The Legal AI Landscape in 2026: What It Is Made Of

We tested the legal AI market instead of reading about it. 65% of vendors are a prompt layer over somebody else's model, and 3% train on legal data.

Key facts

• 65% of the vendors we tested are a prompt layer over a general model; 12% are a harness; 5% train or substantially post-train their own model; 3% train on legal data (HAQQ vendor test, 2026).
• Microsoft Copilot is installed at 76% of law firms, and 94% of firms are using or exploring AI (ILTA 2026 Technology Survey, 500+ firms).
• 7% of in-house legal teams have moved past piloting into measured production, and 83% cannot say whether last year's AI spend paid off (Axiom in-house survey, 2026).
• 10,528 legal tech companies worldwide, 1,821 of them funded, 19 unicorns (Tracxn sector data, 2026).
• Harvey raised $550M at a $15.5B valuation and reports ARR above $400M (TechCrunch, 9 September 2026). Legora raised a $550M Series D at $5.55B, later extended to $600M (Bloomberg Law, 10 March 2026; Crunchbase News, 26 August 2026).
• 2,202 court and tribunal decisions involve AI-hallucinated material (Damien Charlotin's AI Hallucination Cases database, updated 19 October 2026).
• Oman's Personal Data Protection Law became fully enforceable on 5 February 2026, when the grace period under Ministerial Decision 6/2025 ended (CMS, 2026).

There is no shortage of maps of this market. Legaltech Hub lists more than a thousand generative AI products. Venture funds publish category grids. Review sites rank four hundred tools in a table whose pricing column mostly says "Enterprise". All of them tell you what exists.

A buyer needs to know what those products are made of. A directory says a vendor sells contract review. It will not say whether the vendor trained anything, wrote documentation, can be bought without a sales call, or invents case law when pushed. You find that out by signing up, paying and using the product, so that is what we did.

How did we test the legal AI market, and what is our conflict?

HAQQ sells legal AI, so we are a vendor inside the market we describe here. Read every number below knowing who counted it.

We fixed the counting rules before we started. The key one decides what counts as a wrapper: if you removed the underlying general model and the product stopped working entirely, with no retrieval layer, no routing, no legal data and no evaluation of its own, we counted it as a wrapper. That rule catches some very well funded companies, and parts of products we admire.

Every percentage here describes the vendors we tested. Nobody has tested the whole market. Where a number comes from somebody else, we name the source.

The method was slow and plain. 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 product until it broke, then study how it broke. Where a vendor would not sell without a call, try to get in anyway; how often that worked became a finding of its own.

What are legal AI products built from?

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.

Two terms, because they get used loosely. A prompt layer, usually called a wrapper, is an interface and a set of instructions on top of a general model such as GPT or Claude. A harness decides which model handles which step, retrieves the right sources first, calls tools and checks the result before you see it. The engineering sits in the harness, and it is what lets a product fail gracefully instead of making something up with confidence.

A prompt layer can be useful. A well built interface over a frontier model saves lawyers time, and we would say so about a competitor. The trouble starts with price and positioning. Sold at prompt-layer prices, it is a fair trade. Sold as proprietary legal intelligence at several hundred dollars per seat per month, next to a general assistant the firm already pays for, the trade stops making sense once anyone checks. Our teardown of legal AI pricing shows where that shows up in the price list.

Remove the general model underneath, and in two thirds of this market nothing survives.

Who is building legal AI for the person with the legal problem?

Very few. Of everything we tested, 2% address the end client. The rest is built for lawyers.

On the demand side, the World Justice Project estimates that 5.1 billion people have unmet justice needs: a legal problem and no meaningful route to resolving it. That demand cannot be met at the price a billable hour implies, which is exactly the constraint software should relax.

The venture cycle went to tools for the roughly 1.3 million licensed lawyers in the United States and their peers elsewhere. The reasons are sound. Lawyers have budgets, procurement and a clear willingness to pay. Consumers have none of those, and rules on who may give legal advice vary by jurisdiction and will not be settled by a startup.

We are in that 2%, so read this as a disclosed position. We shipped a consumer legal AI app, and it has been much harder than the professional product, for the reasons above. When 98% of the supply points at a sliver of the demand, the market has found one buyer it understands and stopped there.

