Last reviewed: 16 July 2026.
I have no affiliate relationship with the tools reviewed on this page. Vendor statistics are identified as such. Where a vendor does not publish pricing, any figure given is a third-party estimate rather than a confirmed rate card.
Why this guide exists
I have been using AI in my practice almost every working day since 2023. I draft with it, review with it, and write to clients with it, and I’ve written a book on AI for Lawyers. I have also watched it fail: confidently, fluently, and in ways that would have been embarrassing at best and negligent at worst had I not checked.
Most guides to “the best AI tools for legal document review” are written by vendors or by content marketers who have never had to explain to a client why a clause was missed. They rank tools by feature count. They repeat vendor accuracy claims as findings rather than marketing. And several of the guides currently ranking for this term omit the tools that solicitors overwhelmingly actually use, because those tools are not paying anyone a referral fee.
This guide covers the whole field: the general-purpose models sitting in your browser right now, the enterprise platforms that have raised billions, the research platforms grounded in licensed databases, and the practice management systems with AI layered in. It assesses them on public information, including whether the vendor will still exist in two years. And it sets out the framework I use to decide whether a tool goes anywhere near a client matter.
Two cases come first. They matter more than any feature comparison below.
Ayinde: you must check the output
In R (Ayinde) v London Borough of Haringey and Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin), the Divisional Court, sitting under its Hamid jurisdiction, dealt with two matters in which fictitious authorities had been put before the court. In Ayinde, five fake cases appeared in the grounds for judicial review. In Al-Haroun, eighteen non-existent authorities appeared in witness statements, and the solicitor accepted he had relied on his lay client’s research without independently verifying it.
The court was blunt about that. A lawyer is not entitled to rely on their lay client for the accuracy of citations, and it is the lawyer’s professional responsibility to ensure the accuracy of such material. Both matters were referred to the regulators. The court declined to bring contempt proceedings but made clear that its restraint set no precedent.
Strictly, Ayinde decided nothing new. It restated an existing duty. The problem is that the duty was being forgotten, and the courts have continued to encounter the same failure.
Munir: where you put the document matters too
This is the case that should change how you think about every tool on this page, and it has had a fraction of Ayinde’s coverage.
In UK and R (on the application of Munir) v Secretary of State for the Home Department [2026] UKUT 81 (IAC), the Upper Tribunal, again under the Hamid jurisdiction, dealt with two claims in which false citations reached the tribunal. One adviser admitted using ChatGPT. He had also, separately, put draft client emails into ChatGPT to improve them, and uploaded Home Office decision letters to summarise them for clients.
The tribunal made two findings every solicitor using AI needs to hold in mind.
On supervision: a solicitor who delegates work remains responsible for supervising it and for ensuring its accuracy. Supervisors must ensure junior fee earners are aware of the dangers of using non-specialist AI for research and drafting, and must carry out appropriate checks. A supervisor who lets hallucinated citations reach the court in a junior’s work is, the tribunal suggested, likely to be more culpable than a lawyer who fails to check their own.
On confidentiality: uploading confidential documents into what the tribunal called ‘open-source AI tools’ such as ChatGPT is to place that information on the internet in the public domain, and thus to breach client confidentiality and waive legal privilege. Such conduct might itself warrant referral to the SRA and should in any event be referred to the Information Commissioner’s Office.
It is important to clarify three points because that passage is being quoted loosely across the profession.
First, “open-source” is being used colloquially (technically incorrect!). Read it as “public” or “consumer”. The tribunal expressly contrasted these with closed enterprise systems, giving private instances of Microsoft Copilot as its example, and said those do not place information in the public domain.
Second, this was an observation in a Hamid decision, not a binding ruling in a privilege dispute. The High Court is not bound by it. A future court might distinguish between publishing information on the internet and entering it into a contractual, closed-processing arrangement. But it is the first English judgment to draw the link this explicitly, and it is highly persuasive.
Third, the tribunal was not hostile to AI. It said that the use of legal AI by properly trained professionals is a step forward in legal practice and of enormous benefit in focused legal research and large disclosure exercises. The problem is the careless use of confidential material without supervision.
The consequence is severe. Privilege depends on confidentiality. Once confidentiality goes, privilege may go with it. On the tribunal’s analysis, privilege is lost at the point of upload; the user’s intention would not restore the confidentiality on which privilege depends.
Set against that, an April 2026 Censuswide survey commissioned by Access Legal found that 59% of fee earners admitted using unapproved AI applications for client work, including 71% of paralegals and 57% of solicitors, while 68% of firm leaders surveyed believed they had full visibility of AI use and faced zero risk. The sample was 200 UK legal professionals: 100 firm leaders and 100 practitioners. It is a small, vendor-commissioned survey, not research by the Law Society, but the governance gap it describes is credible and serious.
