Top 10 AI Agents for Law Firmsin 2026

The single most useful thing to know about legal AI rankings in 2026 is that the roster changed faster than the writing did.Before any question of which product is better, there is a prior question a managing partner should ask of every list they read: does this list even name the products correctly? Four vendor renames landed inside eighteen months, alongside a billion-dollar acquisition and one company's dismemberment. A ranking written from memory rather than from the register is wrong in at least five places before it reaches a single judgement.

Here is the corrected roster, each item checkable in a public source. Casetext no longer exists as a brand — Thomson Reuters completed its acquisition for $650 million in cash on August 17, 2023, and the 2026 product name is CoCounsel Legal, not CoCounsel 2.0. Leya became Legora in February 2025. Callidus Legal AI became StrongSuit on November 12, 2025 — and its Crunchbase and CB Insights entries still sit at a callidusai slug, which is how the dead name keeps getting copied. Lexis+ AI was replaced outright by Lexis+ with Protégé on February 24, 2026; LexisNexis now describes the older product as its first-generation AI experience. vLex was acquired by Clio for $1 billion in a deal that closed on November 10, 2025, with Vincent AI retained rather than retired. And Relativity is Silver Lake-controlled — visible in the board composition on Relativity's own leadership page rather than in any press release.

None of that is a gotcha. It is the ordinary consequence of a market consolidating at speed, and it is exactly why a ranking has to be re-verified rather than inherited. The vendor-risk section below works through the sharpest case of all: a company that appears on virtually every 2026 legal AI listicle, whose engineering team is now at Microsoft and whose services arm is now at Scissero, and which Companies House nonetheless still lists as Active.

The second thing worth knowing is that legal is unusual: it is one of the few markets where a genuinely independent, peer-reviewed evaluation of AI products exists at all. That evaluation is a decade of credibility in a single paper, and it is also narrower and older than the way it gets quoted. Handling it honestly — reporting what it found, what it covered, and the fact that one of the products it tested has since been retired — is the difference between a ranking a partner can rely on and one they will quietly discount.

And the third: the constraint that decides whether any of this works in your firm is not technical. It is professional-responsibility law, and in California it has now been written specifically about agents. We get to that below, and we build the human-in-the-loop boundary table directly out of the regulator's own language rather than out of ours. If you want the wider cross-industry version of this argument, our survey of the top AI agents across every category in 2026 applies the same evidence standard outside legal.

We scored eight attributes, every one of which you can verify from a public source, and we scored no performance metric of any kind. Here is the methodology in full, stated up front so you can disagree with it before you read the table rather than after.

What we refused to score: accuracy, hallucination rate, hours returned, time saved, matter throughput, realisation lift and return on investment.Not because those things do not matter — they are the only things that matter to a P&L — but because in this market every published figure of that kind was produced by the party selling the product, on a private dataset, with an unpublished methodology. Scoring them would mean ranking vendors on the quality of their marketing departments.

That paragraph is worth unpacking, because it is the load-bearing claim of the whole article. The peer-reviewed study is Magesh, Surani, Dahl, Suzgun, Manning and Ho, "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools", published in the Journal of Empirical Legal Studies (DOI 10.1111/jels.12413) and accepted March 14, 2025. Cite that version, not the 2024 preprint, whose Westlaw figure was superseded. It was preregistered, used 202 hand-constructed legal queries with manual expert coding of every response, and released its dataset with half withheld to guard against model memorisation.

Its definition of a hallucination is the methodological crux and is worth quoting exactly, because vendors dispute it: "A response is considered hallucinated if it is either incorrect or misgrounded. In other words, if a model makes a false statement or falsely asserts that a source supports a statement, that constitutes a hallucination."That means a factually correct answer that cites a source which does not actually support it counts as a hallucination. For a lawyer preparing a brief, that is the right definition. For a vendor's marketing team, it is an unfair one. Both positions are coherent; you should know which one a number is using.

"We demonstrate that the providers' claims are overstated. While hallucinations are reduced relative to general-purpose chatbots (GPT-4), we find that the AI research tools made by LexisNexis (Lexis+ AI) and Thomson Reuters (Westlaw AI-Assisted Research and Ask Practical Law AI) each hallucinate between 17% and 33% of the time."

— Magesh, Surani, Dahl, Suzgun, Manning & Ho, Journal of Empirical Legal Studies (2025)

The claims being tested were not vague. The paper's own endnote reproduces them: LexisNexis marketed "100% hallucination-free linked legal citations"; Casetext said CoCounsel "does not make up facts, or 'hallucinate,' because we've implemented controls to limit CoCounsel to answering from known, reliable data sources"; Thomson Reuters said it avoided hallucinations by relying on trusted Westlaw content. Those are the sentences the study was designed to test, and it contradicted them.