Do law firms pick one legal AI vendor?

Mostly they stack several. 80% of the large firms we looked at either train their own models or run a router across several systems. They assemble four or more tools and decide, task by task, which one answers.

That undoes the model most vendor pitches assume. There is no bake-off with one winner. There is a stack, and the vendor's question becomes which slot it fills.

The ILTA 2026 Technology Survey of more than 500 firms found 94% of firms using or exploring AI, and the most widely installed AI tool was Microsoft Copilot, at 76% of firms. Copilot is a general productivity assistant that came bundled with software firms already owned, and it out-installs every product built for law. We are not publishing a ladder of the legal tools below it: a vendor ranking named competitors on someone else's survey data would be marketing with a citation. The structural point is enough. The most installed AI in law is the one that was already on the desktop.

Adoption figures also overstate progress. 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 are in-house teams while ILTA's 94% describes law firms, so they are two populations. Read together, nearly everyone has started, very few have finished, and most cannot measure the difference. The running tally of adoption research is in legal AI statistics 2026.

Sources: ILTA 2026 technology survey (500+ firms) for the first bar; Axiom 2026 in-house legal survey for the second and third. Different populations, so read the gap as a pattern rather than a single funnel.

What does the legal AI market map look like in 2026?

What do model-lab plugins mean for legal AI vendors?

Every general model provider now sells into legal directly. The latest is OpenAI's Astra for Law, launched on 17 September 2026 with 26 partner plugins.

Two of the largest independent legal AI companies shipped as plugins inside it. That is a rational move, since distribution tends to beat independence in the short run. It also says something about architecture. A plugin is a harness that somebody else hosts, bills for and controls the entry point to. If a vendor's edge is its harness and the harness now runs inside the lab's product, what the vendor defends is a slot in someone else's menu.

We wrote separately about what Astra does on real legal work. In short, the labs' edge is position: they are already where the user works, and in this market that has kept winning.

Is the legal tech market consolidating?

Yes, and the buyers are platforms folding in adjacent workflows. Clio closed its $1B acquisition of vLex on 10 November 2025, combining practice management with a research corpus. Wolters Kluwer bought Brightflag (about $500M) and Libra (about $105M) for spend management and drafting. Filevine bought Pincites to put drafting inside Microsoft Word, Doctrine bought Maite.ai to enter Spain, and Harvey bought Hexus, an onboarding and product-demo company, in January 2026.

The pace picked up through 2026. By late August, Legora had made at least five acquisitions this year and Harvey at least three (Crunchbase News, 26 August 2026), and Harvey announced Guardrails AI, an agent-security company, alongside its September round. Total funding eased slightly from the record: $4.6B in 2025 and more than $2.2B in 2026 by late August, per the same Crunchbase data.

The driver on the buyer side is tool sprawl. AllRize's 2025 survey found 55.3% of firms run between five and ten applications to manage their practice. Platform M&A is partly an attempt to sell fewer, better-connected systems to firms tired of stitching them together.

Do legal AI marketing claims match the products?

Often they do not. 5% of the vendors we tested have documentation we would call adequate: what the system does, what it does not do, what it was evaluated on and what happens when it fails. 80% oversell somewhere between the website and the product, meaning the capability on the landing page differs from the capability we found in the account.

Regulators treat that as a consumer-protection issue. In 2025 the US 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 lawyer. That order is a floor for where this goes.

Accuracy marketing is the sharpest case. Vendors have sold retrieval-grounded research tools on the promise that hallucination was eliminated. Stanford researchers who tested those tools concluded the providers' claims were overstated, and published measured error rates. We are reporting their finding and leaving the original claim where it was made.

Benchmarks are contested too. Vals writes its rubrics with practising lawyers, which is the right method, and vendors still dispute the results. Legal work is evaluated by judgement, and two competent lawyers can score the same answer differently, so a single accuracy number for a legal system is a simplification. We publish our own benchmark method so people can argue with it.

Do legal AI tools still hallucinate in 2026?

Every tool we tested invented something. Not on every answer, but each product in the set at least once produced law that does not exist, claimed work it had not done, or built a confident conclusion on a fact it had made up.