The regulatory floor as solicitors using AI
The SRA has not written AI-specific rules and has said it does not need to. Its Risk Outlook report on AI in the legal market and its compliance tips on AI and technology make the same point: your existing duties already cover it. You remain personally responsible for every piece of work you produce. The SRA expects it to be clear to clients where they are interfacing with AI, expects confidentiality to hold, and expects the COLP to own regulatory compliance when new technology is introduced.
So the question is never “is this tool any good”. It is “can I supervise this tool, and can I put a client document into it”.
What the independent evidence actually shows
Before we look at the tools, let’s look at the evidence. Almost every accuracy claim in this sector is vendor-published. There are three pieces of independent or self-undermining evidence worth knowing, and together they tell a more useful story than any product page.
One. The grounded research platforms hallucinate. LexisNexis marketed “100% hallucination-free linked legal citations”. Thomson Reuters said its tools avoid hallucinations by relying on trusted content. Researchers at Stanford RegLab and HAI tested those claims across more than 200 legal queries in the first preregistered empirical evaluation of these products, later peer-reviewed in the Journal of Empirical Legal Studies. The [paper is public](https://arxiv.org/pdf/2405.20362). Lexis+ AI was the best performer, answering 65% of queries accurately and hallucinating on more than 17%. Westlaw’s AI-Assisted Research was accurate 42% of the time and hallucinated on around a third. GPT-4, ungrounded, hallucinated on 43%. The researchers concluded that providers’ claims were overstated.
Two caveats, in fairness to the vendors. The study tested US-jurisdiction products as they stood in May 2024, and both platforms have been substantially rebuilt since. It would be wrong to quote 17% and 33% as though they were today’s numbers. But the structural point survives: retrieval-augmented generation reduces hallucination, it does not eliminate it, and the errors it leaves behind are not always fabricated cases. They are real cases mischaracterised, or overruled authority cited as good law, which is harder to catch and more dangerous.
Two. Lawyers still beat AI at redlining. The Vals Legal AI Report, run with Legaltech Hub, is the first systematic independent benchmark of legal AI tools against a lawyer control group, using tasks derived from Am Law 100 firms. Harvey, CoCounsel, vLex’s Vincent and Vecflow’s Oliver took part. LexisNexis initially participated, then withdrew from most sections.
The results are more interesting than either the sceptics or the vendors want. AI tools collectively beat the lawyer baseline on four tasks: data extraction, document Q&A, summarisation and chronology generation. Harvey scored 94.8% on document Q&A against a 70.1% lawyer baseline, which is not a marginal difference. But the lawyer baseline beat every AI tool on redlining, at 79.7%. Extraction, document Q&A and retrieval were where the tested systems performed most strongly. Judgment about what a clause should say was materially harder.
Three. The specialist premium is not buying reasoning. A follow-up Vals benchmark on legal research compared three legal AI systems against ChatGPT and a lawyer baseline across 210 questions. Lawyers averaged 71%. ChatGPT scored 80%. The legal-specific tools, grouped, also scored 80%. On five of the nine question types the generalist model was on average more accurate than the legal-specific products.
And then there is Harvey’s own admission, which I will come to below.
How to evaluate any AI tool in this category for use by lawyers in the UK
Tools change every quarter. These questions shouldn’t need to change that often if at all.
Where does the document go, and what happens to it there? Ask for the data processing agreement. Ask whether the underlying model providers are named in it, and whether data retention is contractually agreed rather than asserted on a marketing page. Ask where the data is resident. After Munir, this is not a procurement nicety. It is the difference between a tool you can use on a client matter and one you cannot.
Can it show its working, and at what granularity? When it says a clause is unusual, can it point to the clause? Character-level citation, where you click an assertion and land on the exact source text, is the standard to ask for. Source-level citation, where it tells you which document it came from, is weaker and leaves you re-reading. A tool that can be audited is a tool you can supervise.
Is it deterministic enough? Ask whether the same document, uploaded twice, produces the same analysis. General-purpose models do not. Structured platforms come closer. Inconsistency means you cannot use the output as a checklist, only as a prompt to think.
What is it actually doing? Extraction (finding the assignment clause across 400 leases), analysis (telling you the assignment clause is unusual), and drafting (proposing a replacement) are three different products. The Vals data says the first is where AI is strongest and the last is where it is weakest. Most tools do one well and market all three.
Does it understand English law and English drafting? Subject to contract. Without prejudice. Limitation clause structure under UCTA. A tool trained predominantly on US commercial paper will produce output that looks right and reads wrong.
Will the vendor exist in two years? Not a rude question. A due diligence question. See Robin AI below.