Now the fairness, in the same breath. LexisNexis's Chief Product Officer for North America and UK, Jeff Pfeifer, responded on May 24, 2024: "LexisNexis has not been contacted by Stanford's Daniel Ho, and our own data analysis suggests a much lower rate of hallucination." He also argued the criteria measured answer and citation quality rather than hallucination specifically. LexisNexis has never published the underlying data. Stanford's reply was that it had been denied access to the tools; the paper's endnote records that Thomson Reuters denied three access requests during the initial evaluation and provided access only after the results were released. Stanford then augmented the study, and the augmented results were nearly identical — Ask Practical Law's accuracy moved from 19% to 20%.

And the caveat that matters most in 2026: the products tested no longer describe the products sold.The evaluation was completed in April 2024. The highest-performing system was Lexis+ AI, accurate on 65% of queries — and Lexis+ AI was replaced entirely by Lexis+ with Protégé on February 24, 2026. No successor evaluation of the replacement has been published by anyone. Presenting 17%–33% as a current performance figure for today's products would be exactly the error this article exists to refuse. Presenting it as the best independent measurement the profession has ever obtained, of products that have since moved on, is accurate.

The other benchmark in this market is the Vals Legal AI Report, run by Vals AI with Legaltech Hub, using datasets from eight law firms and an independent-attorney baseline sourced through Cognia Law. It is genuinely useful and it is not equivalent to the Stanford work, because vendors choose which of the seven tasks to enter. In the February 2025 round Harvey entered six, Thomson Reuters entered four, vLex Vincent AI entered six, and LexisNexis withdrew from all but legal research and does not appear in the published report. When Vals separated legal research into its own October 2025 study, neither Thomson Reuters nor LexisNexis opted in, and Harvey, Legora, Luminance, Spellbook and Eve are all absent. Vals also discloses that it has a customer relationship with one or more participants.

For e-discovery the position is starker and completely checkable. NIST's TREC Legal Track ran from 2006 to 2011 and stopped. The archive carries year pages through 2011 and nothing after. That reference corpus underpinned the technology-assisted-review validation literature courts leaned on when they approved predictive coding — and it has had no maintained successor for roughly fifteen years, over precisely the period in which the tools became generative. So when a vendor cites precision and recall, or documents per hour, or growth multiples, there is no neutral corpus those figures could have been measured against.

One more refusal, briefly, because it is the standard opening line of AI articles: the "95% of generative AI pilots fail" figure and the Gartner projection that a large share of agentic projects will be cancelled do not survive a citation check as evidence about your firm. The first rests on a small interview and conference-survey sample about custom-built tools; the second measured investment posture in a poll. Neither is a base rate you should plan against, and neither appears anywhere in our analysis.

Everything below was checked on August 23, 2026.Re-check it yourself: open the vendor's trust centre, open its terms of service, open Companies House or SEC EDGAR, and open the pricing page. Those four tabs will confirm or refute every cell in the next section. If a cell has moved since publication, the vendor changed — not the method.

Every cell below is either a fact traceable to a public source or the literal words "not publicly disclosed". There is no accuracy column, no efficiency column and no score. Two patterns leap out of the table before you read any individual row, and both are more decision-useful than any performance claim would have been.

Corporate status, certifications, contract terms and pricing disclosure checked August 23, 2026. Cells marked as not found or not publicly disclosed record a gap in public disclosure, not evidence that a commitment does not exist privately.

The integration picture is the other half of the buying decision, and for a law firm it is frequently the deciding half. A tool that cannot see your document management system is a tool your associates will paste into, which is a confidentiality problem before it is a productivity one. Here the disclosure quality varies enormously — and the vendor with the best-documented list is not the one with the loudest agent marketing.

Integration claims are recorded at the strength of their evidence: vendor documentation, vendor marketing page, or partner confirmation. Where we could not reach documentation, we print no claim rather than repeating one.

Read those two tables together and the shape of the market is clear. The vendors with the deepest firm-side integration documentation — CoCounsel Legal, Spellbook, Definely — are not the vendors with the strongest AI-specific certification. The vendors with the strongest certification posture — Harvey, Legora — publish the least about how they connect to your DMS. And the two e-discovery platforms with the most mature security programmes, Everlaw and Relativity, document no law-firm DMS connector at all, because they were built as evidence platforms rather than as practice tools. There is no product here that is strong on every axis, which is a more honest finding than a ranked score would have produced.

