In short: we spent months inside the legal AI market instead of reading about it. Of the vendors we tested, 65% are a prompt layer over somebody else's model and 3% are trained on legal data. Only 2% build anything for the person who actually has the legal problem. Every tool we tested invented something at least once, and the most widely installed AI in law firms is not legal AI at all.
There is no shortage of maps of this market. Legaltech Hub counts more than a thousand generative AI solutions in its directory. Venture funds publish tidy category grids. Review sites rank four hundred tools in a table with a pricing column that is mostly the word 'Enterprise'. All of them answer the same question, which is what exists.
None of them answer the question a buyer actually has, which is what this market is made of. A directory will tell you that a vendor sells contract review. It will not tell you whether that vendor trained anything, wrote any documentation, can be bought without a sales call, or invents case law when you push it. Those things show up from the inside, which means signing up, paying, and using the thing.
So we did. This is what the legal AI landscape looks like in 2026 when you stop reading the category grid and start opening accounts.
What we did, and the conflict of interest we should name first
HAQQ sells legal AI. We are a vendor inside the market we are about to describe, which means you should read every number below knowing who counted it.
Two things follow from that. The first is that we fixed the counting rules before we started, and we argued about them for longer than we spent running the tests. What exactly makes something a wrapper rather than a product? We landed on a rule that cuts against us as readily as against anyone else: if you removed the underlying general model and the thing stopped working entirely, with no retrieval layer, no routing, no legal data and no evaluation of its own, it is a wrapper. By that definition some very well funded companies are wrappers, and so are parts of products we admire.
The second is scoping. Every percentage in this piece describes the vendors we tested, not the market as a whole. Nobody has tested the whole market. Where a number comes from somebody else, we say whose it is.
The method was ordinary and slow. Sign up without announcing ourselves. Pay where payment was possible. Run the same body of real legal work through every tool. Read whatever documentation existed. Push each one until it broke, then look carefully at how it broke. For the vendors that would not sell to us without a call, try to get in anyway, which turned out to be its own finding.
The market is mostly a thin layer over somebody else's model
This is the finding everything else hangs off, so it goes first.
What legal AI products are actually built from
HAQQ benchmark, 2026, share of the vendors we tested
Categories are ours and the boundaries are contestable. The rule we used: remove the general model underneath, and if nothing survives, it is a prompt layer.
Some vocabulary, because these words get used loosely. A prompt layer, usually called a wrapper, is a user interface and a set of instructions sitting on top of a general model like GPT or Claude. A harness is the more serious version: it decides which model handles which step, retrieves the right sources first, calls tools, and checks the result before it reaches you. The difference between the two is where the engineering lives, and it is the difference between a product that degrades gracefully and one that confidently makes something up.
Sixty-five percent being a prompt layer is not automatically damning. A well built interface over a frontier model is genuinely useful, and pretending otherwise would be dishonest. The problem is pricing and positioning. A prompt layer sold at prompt-layer prices is a fair trade. A prompt layer sold as proprietary legal intelligence, at several hundred dollars per seat per month, against a general assistant the firm already pays for, is a trade that stops making sense the moment anybody checks. We wrote about where that collapse shows up in the price list in our teardown of legal AI pricing.
Remove the general model underneath, and in two thirds of this market nothing survives.
Almost nobody is building for the person with the legal problem
Of everything we tested, 2% address the end client. Everything else is built for the lawyer.
Set that against the demand side. The World Justice Project estimates that 5.1 billion people have unmet justice needs, meaning they have a legal problem and no meaningful route to resolving it. That is not a market the legal profession is failing to serve out of indifference. It is a market that cannot be served at the price a billable hour implies, which is precisely the constraint software is supposed to relax.
The industry has instead spent its entire venture cycle building tools for the roughly 1.3 million people in the United States who already have a law licence, and their equivalents elsewhere. There are good reasons for this. Lawyers have budgets, procurement processes and an obvious willingness to pay. Consumers have none of those, plus regulatory constraints on who may give legal advice that vary by jurisdiction and are not going to be resolved by a startup deciding they should be.