What is the total cost, honestly? Licence, plus playbook configuration, plus document management system integration, plus training, plus the hours before it starts paying for itself, plus the renewal uplift. Ask for a contractual cap before you sign; an uncapped increase can materially change the arithmetic at renewal.
Category one: general-purpose models
This is what most solicitors are actually using, so it goes first.
ChatGPT, Claude, Microsoft Copilot
The distinction that matters is not between the three brands. It is between consumer tiers and enterprise tiers, and Munir is the reason.
Consumer tiers. Free and personal-subscription accounts operate under consumer terms rather than a firm-negotiated processing arrangement. Some providers allow a user to disable model training, but a training toggle is not a substitute for a data processing agreement, agreed retention, organisational controls or a documented privilege assessment. This is the category the Upper Tribunal was describing. Do not put confidential client documents, client names or matter facts into an unapproved personal account. Not the decision letter. Not the draft advice. Not the witness statement. Not “just to improve the wording of this email”. That last one is exactly what the adviser in Munir did.
Enterprise and business tiers. Materially different. A data processing agreement, contractual no-training commitments, tenant isolation, admin controls, audit logs, and in some cases zero data retention and UK or EU data residency. Microsoft 365 Copilot operating inside your own tenancy is the example the tribunal itself gave of a closed system. Claude and ChatGPT both have enterprise offerings with comparable contractual protections.
That does not settle the privilege position. It makes it defensible. Reduced risk is not no risk, and there is no English authority yet on enterprise tools. What it does mean is that you have a contractual answer when the ICO or the SRA asks where the document went.
What they are good at. Analysis, drafting, and reading long documents. Copilot’s advantage is that it inherits your existing Microsoft 365 permissions, retention and eDiscovery controls, which for a firm already on M365 removes most of the procurement question. It is not a legal research tool: no legal database, no citation capability. Claude and ChatGPT are stronger on sustained analytical reasoning over a long contract.
What they are bad at. Authority. A general-purpose model with no retrieval will invent case law that sounds exactly like real case law. Never ask an ungrounded model for a case. Ask it to reason about a document you have given it, and check what it says against the document.
The verdict. Commercial terms, structured instructions, human verification of every assertion. That combination is legitimate and effective, and the Vals research data suggests it is not obviously worse than the specialist alternatives at the reasoning itself. Putting confidential client material into an unapproved consumer account creates a serious risk of breaching confidentiality, losing privilege and triggering regulatory or data-protection reporting obligations.
Claude for the legal industry
Anthropic made the most consequential move into legal AI since this guide was drafted. In May 2026 it released what is generally being called Claude for Legal: more than twenty connectors to legal systems and twelve open-source plugins for particular kinds of legal work.
It is important to understand what this is. Claude for Legal is not a separate legal model, and it does not come with its own authoritative legal database. It is an orchestration layer around Claude: reusable instructions, access to external systems and integrations with the software in which lawyers already work.
The plugins cover commercial, corporate, employment, privacy, product, regulatory, AI governance, intellectual property and litigation work, as well as legal clinics and law students. Each begins with a setup process intended to capture the organisation’s playbook, escalation thresholds, risk appetite and house style. The commercial plugin, for example, reviews agreements against your stated positions and prepares business-facing explanations. The litigation plugin handles matter intake, chronologies, legal holds, privilege logs and first drafts.
The more significant development is the connector ecosystem. Claude can now connect to iManage, NetDocuments, DocuSign, Ironclad, Datasite, Everlaw, Relativity and other systems. Research can be grounded through services including Thomson Reuters CoCounsel, Midpage and CourtListener, while Harvey itself can be accessed through a Claude connector. Claude also works within Word and Outlook, carrying context from a redline into the accompanying email or closing checklist.
That architecture addresses two of the weaknesses of a general-purpose model. Documents can remain within governed systems rather than being downloaded and pasted into a browser, and legal research can be routed through an authoritative external source. Existing permissions are generally respected by the connectors.
But do not confuse a plugin with verification. Most of Anthropic’s legal plugins are sophisticated instruction sets. They can impose a better process, but they do not make the underlying model legally authoritative. A contract-review plugin can tell Claude to identify every change and explain its reasoning; it cannot guarantee that Claude has found every problem. Research is only grounded where an appropriate research connector is enabled, and the quality of the citation depends on that connected service.
The confidentiality analysis also remains configuration-specific. The fact that the plugins are available to paid Claude users does not make every Claude account suitable for client work. A firm must still establish which commercial terms apply, what retention period has been configured, which connectors send information to third parties, and whether those third parties are covered by its contractual and regulatory assessment. Adding connectors can reduce insecure copying and pasting while increasing the number of processors through which matter information passes.
