Home / Articles / The legal AI services map: 44 companies scored on who does the work, who stands behind it, and whether an agent can hire them, with the evidence in a public vault

The legal AI services map: 44 companies scored on who does the work, who stands behind it, and whether an agent can hire them, with the evidence in a public vault

By Dinis Cruz · 2026-10-28 · article v1.0.0 · newsroom v0.13.122 · legalailaw-firmsagentsmcpstartupsregulationvaultsevidencearticle

Abstract: Which companies sell legal work done by AI, by AI and lawyers together, or by lawyers with AI help, and can an agent hire any of them? I asked this site's agent to map them in the style of an analyst's quadrant, but with a method fixed before any company was scored, a formula anyone can re-run, and the evidence for every score kept in a public vault. Three scores, from public evidence only. Automation: how much of the work AI does. Accountability: whether a regulated firm and a named lawyer stand behind the output. Agentic readiness: an API, an MCP server, llms.txt, whether AI crawlers are allowed, published prices. Of 44 companies, 24 services and one reference are on the map; tools sold to lawyers are scored apart. Two services sit where AI does the work and a lawyer stands behind it: Garfield.Law, the UK's first SRA-authorised AI law firm, and Crosby. Three expose an agent interface: Soxton, whose MCP server lets an agent file an incorporation, LegalZoom and Rocket Lawyer. None of them has all three, which is where our own proposal, legal on demand for startups, sits as a target, not a competitor.

In short

• The question: who sells legal work done with AI, who stands behind it, and can an agent hire them?
• The method was fixed before scoring: three scores, fourteen indicators, public evidence only, no evidence scores zero.
• 44 companies researched: 24 services and one reference mapped; 10 tools for lawyers scored apart; 9 listed, not mapped.
• Two services are where AI does the work and a lawyer signs: Garfield.Law and Crosby.
• Three services can be used by an agent: Soxton, LegalZoom and Rocket Lawyer, through MCP servers or APIs.
• No service has all three, which is the gap our proposal, legal on demand for startups, is meant to fill.
• The UK is ahead on regulation: the SRA has authorised two firms where software does the work.
• The evidence is in a public vault: screenshots, robots.txt, llms.txt, every score with its URL and quote.

Why a map

In legal on demand for startups I proposed a service where agent teams do the reading and drafting and a regulated firm's lawyers judge and sign. Before building it, I wanted to see who is already doing some part of it, and whether any of them could be used by an agent rather than a person with a browser.

Analyst quadrants are useful and hard to check. So this one is built the other way round. The method, the criteria and the formula were written down and dated before any company was scored. Every indicator records a value, a URL and a quote. No evidence means the lowest value, never a guess. And everything is in a vault anyone can open, with each company invited to correct its entry.

The method

Three scores, each from 0 to 100, from public evidence only: the company's own pages, regulators' registers and reputable reporting.

• Automation (A): how much of the work AI does. Who produces the work (AI only, AI with lawyer review, lawyers with AI tools, lawyers only), weighted 40%; the turnaround claimed; whether it is self-serve from start to finish; and whether the price follows the task or the hour.
• Accountability (R): who stands behind the output. A regulated provider, 35%; a named lawyer reviewing or signing as standard, 30%; professional indemnity or malpractice cover stated, 15%; a published evaluation or accuracy measure, 20%. Minus 15 if a regulator or a court has acted against the service's claims.
• Agentic readiness (G): can an agent use it. A public API, 35%; an MCP server or agent integration, 25%; a real llms.txt, 15%; robots.txt that lets AI crawlers in, 15%; published prices, 10%.

Tools sold to lawyers (Harvey, Legora, CoCounsel and others) are scored for automation and agentic readiness only and shown apart, because the lawyer delivers the work, not the tool. Companies with no AI in the delivery, and ones that have closed or collapsed, are listed, not mapped. The full method, with every rule of application, is METHOD.md in the vault.

What the map shows

Two services are in the top right, where AI does the work and a lawyer stands behind it:

• Garfield.Law, which describes itself in its llms.txt as "the UK's first SRA-regulated AI law firm". Debt-recovery letters and small claims, priced per task (a letter before action is £7.50), with "Document creation time: 5 minutes". The SRA authorised it in May 2025, with the client approving each step and named solicitors accountable.
• Crosby, "a human-AI hybrid law firm" in New York, with a "median review time of 58 minutes" and "Fixed pricing by document, not by the hour". The prices themselves are not published, so it scores lower on automation than it might.

Above them on accountability, and further left, are lawyers with AI help. General Legal scores 100 for accountability: it says it "is a US law firm, not a software product", its lawyers "carry malpractice insurance", and it has published its own evaluation of the gap between what models produce and what a lawyer would send. Lawhive owns an SRA-regulated firm and has bought another. Carta Law (formerly Avantia), Eudia Counsel and KPMG Law US are regulated firms under the UK or Arizona rules for non-lawyer ownership.