We are in that 2%, so treat this as a disclosed position rather than a neutral observation. We shipped a consumer legal AI app and it is much harder than the professional product, for exactly the reasons above. But a market where 98% of the supply points at 0.1% of the demand is not a mature market. It is a market that has found one buyer it understands.
Firms are not choosing a vendor. They are stacking them.
80% of the large firms we looked at are either training their own models or running a router across several. They are not evaluating vendors in order to pick one. They are assembling four or more and deciding, per task, which one answers.
This breaks the mental model most vendor pitches are built on. There is no bake-off with a winner. There is a stack, and the question is which slot you occupy in it, which is a much less comfortable question than who is best.
The other half of the same picture comes from outside our benchmark. The International Legal Technology Association's 2026 technology survey, which covers more than 500 firms, found 94% of firms using or exploring AI, and found the single most widely installed AI tool in law firms to be Microsoft Copilot, at 76% of firms. Copilot is not a legal product. It is a general productivity assistant that arrived bundled with software firms had already bought, and it out-installs everything built specifically for law.
We are deliberately not publishing a ladder of the purpose-built tools underneath that figure, because a vendor ranking its named competitors on somebody else's survey data is not analysis, it is marketing with a citation. The structural point stands on its own: the most installed AI in law is the one that was already there.
Adoption numbers are also doing less work than they appear to. Axiom's 2026 survey of in-house legal teams found only 7% had moved past piloting into measured production, and 83% could not say whether the previous year's AI spend had paid for itself. Those two figures describe in-house departments while the 94% describes law firms, so they are two populations rather than one funnel. Read together they still say the same thing: nearly everyone has started, almost nobody has finished, and most cannot prove the difference. We keep a running tally of the adoption research in legal AI statistics 2026.
The labs arrived, and a plugin is just a packaged harness
Every general model provider is now coming for legal directly. The most recent is OpenAI's Astra for Law, launched on 17 September 2026 with 26 partner plugins.
The detail worth sitting with is that two of the largest independent legal AI companies shipped as plugins inside it. That is a rational move and probably an unavoidable one, since distribution beats independence in the short run. It is also a quiet admission about the architecture. A plugin is a harness that somebody else hosts, bills for and controls the entry point to. If your differentiation is the harness, and the harness now runs inside the lab's product, the thing you are defending is a slot in someone else's menu.
We wrote separately about what Astra does and does not do on real legal work. The short version is that the labs are not better at law than legal vendors. They are better at being where the user already is, and in this market that has repeatedly mattered more.
The claims outrun the products, and the regulators have noticed
5% of the vendors we tested have documentation we would call adequate. Not a marketing page with a features grid. Documentation: what the system does, what it does not do, what it was evaluated on, what happens when it fails. 80% oversell somewhere between the website and the product, meaning the capability described on the landing page is not the capability we found in the account.
This is not a victimless style problem. In 2025 the United States Federal Trade Commission finalised a $193,000 order against DoNotPay and barred it from marketing itself as a robot lawyer, because it could not substantiate the claim that its AI performed like a human one. That case is the floor, not the ceiling, of where this goes.
Accuracy marketing is where it gets sharpest. Vendors in this category have marketed retrieval-grounded research tools on the promise that hallucination had been eliminated altogether. Stanford researchers who tested those tools concluded that the providers' claims were overstated, and published measured error rates in place of the marketing. We are reporting that finding rather than repeating the claim it was made about.
Benchmarks themselves are contested, and honestly so. Vals writes its rubrics with practising lawyers, which is the right way to do it, and vendors still dispute the methodology. That is not a scandal. Legal work is evaluated by judgement, the same way law is, and two competent lawyers can score the same answer differently. Anyone selling you a single accuracy number for a legal system is selling you a simplification. We publish our own benchmark method precisely so it can be argued with.
