures — and it has grown quickly, reaching a reported $11 billion valuation and over a thousand customers across roughly 60 countries. Notably for this discussion, Harvey has been expanding its jurisdictional coverage aggressively: from around six jurisdictions in mid-2025 to 60-plus a year later, moving out of pure common-law markets into civil-law jurisdictions across Europe, Latin America and Asia-Pacific. Its "Data Factory" now ingests several hundred legal data sources.

CoCounsel, from Thomson Reuters, is the incumbent play. Its strength is that it sits inside the Westlaw stack, so research, drafting and the underlying corpus are one system rather than three. For US legal research it is formidable. Its acknowledged weakness is jurisdictional depth outside the United States.

Lexis+ AI (and Protégé) is the LexisNexis counterpart, grounded in the LexisNexis corpus. Grounding matters more than marketing here. Stanford's benchmark study, AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries, tested purpose-built legal research tools against more than 200 queries and found Lexis+ AI and Thomson Reuters' Ask Practical Law AI produced incorrect information more than 17 per cent of the time, while Westlaw's AI-Assisted Research hallucinated more than 34 per cent of the time. The study counted two failure modes: answers that are factually wrong, and answers that are legally correct but cite sources which do not support them — "misgrounded" responses.

Those measurements are from an earlier generation of these products and the vendors have iterated since, so do not read them as current scores. Read them as evidence of the failure mode: a tool grounded in a real corpus, answering questions about its own jurisdiction, still got it wrong often enough that unverified output was unsafe.

Legora (formerly Leya, out of Stockholm) has emerged as the main European enterprise alternative to Harvey, with a similar large-firm focus.

There are others, and the list will be different in six months. The pattern, though, is stable.

The pattern: these are corpus products, not chat products

The thing that separates a serious legal AI tool from a general chatbot with a nice interface is not the model. Everyone has access to broadly similar frontier models. The difference is what the system is grounded in — whether, when it tells you something, it is retrieving from an actual body of legislation and case law and citing it, or generating plausible text.

That distinction is why the Stanford numbers above matter, and it is why Thomson Reuters and LexisNexis are competitive at all: they have spent decades building the corpus. Harvey's answer to the same problem is to industrialise ingestion — hence "Data Factory" and the jump from six jurisdictions to sixty.

Which leads to the obvious question.

Where does Mauritius sit in all this?

Nowhere, in practice, and for reasons that are not really the vendors' fault.

The market is too small to prioritise. A vendor choosing its next jurisdiction weighs the cost of acquiring, structuring and maintaining a national corpus against the addressable revenue. Mauritius does not win that calculation against Brazil, Germany or India. Every jurisdiction on a global roadmap is a jurisdiction chosen because it pays for itself.

The corpus is genuinely hard to build. Mauritian primary sources are public but not neatly packaged. Legislation is available through the Supreme Court's legislation resources, with recent material from 2008 onward in full text. Judgments are searchable through the Supreme Court judgment search, and material from a number of other courts is freely available — but Supreme Court decisions have historically been distributed on a subscription basis, and the Supreme Court Library holds a great deal that is not online at all. This is a normal state of affairs for a jurisdiction of this size. It is also exactly the kind of fragmentation that makes automated ingestion expensive.

And the law itself does not fit a single template. Mauritius is a mixed jurisdiction: French-derived civil law, with the Code Civil Mauricien still governing much of the private-law field, sitting alongside English-derived commercial and procedural law and a substantial body of local statute. A system trained and tuned for common-law research does not transfer cleanly. Neither does one built for a pure civil-law jurisdiction. Harvey's expansion into civil-law markets is real progress on that front, but "handles civil-law reasoning" is not the same as "knows the Workers' Rights Act 2019."

What this means in the room

The practical failure mode is not that a global tool refuses to answer a Mauritian question. It is that it answers confidently using the nearest jurisdiction it does know.

Ask a general legal AI about notice periods on termination and you will get something reasonable-sounding and structured around English or US employment concepts. What you will not get is section 63(4) of the Workers' Rights Act 2019, which fixes a 30-day minimum notice period notwithstanding any agreement to the contrary — a provision that overrides the contract you are looking at.

Ask about the cost of an unjustified termination and you may get a discussion of reasonable notice and mitigation. What you need is three months' remuneration per 12 months of continuous employment under sections 69 to 71, plus interest of up to 12 per cent — a figure that is usually far larger than the intuition imported from elsewhere.

Ask about data protection compliance and you will get GDPR. Mauritius tracks the GDPR closely, which is precisely what makes this dangerous: the Data Protection Act 2017 requires registration with the Commissioner before acting as a controller or processor, an obligation the GDPR deliberately abolished. There is nothing in a GDPR-derived programme to copy across, so it is simply absent.

In each case the answer is fluent, well-organised, and wrong in a way that is hard to spot unless you already knew the answer. That is the worst possible failure mode for a research tool.

The honest comparison

If you work at a firm with a US or UK practice, the global tools are worth their price and this article is not an argument against them. CoCounsel inside Westlaw is a genuinely strong research environment for American law. Harvey is doing serious work on multi-jurisdictional coverage.

But three things follow for a Mauritian practice:

1. Coverage is the specification, not a feature. When you evaluate any legal AI tool, the first question is not what the interface looks like or which model sits underneath. It is: which body of law is this grounded in, and can it show me the section it relied on? If it cannot cite Mauritian primary sources, it is a drafting assistant, not a research tool.

2. Verification is not optional at any price point. The best result in Stanford's benchmark was still better than one wrong answer in six, and that was a well-resourced product answering questions about its own jurisdiction. Outside the corpus, nobody is measuring the error rate at all.

3. The gap is structural, so it needs a local answer. Nobody is going to build the Mauritian corpus as a side effect of expanding into Brazil.

Where Jenny fits

That last point is the reason Jenny exists.

Jenny is a legal AI assistant built specifically on Mauritian law — the Companies Act 2001, the Workers' Rights Act 2019, the Data Protection Act 2017, the Income Tax Act, the Divorce and Judicial Separation Act 1981 and the relevant articles of the Code Civil Mauricien, with more being added. It works in English and French, because Mauritian practice does.

The design principle is the one this whole article turns on: answers are grounded in retrieved statutory text and cite the sections they rely on, so you can check them. Where the source material does not support an answer, the right behaviour is to say so rather than to fill the gap with the nearest foreign analogue.

It is not trying to be Harvey. It is trying to be correct about the jurisdiction you actually practise in — which, as it turns out, is the part the global tools cannot buy their way into.

This article is general information about Mauritian law, not legal advice. Legislation is amended and courts reinterpret it. Verify the current text of any provision and consult a qualified Mauritian legal practitioner before acting.