My read. This is unusually important because it makes the specialist platform’s main advantage available much closer to the foundation model. Anthropic is effectively giving firms a toolkit for building their own Harvey-like workflow layer, using open plugins that can be inspected and adapted rather than a playbook locked inside one vendor.
For a small firm with strong internal processes, that is compelling. It is also strikingly close to the approach I describe below: start with a capable model, encode the solicitor’s own standards as structured instructions, connect it only to trusted sources, and keep the solicitor responsible for the final judgment.
It does not replace Harvey, Legora or Luminance where hundreds of users need controlled, repeatable workflows over thousands of documents. It does, however, make the gap much narrower for an individual solicitor or small team. [Anthropic’s announcement and complete connector list are here: https://claude.com/blog/claude-for-the-legal-industry), and the legal plugins themselves are publicly available here: https://github.com/anthropics/claude-for-legal).
OpenAI’s legal vertical: not yet a product
OpenAI is heading in the same direction, but buyers should distinguish an announced strategy from a tool they can purchase.
In June 2026, Jason Boehmig, the lawyer who founded contract-management platform Ironclad, joined OpenAI to lead product development for its “legal vertical”. That is a serious appointment and a clear indication that OpenAI intends to build specifically for law firms, in-house teams and the wider legal technology ecosystem.
There is not, however, a released product called ChatGPT for Legal at the time of writing. OpenAI has announced no legal-specific feature set, authoritative research database, price, launch date or separate contractual terms. It would therefore be misleading to rank ChatGPT for Legal alongside Harvey, Legora or Claude’s released legal plugins.
What exists today is ChatGPT Enterprise and ChatGPT Business. They provide organisational administration, contractual protection for business data, connections to workplace systems, custom GPTs and agentic workflows. A firm can build legal processes on top of them, but the legal knowledge, playbook and verification system must come from the firm or from a connected third-party service. ChatGPT does not become a grounded legal research platform merely because it is deployed inside a commercial workspace.
The appointment matters for a different reason. Both leading foundation-model companies are now moving down the stack towards the workflows that legal AI vendors have treated as their defensible advantage. Anthropic has already shipped the connectors and plugins. OpenAI has hired one of the most experienced legal-workflow founders to build its response.
My read. Put OpenAI’s legal vertical on the watchlist, not the comparison table. Until there is a product to test, a contract to inspect and outputs to verify, there is nothing responsible to recommend. [Boehmig’s appointment and announcement are reported here](https://www.lawnext.com/2026/06/ironclad-founder-jason-boehmig-joins-openai-to-develop-products-for-the-legal-sector.html).
Category two: the enterprise platforms
Harvey
The category leader by every commercial measure. Founded 2022 by a former securities litigator and a Google DeepMind researcher. As of March 2026: a $200m round at an $11bn valuation, more than 100,000 lawyers across 1,300-plus organisations in 60 countries, a majority of the Am Law 100, and reported ARR around $190m. It has offices in Dublin, Paris and London, and roughly 30% of its customers are in EMEA.
The product is Assistant (chat and analysis), Vault (secure document store with grounded Q&A, syncing from iManage, SharePoint and Box), Workflows (multi-step matter-specific automation), Knowledge (firm-specific grounding on your own precedents), Agent Builder, and a Word add-in. In the Vals benchmark it was the standout performer, taking top scores in five of six tasks it entered and beating the lawyer baseline on four.
Now the thing Harvey’s own blog says, which no vendor guide will tell you. Harvey used to sell a proprietary legal model as its differentiator. It has stopped. In its post Expanding Harvey’s Model Offerings, the company states that within less than a year, seven models, including three from outside OpenAI, now outperform the originally benchmarked Harvey system on BigLaw Bench, Harvey’s own evaluation suite. Its explanation is that general foundation models have improved at baseline legal reasoning, which reduces the effort of adopting new models, so optimisation can focus on task execution, firm knowledge and collaboration rather than baseline reasoning.
Read that again as a buyer. Harvey is telling you, in its own words, that the reasoning is commoditised and that what you are paying for is the orchestration layer.
That is not a criticism. The orchestration layer is real and, at scale, valuable. Harvey’s own head of applied research publishes benchmark results on frontier models as they release, and the company now routes tasks to Claude and Gemini alongside OpenAI through a model selector. Its LexisNexis alliance, announced June 2025 and described by Artificial Lawyer as possibly the most important legal tech move in a decade, brings primary law and citations inside the Harvey interface. That aims to address the Ayinde problem directly.
Cost. Harvey publishes nothing. Third-party buyer reports put some mid-market contracts at roughly $1,000 to $2,000 per seat per month, sometimes with 20 to 25-seat minimums and twelve-month terms, though large firms with leverage reportedly pay far less per seat. None of those figures is a confirmed Harvey rate card, and terms vary materially. A 50-lawyer firm deploying across the team is nevertheless likely to be looking at six figures annually before anyone drafts a document. Bespoke custom-model builds, per Harvey’s CEO, can exceed $5m.