Hallucination is still the common denominator
Every tool we tested invented something. Not every answer, and not constantly, but every product in the set at least once produced law that does not exist, claimed to have done work it had not done, or built a confident conclusion on a fact it had manufactured.
The external measurements agree on the shape if not the exact figure. Stanford's audit put Lexis+ AI above 17% and Westlaw's AI-assisted research above 33% on the tool versions tested. General models without retrieval have measured between 58% and 88% on legal citation tasks in the Journal of Legal Analysis work. Retrieval grounding helps a great deal. It does not solve the problem, and the errors that survive grounding are the dangerous ones, because they look like real citations attached to the wrong proposition.
The consequences are now counted. Damien Charlotin's public database of court decisions involving AI-hallucinated filings held 1,598 cases on 9 June 2026, the date our audit of it reports. As of 19 September 2026 it holds 2,044, up from roughly 200 a year ago. That curve is the actual state of legal AI safety, and it is the reason our hallucination audit is a living document rather than a one-off post.
One less discussed failure sits next to it. 75% of the tools we tested write in their own hardcoded voice, regardless of whose firm, whose client, or whose house style. Nothing in this market produces a digital twin of how a particular lawyer actually writes. For a profession where the phrasing of a clause is the product, that is a strange thing for an entire category to have skipped.
Try HAQQ AI Free
Experience AI-powered legal drafting and research
The pricing wall is mostly theatre
Only 10% of the vendors we tested are genuinely self-serve. Of twenty vendors we checked for a published price, 14 publish none at all. The category standard is a contact form.
And yet 75% of them were straightforward to get into without ever speaking to a human. Trials, partner routes, self-service tiers behind a sales page, and in several cases simply completing a checkout the marketing site did not link to. We did this on every single vendor we tested.
That gap is the interesting part. A sales-gated price usually signals either genuine enterprise complexity or price discrimination. When the same product turns out to be buyable in four minutes by anyone who looks slightly harder, it was the second one. The full accounting of who publishes what, and what it costs when you get past the gate, is in legal AI pricing in 2026.
MENA does not buy the landscape. It builds it.
Most of this market is localised one jurisdiction at a time, which is unsurprising, since law is. What is less widely understood outside the region is that in the Middle East and North Africa the buyer is frequently the state, and the state builds.
The United Arab Emirates launched in July 2026 what it describes as the first fully integrated AI judicial platform. Saudi Arabia runs Najiz, which carries roughly 160 justice services, alongside a state-owned AI company behind an Arabic language model. These are not procurement decisions waiting for a vendor to win them. They are infrastructure programmes that a vendor can at best supply a component to.
At the firm level the same instinct repeats. Firms here build their own AI and then sell it to other firms, which is a route almost nobody takes in London or New York and which several MENA practices now treat as a second revenue line. We covered the regional picture in more depth in legal AI in MENA and legal AI in Saudi Arabia.
The implication for anyone mapping this market from outside the region is simple. A vendor landscape drawn from Western funding data will miss the largest legal AI deployments in MENA entirely, because they never appear as a funding round.
What the market looks like from above
Tracxn's sector data counts 10,528 legal tech companies globally, 1,821 of them funded, and 19 unicorns. Against the composition data above, that is a great many companies for a market where 8% of the products we tested contain a model anybody trained.
Two vendors dominate the funding narrative. Harvey, reported at roughly $11 billion in valuation and around $300 million in ARR, and Legora at roughly $5.55 billion and around $150 million, raised $750 million between them inside 15 days. Those revenue figures are press-reported rather than audited, and neither company publishes them itself, so treat them as the market's best guess. We track the quarterly movement in our legal AI market report.
Scale follows funding. 90% of the funded vendors we tested are past 200 people, and only 20% publish an ARR figure of any kind. A category this large that discloses this little is a category where buyers cannot compare vendors on anything except the demo.