External measurements agree on the shape. Stanford's audit put Lexis+ AI above 17% and Westlaw's AI-Assisted Research above 33% on the versions tested. General models without retrieval measured between 58% and 88% on legal citation tasks in Dahl et al. (Journal of Legal Analysis). Retrieval grounding helps a lot and leaves a residue, and the errors that survive grounding are the dangerous ones, because they look like real citations attached to the wrong proposition.

Sources: Stanford RegLab's evaluation of commercial legal research tools for the first two bars, on the versions tested at the time; Dahl et al., Journal of Legal Analysis, for the general-model range.

Courts now count the consequences. Damien Charlotin's public database of decisions involving AI-hallucinated material held 1,598 cases on 9 June 2026, 2,046 on 21 September and 2,202 at its 19 October 2026 update, up from roughly 200 in mid-2025. Our hallucination cases tracker re-checks that count every month.

Source: the public AI Hallucination Cases database maintained by Damien Charlotin, read by HAQQ on 20 October 2026 (database updated 19 October). It counts only decisions a court wrote up, so treat the last bar as a floor.

A quieter failure sits next to it. 75% of the tools we tested write in their own hardcoded voice, whoever the firm, the client or the house style. None of them reproduce how a particular lawyer writes. In a profession where the phrasing of a clause is the product, a whole category skipped that.

Why do legal AI vendors hide their prices?

Only 10% of the vendors we tested are fully self-serve. Of twenty vendors we checked for a published price, 14 publish none at all. The category default is a contact form.

First two bars from the twenty vendors we checked for a published price. The third is the share of the full test set we were able to get into and use unaided.

Yet 75% of them were easy to get into without speaking to anyone: trials, partner routes, self-service tiers behind a sales page, and in several cases a checkout the marketing site did not link to.

A sales-gated price usually means either enterprise complexity or price discrimination. When the same product can be bought in four minutes by anyone who looks a little harder, it was price discrimination. Who publishes what, and what it costs past the gate, is in legal AI pricing in 2026.

How is legal AI regulated and built in the Middle East?

In the Middle East and North Africa the largest buyer of legal technology is often the state, and the state tends to build. Two moves in 2026 show the direction. On 22 January 2026, at the World Economic Forum in Davos, the UAE Cabinet's General Secretariat published its first whitepaper on regulatory intelligence, with Presight (a G42 company) and PwC. It proposes a Unified Regulatory Digital Twin, a live model of the UAE's regulatory system, and a Sovereign Governance-in-the-Loop (SGiL) framework in which AI assists legislators and humans make the decisions.

On 28 July 2026 the Abu Dhabi Judicial Department, with the Department of Government Enablement, announced an AI Judicial Platform to support judicial decision-making under full human supervision, with a first phase from September 2026 and a phased rollout over 18 months (WAM, via Khaleej Times). Saudi Arabia's Ministry of Justice runs Najiz, which offers about 160 electronic justice services (Saudipedia, October 2025).

These are infrastructure programmes. A vendor can supply a component; it does not win the platform. At firm level the same instinct repeats: several MENA practices build their own AI and then license it to other firms, a route few London or New York firms take.

Data protection law sets the floor for any legal AI that touches client data in the region:

For vendors selling into the Gulf, regulator posture, hosting location and Arabic-language quality decide who gets shortlisted. Regional vendors compete by being closer to the language, the law and the regulator, since outspending global incumbents on models is not an option. HAQQ's own answer is EU, US or Middle East hosting, private cloud or on-premise, and full right-to-left Arabic. We cover the regional market in legal tech in the Middle East and legal AI in Saudi Arabia.

One consequence for anyone mapping the market from outside: a vendor map drawn from Western funding data misses the largest legal AI deployments in MENA, because they never appear as a funding round.

General legal information, not legal advice; check the current text of each law with the regulator or local counsel before relying on it.

How many legal AI companies are there?

Tracxn's sector data counts 10,528 legal tech companies globally, 1,821 of them funded, and 19 unicorns (2026). Set against our composition data, that is a lot of companies for a market where 8% of the products we tested contain a model anybody trained.

Source: Tracxn sector data, 2026. Bars are proportional to the first: unicorns are 0.2% of the field, and the bar is drawn at a visible minimum rather than to scale.

