25 legal AIs. Only three show a price.

Legal work has crossed the adoption line faster than its measurement. The Thomson Reuters Institute's 2026 AI in Professional Services Report, drawn from more than 1,500 respondents in 27 countries across law, tax, accounting and risk, found that 40% of organisations now use generative AI, up from 22% a year earlier. Fifteen per cent have adopted some form of agentic AI. Only 18% say they track the return on their AI tools.

For a profession that bills by evidence, that last number is the uncomfortable one. Clients ask whether AI makes legal work cheaper, faster or better. Most firms and legal departments cannot yet answer with data.

ETHORITY reviewed 25 legal and compliance AI offerings for the Good AI Prism in October 2026, reading each provider's pricing, security and legal pages. Here is what the market puts on the record, and what buyers should demand before relying on it.

The market: research, review, contracts, compliance

The Legal & Compliance area map covers four jobs. Legal research and drafting assistants: Harvey, Legora, CoCounsel, Lexis+ with Protégé, Libra by Wolters Kluwer, Otto Schmidt Answers, Noxtua and DeepJudge. Document review and e-discovery: Relativity aiR for Review and Everlaw. Contract review and lifecycle management: Juro, Legartis, LegalOn, Spellbook, Robin AI, Definely, Luminance, Ironclad, Icertis Vera, Sirion and LEGALFLY. Regulatory compliance and workflow: Norm Ai, Corlytics, Bryter and Clio Work.

German-language legal content has its own cluster. Libra builds on Wolters Kluwer content, Otto Schmidt Answers on the publisher's own commentary, and Noxtua runs on European clouds. For German law, the licensed content behind a tool matters as much as the model.

3 of 25 legal AI offerings publish a price. Libra, Legartis and Otto Schmidt Answers do; the other 22 quote on request.

Prices: per user, per year, mostly private

The three published prices share a unit: per user. Libra lists its Starter plan at €900 per user per year, with Professional at €2,000 and Team at €2,400, and charges partner content separately. Legartis lists Professional at €2,500 per user per year. Otto Schmidt Answers costs €79 per user per month at list price. The large research and contract platforms quote.

Per-user pricing is easy to budget and hard to judge, because the value depends on how much work moves. A practice group that runs every first-draft contract review through a tool gets a different return from one that uses it twice a month. Measure review time and verified-citation rate per matter type before and after, which is exactly what the Global Market Study recommends for this function.

Scenario. For a team of ten lawyers, list prices run from €9,000 a year (Libra Starter at €900 per user) to €9,480 (Otto Schmidt Answers at €79 per user a month) and €25,000 (Legartis Professional at €2,500). The price gap matters less than the review time per matter and the verified-citation rate each tool reaches on your own files.

Accuracy: the number that decides liability

The risk in legal AI is not that a tool fails to answer. It is that it answers confidently and wrongly. Studies summarised by LLRX report hallucination rates of 17% to 33% in legal research tools. Those are tools built for lawyers, with retrieval over legal sources, not general chatbots.

The response is procedural, not technical. Require source citations for every statement of law. Classify the evidence behind each citation. Keep a person who signs for every output that reaches a client or a court. And measure the verified-citation rate per tool and per matter type, so that trust follows evidence rather than demos.

17–33% hallucination rates reported for legal research tools built for lawyers (studies summarised by LLRX, Feb 2026).

Where client data goes, and what it trains

Client confidentiality makes hosting and training terms decisive, and legal providers document them more often than most. Eleven of 25 document an EU hosting option. Harvey lets customers keep data in the EU or Switzerland, the US or Australia. Noxtua runs its own AI infrastructure on European clouds including IONOS and Deutsche Telekom. Bryter keeps EU and EEA customer data on servers in Frankfurt. Libra states that user data is stored in Germany.

On training, 16 of 25 document that customer data is not used to train models, or only with opt-in. One, Legartis, offers an opt-out from model training on its Team and Enterprise plans, which makes the default worth checking on smaller plans. For a law firm, a training clause is a professional-secrecy question, not a feature comparison.

Agents arrive in the contract workflow

The 15% of organisations already using agentic AI are a signal of where contract tools are heading. Several providers in the map now market agents that redline against a playbook, route a contract for approval or answer a business user's question about a clause without a lawyer in the loop. That changes the governance question from "is the answer right?" to "who authorised the action?"

A sensible boundary has three lines. Agents may propose redlines, but a named lawyer accepts them. Agents may answer from an approved playbook, but questions outside it go to a person. Agents may route and file, but they do not sign, send to the counterparty or change a template without approval. Written down, those three lines are short enough to fit into a policy and concrete enough to test in a pilot.

The compliance calendar is now fixed

For compliance teams, 2026 turned AI principles into dates. Under the AI Act, as amended by Regulation (EU) 2026/1744, the calendar reads as follows. Prohibited practices and the AI literacy duty have applied since 2 Feb 2025. General-purpose model obligations since 2 Aug 2025. Article 50 transparency duties since 2 Aug 2026. Machine-readable marking for generative systems already on the market, and a new prohibition on non-consensual intimate imagery, from 2 Dec 2026. High-risk duties for Annex III uses, including employment, credit, education and the administration of justice, from 2 Dec 2027. High-risk duties for AI in regulated products from 2 Aug 2028.

Boards have noticed. In the Allianz Risk Barometer 2026, 32% of respondents named AI as a top business risk, now in second place. Our article on why AI Act readiness is not operational control explains how to connect each obligation to a system, a control and a piece of evidence.

What good governance looks like in a legal team

Legal and compliance teams usually own AI governance for the whole organisation, and they use AI themselves. Both roles benefit from the same artefacts: an inventory of AI systems with owners and intended uses, a register of provider terms, a review rule for outputs that leave the building, and a record of who was trained on what. The difference between a policy and governance is whether these artefacts are current when someone asks.

What to do Monday

1. Set the accuracy rule. No AI output reaches a client or a court without verified citations and a named reviewer.
2. Measure two numbers. Verified-citation rate and review time per matter type, before and after the tool.
3. Read the training clause as a secrecy clause. Default setting, plan level and opt-out route, in writing.
4. Put the AI Act dates in the compliance calendar. Next stops: 2 Dec 2026 for marking, 2 Dec 2027 for Annex III.
5. Start the inventory. List every AI system, its owner and its intended use; it is the first artefact every regulator asks for.

The HR & Recruiting topic area covers the employment uses that carry high-risk duties from Dec 2027.