The landing pages tell you what the category believes converts. The four elements that recur are Security, a prompt library, testimonials, and a Labs section claiming deep AI research. Security leads, and there is a reason for the timing: generative AI has just entered the legal industry's tracked security-risk list, straight in at number two. Buyers are frightened, and the market has correctly read that fear as the first objection to clear. Our own security posture is published for the same reason, and we would rather say that plainly than pretend we are above the pattern.
On the right task, it already beats the lawyer
It would be easy to read the preceding sections as a case that legal AI does not work. That is not the finding.
In the Vals legal AI report, the strongest system on document question answering scored 94.8%, which is 24.7 points above the human lawyer baseline measured in the same study. On redlining, that same system lost to the human control group. Both results belong to one product in one evaluation.
That is the most useful single fact in this entire landscape. Capability is not a property of the tool, it is a property of the pairing between the tool and the task. Point one of these systems at finding what a 400-page document says and it outperforms a trained professional who is tired. Point the same system at deciding which of two acceptable clause formulations protects your client better, and it produces something that reads like an answer and is not one.
Most failed legal AI pilots we have seen are not capability failures. They are aiming failures. A firm buys a system, aims it at the hardest judgement-heavy work in the building because that is where the pain is, watches it underperform, and concludes the category is immature. The task-by-task breakdown is more useful than any overall score.
What we take from this
The lesson we keep returning to is about distribution rather than intelligence, and it cost us something to learn.
Our first serious product asked lawyers to come to a new tab and do their work there. It was a better tool than what they had. It lost anyway, and it lost to habit. Every product in this market that asked a lawyer to leave Word, Outlook or the document management system has lost the same way, which is why a general productivity assistant that was already installed out-installs every purpose-built legal AI product built by people who think about law all day.
That reframes what is actually scarce here. It is not model quality, which is converging and is largely bought from four companies. It is not legal knowledge, which is written down. It is the right to occupy the place where the work already happens, plus the discipline to be honest about which tasks the system is good at. Those are the two things nearly all of the 10,528 companies are competing for, whether or not they describe it that way.
We are one of those companies and we do not get to exempt ourselves from any of the above. What we can do is publish the counting, including the parts that are uncomfortable for us, so that the next person mapping this market starts from something better than a category grid.
Key Takeaways
- Two thirds of the market is a prompt layer. Of the vendors we tested, 65% stop working entirely if you remove the general model underneath, and 3% are trained on legal data.
- Buyers are stacking, not choosing. 80% of the large firms we looked at run several systems or their own router. There is no bake-off with a winner.
- The most installed AI in law firms is not legal AI. ILTA puts Microsoft Copilot at 76% of firms, ahead of anything built for law, because it was already on the desktop.
- Adoption is near universal and completion is rare. 94% of firms are using or exploring AI, while only 7% of in-house teams have moved past piloting and 83% cannot prove last year's spend paid for itself.
- Hallucination is still the common denominator. Every tool we tested invented something, external audits put grounded legal tools between 17% and 33%, and the court-decision tracker has gone from about 200 cases to 2,044 in a year.
- The pricing wall is mostly theatre. 14 of 20 vendors publish no price, yet 75% were buyable without ever speaking to a salesperson.
- Almost nobody serves the end client. 2% of what we tested addresses the person with the legal problem, against 5.1 billion people with unmet justice needs.
- Capability is about aiming. The same system scored 94.8% on document question answering, 24.7 points above the human baseline, and lost to humans at redlining. Most failed pilots are aiming failures, not capability failures.
Sources and further reading
- Legal AI Statistics 2026: how many lawyers actually use AI
- Legal AI Pricing in 2026: every published price, and every vendor that publishes none
- AI Legal Hallucination Audit: the court cases with fabricated citations
- Our quarterly legal AI market report
- Legal AI in MENA in 2026
- How we run our legal AI benchmark
- World Justice Project: measuring the global justice gap
- Stanford RegLab and HAI: assessing the reliability of AI legal research tools
- Dahl et al., Large Legal Fictions, Journal of Legal Analysis
- Damien Charlotin: AI hallucination cases database
- FTC final order against DoNotPay
- Vals Legal AI Report