My read. Harvey is correctly priced for a customer comfortable spending a million dollars on a software pilot. If you are AmLaw or a Fortune 500 legal department with procurement, a partner champion and high-volume repeating workflows, it earns the money. If you are a UK SME practice, it is probably not for you, and Harvey is not pretending otherwise. The interesting fact for the rest of us is not Harvey’s price. It is Harvey’s public concession that the model underneath is no longer where the advantage lives.
Legora
Harvey’s direct rival, and the one to watch from a European perspective. Stockholm-founded in 2023 as Leya, rebranded 2025, now around $5.6bn post-money after a $600m Series D with NVIDIA’s venture arm and Atlassian joining. Reported ARR above $100m. Over 1,000 customers across 50-plus markets, including Linklaters, White & Case, Cleary Gottlieb, Bird & Bird, Dentons and Barclays.
The differentiated feature is Tabular Review, and it is the most genuinely useful idea in this article. Drop a folder of contracts in. Each document becomes a row. Each question you write becomes a column. “Pull the governing law clause, the liability cap and the change of control trigger from these 200 leases” returns as a sortable, filterable grid, each cell linked back to its source. For anyone who has spent a weekend tagging exceptions across a data room, that is the feature that justifies the spend.
Around it sit Workflows (multi-step agentic processes built in natural language, which can call a Tabular Review, run a web search and verify citations against a legal database within one run), Legal Research with cited synthesis, Word and Outlook add-ins with native track changes, automated Playbooks, and Portal, a white-labelled client-facing workspace.
Security and residency. ISO 42001 (the AI governance standard), ISO 27001, SOC 2 Type 2, GDPR, AES-256, EU and US data residency, bring-your-own-key, zero AI training or retention. For a UK or EU firm reading Munir carefully, that certification stack is the point.
The caveats. Third-party reviewers report Legora’s citation granularity is source-level rather than character-level, so you still open each authority. Pricing is not published; some buyer guides report around $3,000 per seat per year with a ten-seat minimum, plus consumption pricing for its most capable agent tier. Those figures are not a confirmed Legora rate card. It is a firm-side tool built for firm-scale document sets, and demo-only procurement makes it hard work for a small practice.
My read. If your practice involves genuine document volume, cross-border work, or multilingual review, Legora is the more interesting of the two giants and its EU posture is a real advantage under English privilege analysis. Note also what its CEO says when asked whether the model makers will eat him: foundation models are improving quickly, but the value is in how they are applied. That is the same argument Harvey now makes. It may well be right. It is also exactly what you would say if the reasoning had commoditised.
Category three: research platforms grounded in a licensed database
Lexis+ with Protégé (formerly Lexis+ AI)
Renamed in Q1 2026; LexisNexis’s own UK page confirms it. Now positioned as a workflow platform rather than a chat box: agentic drafting, DMS integration with iManage and SharePoint, and Protégé Vault for large matter document sets, supporting up to 100,000 documents, with outputs linking back to exact document passages.
The feature that matters most here is Shepard’s Verify Trust Markers, which identifies legal citations in AI-generated and human-drafted content, checks them against LexisNexis sources, and flags citations that cannot be verified as existing. That is a direct engineering response to the Ayinde problem, and it is the right one. LexisNexis’s CEO frames the whole platform around the proposition that legal AI must produce work lawyers can verify and defend. After the Stanford study, they would say that. It is still the correct target.
CoCounsel Legal UK (Thomson Reuters)
Built on the $650m Casetext acquisition, grounded in Westlaw and Practical Law, integrating with Microsoft 365, document management systems and HighQ. Deep Research runs multi-step agents across the databases. It reviews and analyses contracts, extracts key terms, builds chronologies and connects insights across thousands of documents. Every result carries citations. Among the first AI systems certified to ISO/IEC 42001.
In the Vals benchmark it was the only vendor other than Harvey to take a top score, averaging 79.5% across the four tasks it entered and exceeding the lawyer baseline on each by more than ten points.
Cost, both platforms. Neither publishes a rate card. Third-party buyer reports put some comparable all-in seats in the range of a few hundred pounds per user per month once the AI tier is bundled with the underlying database subscription, so a three-solicitor firm can clear £10,000 to £18,000 a year before billing an hour. Treat that as an order-of-magnitude estimate, not a quote. Both negotiate, hardest near renewal. Put the other platform’s written proposal on the table. Ask for the AI tier inside the seat price rather than as an add-on. Strike or shorten the auto-renewal clause before you sign, not after.