Two vendors dominate the funding story. In March 2026, Harvey ($200M at $11B) and Legora ($550M at $5.55B) raised $750M between them within 15 days. Since then Harvey has raised $550M at $15.5B and reports ARR above $400M and more than 3,000 customer organizations (TechCrunch, 9 September 2026). Revenue figures are company-reported and unaudited.

Scale follows funding. 90% of the funded vendors we tested have more than 200 staff, and only 20% publish an ARR figure of any kind. A category this large that discloses this little leaves buyers comparing vendors on the demo.

Landing pages show what the category thinks converts. Four elements recur: security, a prompt library, testimonials and a Labs section claiming deep AI research. Security leads, and the timing makes sense: generative AI has just entered the legal industry's tracked security-risk list at number two. Buyers are worried, and vendors have read that worry as the first objection to clear. We publish our own security posture for the same reason, and we are part of the pattern.

Is legal AI better than a lawyer yet?

On some tasks it is. In the Vals Legal AI Report, the strongest system on document question answering scored 94.8%, 24.7 points above the human lawyer baseline in the same study. On redlining, the same system lost to the human control group. Both results come from one product in one evaluation.

That is the most useful single fact in this market: capability belongs to the pairing of tool and task. Point a system at finding what a 400-page document says and it beats a tired professional. Point it at choosing which of two acceptable clause wordings protects your client better, and it produces something that reads like an answer and does not hold up.

Most failed legal AI pilots we have seen were aimed badly. A firm buys a system, points it at the hardest judgement-heavy work in the building because that is where the pain is, watches it underperform and decides the category is immature. The task-by-task breakdown is more useful than any overall score.

What should a buyer take from this?

Our own lesson is about distribution, and it cost us to learn it. Our first serious product asked lawyers to open a new tab and work there. It was a better tool than what they had, and it lost to habit. Products that ask lawyers to leave Word, Outlook or the document management system keep losing the same way, which is why a general assistant that was already installed out-installs every purpose-built legal product.

Model quality is converging and is mostly bought from four companies. Legal knowledge is written down. What stays scarce is the right to sit where the work already happens, plus the discipline to say which tasks a system is good at. When you evaluate a vendor, ask four questions: what model it runs on and what survives if you remove it; what it was evaluated on, with the documentation; where it runs inside your existing tools; and what it costs without a sales call.

We are one of the 10,528 companies and the findings above apply to us too. What we can do is publish the counting, including the parts that are uncomfortable for us, so the next person mapping this market starts from more than a category grid.

Key takeaways

• Two thirds of the market is a prompt layer. Of the vendors we tested, 65% stop working entirely without the general model underneath, and 3% train on legal data.
• Large firms run stacks. 80% of the large firms we looked at run several systems or their own router.
• The most installed AI in law firms is general-purpose. ILTA puts Microsoft Copilot at 76% of firms, because it was already on the desktop.
• Adoption is wide and completion is rare. 94% of firms are using or exploring AI; 7% of in-house teams are past piloting.
• Hallucination is common to every tool we tested. External audits put grounded legal tools at 17% to 33%, and the court-decision count reached 2,202 on 19 October 2026.
• Consolidation is the 2026 pattern. Clio and vLex, Wolters Kluwer and Brightflag, and at least eight 2026 acquisitions by Legora and Harvey alone.
• In MENA the state builds. The UAE regulatory digital twin, the Abu Dhabi AI Judicial Platform and Najiz are infrastructure, and Oman's PDPL has been fully enforceable since 5 February 2026.
• Aim the tool at the right task. The same system beat lawyers by 24.7 points on document Q&A and lost to them on redlining.

Sources and further reading

• Legal AI Statistics 2026
• Legal AI Pricing in 2026: every published price, and every vendor that publishes none
• AI hallucination cases: the sanctions tracker
• Legal tech in the Middle East
• 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
• Crunchbase News: legal tech funding down slightly from all-time high (26 Aug 2026)
• Legal IT Insider: Clio completes $1bn vLex acquisition (10 Nov 2025)
• UAE Cabinet: first whitepaper on regulatory intelligence, Davos 2026
• Khaleej Times: Abu Dhabi AI judicial platform
• CMS: Oman Personal Data Protection Law entering the enforcement phase
• Al Tamimi: Egypt issues Executive Regulations of the Personal Data Protection Law
• Saudipedia: what is the Najiz e-platform