Below the line are automation without a signature, and the platforms. LawFairy, the SRA's second technology-only firm, runs a rules engine with no generative AI, scores 100 for automation, and states no lawyer sign-off. Rocket Lawyer, LegalZoom and SeedLegals are platforms with lawyers available, not lawyers signing as standard. DoNotPay is at the bottom: the FTC's final order of 2025 imposed $193,000 of relief and bars claims about its "AI lawyer" that it cannot back.

Can an agent hire them?

This was the question I most wanted answered, and the answer is: very few, and none of the AI-native law firms.

• Soxton runs "a remote MCP server" so that "an authenticated agent can request all three on a human's behalf, including preparing, signing, and submitting an incorporation". Its llms.txt publishes the prices: "One-time filing fee: $400. Registered agent fee: $300 per year." It is not a law firm itself; its attorneys work through one.
• LegalZoom declares an MCP endpoint in its llms.txt, with tools including request_attorney_review.
• Rocket Lawyer calls itself "an AI-native legal membership", and its llms.txt says partners can build through its APIs and its MCP server.

Among the tools for lawyers, Wordsmith, Harvey and Ironclad publish APIs, and several have MCP servers. On the open web, about 19 of the 44 sites have a real llms.txt. Most do not mention AI crawlers in robots.txt at all. Avantia Law's old site blocks all the main ones, though Carta Law now lives on carta.com, which does not; Genie AI explicitly lets them in.

The rules are different in each country

The UK is ahead. The SRA authorised Garfield.Law in May 2025 as the first firm where AI does the work, and LawFairy in February 2026 as the second; the International Bar Association noted that no other jurisdiction had followed. In the US, the rules against unlicensed practice of law are the barrier: in Upsolve v. James, in September 2025, the appeals court upheld New York's rules. Arizona's licensing route for firms with non-lawyer owners and Utah's regulatory sandbox, narrowed in 2025, are how firms combining AI and lawyers get outside ownership. In Germany, the 2021 Smartlaw ruling held that a contract generator is not a prohibited legal service.

Walking back from the North Star

The orange mark is our own proposal, scored by the same rules as a design, not a company: agent teams do the work, a regulated firm's lawyers sign it with indemnity cover, every output is sourced, and an agent can hire it through an API and an MCP server. It is drawn as a target and never ranked.

Walking back from it, the gaps are clear:

1. Nobody combines a regulated signature with an agent interface. Garfield.Law and Crosby have the signature; Soxton, LegalZoom and Rocket Lawyer have the interface.
2. Nobody publishes an independent accuracy audit. The best evidence is General Legal's own evaluation and Lawhive's company-reported exam score. Stanford found that leading legal research tools hallucinated in 17 to 33% of queries; a service that wants to stand out can publish its error rate.
3. Prices per task are rare. Garfield.Law, General Legal, Soxton and SeedLegals publish them; most firms do not, which also means an agent cannot compare them.
4. The UK is the place to start a firm where AI does the work, because the regulator has already said yes twice.

The evidence vault

Everything is in a public vault: METHOD.md, written before scoring; scores.csv and scores.json, with every indicator and its reason; the research record for every company, with each quote marked as verbatim, extracted or paraphrased from a search result; and, for every company, screenshots of the pages the claims come from, robots.txt and llms.txt as fetched, and a manifest of SHA-256 hashes with the capture time. The captures are of public marketing pages, kept for review and correction; copyright stays with their owners.

Open it in the vault browser, or clone it with pip install sgit-ai and sgit clone sgit_public_read_a0a1745807e255241d391b63d3f0e5b8bd96f942494ef91435f19c02a5436233:m3grufkt legal-ai-map. If a company thinks its score is wrong, the method says how to show it.

What this does not claim

• It measures what is published, not quality. A firm that does great work and says little about it scores low.
• No evidence scores zero, which favours companies that publish; that is the point, but it is a choice.
• Some research quotes are fragments or search summaries, and are marked as such in the vault; the article quotes only text fetched verbatim from the companies' own files.
• Funding figures and several claims are unverified and are not used in any score.
• The North Star is ours. I have an interest in the gap this map shows.

Where this comes from

Each company's own pages, llms.txt and robots.txt, fetched on 28 October 2026 (URLs in the vault); the SRA on Garfield.Law; the IBA on the SRA's approach; the FTC's final order on DoNotPay; Stanford RegLab and HAI on hallucination in legal research tools (2024). Research by this site's agent, under my direction.

Peer review Not reviewed yet. Review it

Want the next issue by email. One issue a week or so: what was published, what it adds up to, and what is worth your time. Subscribe to the SGit Newsroom →

← All articles