Category four: specialist contract and document platforms
Luminance
The most technically serious UK-origin platform in the category. Founded in Cambridge in 2015 by machine learning researchers, refined with Slaughter and May, deployed across many hundreds of organisations in over seventy countries. Diligence for M&A, Corporate for in-house workload, Discovery for eDiscovery with technology-assisted review and PII detection. Multi-model mixture-of-experts architecture with cross-validation before outputs surface. Heavily Word-oriented.
Fits large-volume transactional work and contract portfolio analysis. Does not do litigation: no case management, no legal research, no court filing. Quote-based pricing, deployment measured in weeks to months. A published assessment aimed at solo practitioners concluded it works best as an efficiency enhancer for high volumes of standardised documents, with significant manual intervention still required for setup.
Spellbook
Word-native contract AI, around 4,500 teams. Sits in a sidebar and reads the document in real time: redlining through Word’s native track changes, missing-clause detection, risk flagging, clause benchmarking, configurable playbooks. Zero data retention agreements with the underlying model providers, SOC 2 Type II, GDPR.
No published rate card; third-party estimates cluster around $99 to $199 per user per month, usually billed annually. Treat that range as indicative rather than confirmed. Contracts only: no case law research, no litigation support. Spellbook’s own guidance says AI output can be wrong and must be reviewed by a lawyer, which is more honest than most of its competitors manage.
The buying rule. If your team drafts ten or more contracts a month against similar paper and lives in Word, the workflow earns its cost. If contracts are one slice of your work, a general-purpose model at a fifth of the price gets you most of the way.
Summize and Juro
Two UK contract lifecycle platforms with opposite bets. Summize (Birmingham, substantial Series B in January 2026) embeds into Word, Outlook, Teams and Salesforce rather than making you move into a portal, and runs a multi-agent review with validation steps between agents. Juro (London, founded by Magic Circle lawyers) is browser-native, with zero-retention APIs and IASME certification.
Word-native versus browser-native is the first thing to decide, and it is not cosmetic. It is the single biggest determinant of whether legal tech survives month three. Note also that Juro is optimised for contracts born inside the platform; third-party paper and legacy documents are less comfortable territory, and a solicitor’s working life consists of other people’s documents in other people’s formats.
Genie AI
London, 2017, positioned for founders and lean in-house teams rather than law firms. Conversational drafting and review, template and clause library, red-amber-green risk flagging, ISO 27001, no training on customer data. Transparent pricing: a free tier, then around £30 a month. Reviewers note it can misinterpret complex clauses and that document comparison has been a gap. It is built to reduce dependence on lawyers, not to make lawyers faster. For a solicitor it is the wrong shape.
Category five: practice management with AI inside
Clio
Worth attention because of what it has bought. Clio is a cloud practice management platform serving over 200,000 legal professionals. In 2025 it acquired vLex for $1bn, the largest deal in legaltech history, bringing in Vincent AI and a proprietary database of over a billion editorially enriched legal documents across 110-plus countries. It separately acquired UK-based ShareDo, now rebranded Clio Operate for large firms, and document automation tool Lawyaw. It closed a $500m Series G at a $5bn valuation.
The strategic logic is the same one that matters after Munir: research, drafting, matter management and billing inside one governed environment, so nothing gets pasted into a browser. Clio Duo handles summarisation, client communications, smart recommendations and, usefully, an audit log of AI activity. Vincent brings multimodal analysis, legal theory testing, contract analysis and citation-linked research.
In the Vals benchmark Vincent participated across the tasks and performed respectably rather than at the top. Clio Duo is priced as an add-on to the Clio subscription, in the region of $49 to $59 per user per month on top of base pricing.
My read. For a small or mid-sized firm this is the most coherent bet in the category, because it solves the governance problem and the tooling problem at once. The risk is concentration: practice management, research, drafting and payments through a single vendor is a lot of eggs.
LEAP
A practice management system. Around 20,000 legal professionals across 2,500 UK and Ireland firms, and a Law Society strategic partner. Matter AI reads matter correspondence and surfaces information from within the file. Generator drafts documents. LawY is a legal research assistant with an optional human verification step: you submit the AI answer, a qualified lawyer reviews and where necessary amends it, and you are notified.
Same strategic argument as Clio: the AI operates on the matter file where the file already lives.
But. LawY’s marketing has described its verifiers as “eliminating any risk of error or misinformation”. No human verification process can honestly claim that, particularly with a stated turnaround of one to two business days and verification unavailable on some output types. Human-in-the-loop is a genuine and valuable design choice. It is not a guarantee, and it does not transfer your responsibility under the SRA Code to somebody else’s reviewer. Platform lock-in applies: the AI works within LEAP, so you must be on LEAP. Pricing on enquiry.
Access Legal’s CaseMatters Evo and Smokeball’s Archie AI occupy similar ground. When reading their comparison guides, note that each ranks its own product first. So does LEAP’s.
The thing nobody in the vendor guides will tell you
In late 2025, Robin AI collapsed.
Robin AI was a London contract review platform founded in 2019 by a former Clifford Chance disputes lawyer and a machine learning researcher. Around $10m annual recurring revenue. Thirteen Fortune 500 clients. ISO 27001 and SOC 2. Word integration. Roughly $70m of venture funding. By every metric a buying guide applies, a serious platform.
It failed to close a $50m round. It made layoffs in London and New York. [HMRC issued a winding-up petition](https://www.thelawyer.com/robin-ai-faces-hmrc-winding-up-petition-as-it-seeks-rescue-buyer/) in early November 2025. In December, [Scissero acquired its managed services team](https://legaltechnology.com/2025/12/10/scisseros-acquisition-of-robin-ai-the-combined-offering-fits-hand-in-glove/). Microsoft subsequently absorbed a group of its engineers. Nobody acquired the software product.
Consider what that means for a firm that had ingested its contract portfolio into Robin’s repository, configured its playbooks, trained its people, and built its review workflow around the platform.
So add a question, and ask it before you ask about features. What is your exit plan if this vendor disappears? Can you export your documents and your extracted data in a usable format? Is your playbook portable, or does it exist only as configuration inside their system? How much of your workflow lives in the tool and how much lives in your people?
The current funding environment does not make this question less urgent. Harvey is valued at nearly twice the size of the entire legal AI software market it operates in. Legora is valued at roughly the size of that market. Those bets may prove right. They are still bets, and it is your matter file inside them.
What I actually use, and why
The honest answer is unfashionable, and the independent evidence supports it better than it did a year ago.
I do not use a dedicated legal AI review platform. I use a frontier general-purpose model on commercial terms, with a structured set of instructions built around it that enforce the discipline the platforms sell as a feature. Generally, Claude (on their Pro Max plan) at the time of writing, but ChatGPT’s Work (released this week) is looking decent too.
Consider what the evidence above actually adds up to. Harvey, the $11bn category leader, publishes that seven frontier models now beat its own proprietary system on its own benchmark, and has repositioned around workflow orchestration. Legora’s CEO makes the same argument. Vals found a generalist model matching legal-specific tools on research accuracy across 210 questions. And on redlining, the task closest to what a commercial solicitor actually does all day, the lawyer baseline still beat every AI tool tested.
What the specialist platforms add over the underlying model is workflow, playbook configuration, integration, security certification, and consistency. Those are real. At firm scale, with a data room of 2,000 documents and twelve people who need the same answer, they are worth a great deal, and Tabular Review is a genuinely better idea than anything I could build. What they cannot give you is the professional judgment about whether the output is right. That remains yours under the SRA Code, and it remained yours in Ayinde, where the failure was not the tool’s.
Two things make my approach legitimate rather than reckless, and the recommendation is worthless without them.
The first is the tier. Commercial terms, with a data processing agreement, contractual no-training commitment, and appropriate retention settings. Not a consumer account. Munir is an emphatic warning about what may happen when you paste a client document into a public tool, and no productivity gain is worth exposing your client’s confidentiality or privilege.
The second is the instruction set. A structured review that goes clause by clause against your own stated positions, produces a risk-rated table, identifies missing provisions rather than only bad ones, runs an adversarial read of what the other side will argue, and flags every assertion it cannot verify against a source you supplied. It never cites authority you have not given it. Then you read it, as the supervising solicitor, and you decide.
That last part is the whole discipline in one sentence. The tool proposes. You verify. Nothing leaves your desk that you have not checked.
I have written all of this up properly: the framework, the verification standard, the actual prompts and installable skills I use for contract review, lease review, settlement agreements, attendance notes and pre-action correspondence, and the reasoning behind each.
Get The Lawyer’s Prompt
I wrote a book on AI for lawyers. It is based on hundreds of hours of practical use, five paid CPD seminars, and everything I have learned over the last few years.
I have since expanded it into a practical toolkit for sole practitioners, consultants and small law firms.
I have created fifteen solicitor-trained skills for Claude, plus the book that explains the discipline behind them. They are built around the verification standard reinforced in Ayinde: every authority sourced, anything unverifiable flagged, case law never invented.
Contract review, lease review, settlement agreements, SPAs, attendance notes, pre-action correspondence, client emails and much more In your house style, to one standard, every time.
Get The Lawyer’s Prompt, £99 →
There are free resources on the site too, no sign-up required: an AI policy generator that builds a tailored internal policy for your firm covering the SRA Code, the Ayinde verification duty, confidentiality and supervision; an AI verification checklist; and plain-English guides on AI for solicitors, AI contract review, and what Ayinde actually decided.
Build your firm’s AI policy, free →
Frequently asked questions
Can solicitors use ChatGPT? For anything involving confidential client material, not on an unapproved consumer account. In *Munir*, the Upper Tribunal observed that uploading confidential documents into a public AI tool such as ChatGPT places the information in the public domain, breaching confidentiality and waiving privilege. It said such conduct might warrant referral to the regulator and should be referred to the ICO. That was an observation in a Hamid decision, not a binding determination in a contested privilege dispute, but it is a warning no firm can sensibly ignore. Commercial tiers with a data processing agreement, no-training commitment and organisational controls are a different proposition, and the tribunal expressly distinguished closed systems. The verification duty applies either way.
What is Claude for Legal? It is not a separate model or a proprietary legal database. It is a collection of legal workflow plugins and connectors operating around Claude. The plugins encode repeatable processes for areas including commercial, corporate, employment, privacy and litigation work. The connectors can bring governed sources such as document management and legal research systems into the workflow. It is a significant addition, particularly for small teams, but the plugins remain instructions rather than a guarantee of accuracy.
Is there a ChatGPT for Legal? Not yet. OpenAI appointed Ironclad founder Jason Boehmig in June 2026 to lead product for its legal vertical, but it has not announced a product name, feature set, price or launch date. ChatGPT Business and Enterprise can be configured for legal workflows today; that is not the same as a released legal-specific product.
Does using Microsoft Copilot avoid the problem? Private, properly configured instances of Copilot inside your own Microsoft 365 tenancy are the example the Upper Tribunal itself gave of a closed system that does not place information in the public domain. That is a strong position, but reduced risk is not no risk, the privilege point is not settled by binding authority, and consumer Copilot tiers do not carry the same protection. Check your Product Terms and your DPA rather than assuming.
Is Harvey worth it? For an Am Law firm with procurement, budget authority and high-volume repeating workflows, the arithmetic can work: reported adoption is high, and at those charge-out rates the time saved may cover the licence several times over. For a UK SME practice, the reported cost and minimum commitments are difficult to justify, and Harvey is not primarily built for that market. Note that Harvey itself now says the differentiator is workflow orchestration, not the model.
Harvey or Legora? On public information: Harvey for US-dominant common law practices and depth of AmLaw integration, Legora for European data sovereignty, multilingual work, and high-volume structured extraction via Tabular Review. Both are demo-only, quote-based, and priced for firms with a procurement function.
Is AI contract review accurate enough to rely on? No tool should be relied on without verification. The independent Vals benchmark found AI beating a lawyer control group on data extraction, document Q&A, summarisation and chronology generation, and losing to lawyers on redlining. The Stanford study found the leading grounded research platforms hallucinating between roughly 17% and 33% of the time on the versions tested in 2024, and concluded vendor claims were overstated. Use AI where it is strong. Do not delegate judgment.
Which AI tool is best for a small law firm? Probably none of the enterprise platforms; their economics assume you are displacing hundreds of associate hours. For a small firm or a consultant, a frontier general-purpose model on commercial terms with a disciplined structured instruction set gets most of the value at a fraction of the cost, provided the discipline is genuinely there. If contracts dominate and you live in Word, look at Spellbook. If you want the AI inside the matter file, look at Clio or LEAP.
Can I put client documents into an AI tool? Only with a clear lawful basis, appropriate contractual protection, and where relevant informed client consent. Get the DPA in writing. Check data residency, whether the underlying model providers are named, and whether zero data retention is a contractual term rather than a marketing claim. Never into a public consumer tool. If it has already happened, treat it as an incident: it may need reporting to the ICO and, on the Upper Tribunal’s view, potentially to the SRA.
Not a solicitor?
This guide is written for lawyers. If you are a business owner, consultant or employee who simply needs a contract looked at before you sign it, the tools above are the wrong shape and the wrong price.
If you’re looking for a contract review, I can plainly help.
QuickLegalCheck does AI contract reviews from £49, with a plain-English report in minutes. It is not legal advice, it is not a regulated service, and for anything complex or high-value you should instruct a solicitor. But it is faster and cheaper than signing something you have not understood. Disclaimer: I helped build the system.
——————–
Steven Mather is a solicitor specialising in business sales and acquisitions, commercial contracts and corporate work for owner-managed businesses. Nothing on this page is legal advice, and nothing here is a recommendation to buy any particular product (except my book AI for Lawyers). All pricing figures are third-party estimates unless stated otherwise; most vendors in this sector do not publish rate cards. Do your own due diligence, particularly on vendor stability and data protection terms.