Each entry below states what the product does, what is verifiable, what is not disclosed, who it fits, who it does not, and who owns it. The ordering reflects the verifiable-attribute picture, not a performance judgement — there is no performance judgement available to make.

Fifteen vendors passed our corporate verification and one did not, so there was no need to pad the list — and no excuse to write any entry from memory. Five verified vendors sit outside the ten either because they solve a narrower problem or because they are built for in-house legal rather than for a firm. They are worth knowing about, and in two cases they are better answers than anything in the ten.

An AI agent may draft, research, extract, summarise and queue; a licensed lawyer must review, verify and sign. That boundary is not a best practice — it is the rule, and in California it is now written specifically about agents. No product decision in this article matters as much as this section, because the constraint binds identically regardless of which vendor you choose.

Start with the baseline. ABA Formal Opinion 512, issued by the Standing Committee on Ethics and Professional Responsibility on July 29, 2024, is the profession's reference text on generative AI. Its opening summary is the whole duty set in one sentence: lawyers using these tools must fully consider their obligations of competence, confidentiality, communication, supervision of employees and agents, advancing only meritorious contentions, candour toward the tribunal, and reasonable fees.

"a lawyer's reliance on, or submission of, a GAI tool's output—without an appropriate degree of independent verification or review of its output—could violate the duty to provide competent legal representation as required by Model Rule 1.1."

— ABA Formal Opinion 512, “Generative Artificial Intelligence Tools” (July 29, 2024)

Read the calibration alongside it, because 512 is more practical than its reputation suggests and this passage is the most operationally useful sentence in the opinion. It expressly contemplates sampling-based validation: if a lawyer uses a tool to review and summarise numerous lengthy contracts, the lawyer would not necessarily have to manually review the entire set to verify the results, if the lawyer had previously tested the tool's accuracy on a smaller subset by manually reviewing those documents, comparing them to the tool's summaries, and finding the summaries accurate. That is a workable protocol you can write into a firm policy — and it is the difference between a defensible deployment and a theatrical one.

On confidentiality, 512 is stricter than most firms have yet operationalised: because many self-learning tools are designed so their output could lead directly or indirectly to disclosure of information relating to a client's representation, a client's informed consent is required before inputting such information — and merely adding general boiler-plate to engagement letters purporting to authorise GAI use is not sufficient. It also sets a baseline every lawyer should recognise as a work instruction: read and understand the terms of use, privacy policy and related contractual terms of any tool you use, or consult someone who has.

And on supervision, Rules 5.1 and 5.3 turn vendor diligence into an ethical duty rather than a procurement preference. Managerial lawyers must establish clear policies on permissible use; supervisory lawyers must make reasonable efforts to ensure compliance. 512 then applies the cloud-and-outsourcing checklist to AI providers directly: reference checks and vendor credentials, understanding security policies and protocols, confidentiality agreements, understanding the vendor's conflicts-check system to screen for adversity among firm clients, and the availability of a legal forum for relief under the vendor agreement. That paragraph is, in effect, the specification for the comparison table above.

Its definition of agentic AI is better than any vendor's: systems that can autonomously perform tasks or workflows without human prompting, built not just with multistep functionality but able to pursue goals, plan, select tools and execute tasks without immediate human intervention — with examples including pleading revisions across multiple iterations, preparing discovery responses, coordinating review of documents, and autonomously facilitating client intake. If you want a working definition to hold vendors against, use that one.

"Lawyers must not deploy agentic systems in a manner that allows the system to make substantive legal determinations, communicate legal advice, prepare and file pleadings, or otherwise act in a representative capacity without meaningful lawyer supervision and review."

— State Bar of California, COPRAC — 2026 Practical Guidance

The guidance adds three things a firm should build directly into policy. First, autonomy does not dilute duty: this autonomy does not satisfy a lawyer's duty to exercise independent judgment, nor does it alter existing ethical obligations, and professional judgment cannot be delegated to AI. Second — and this is the scaling principle the boundary table below is built on — the greater the degree of autonomy afforded to an AI system, the more important it is for a lawyer to implement supervisory controls and verification mechanisms.Oversight is not a fixed overhead; it scales with what you let the system do. Third, on filings specifically: where agentic AI is used in connection with court filings, no document may be transmitted to the court without lawyer review and approval, and lawyers must not permit AI systems to autonomously file documents, communicate with the court, or make representations on the lawyer's behalf.

The confidentiality passage is the one that should change your architecture. Agentic systems may be configured with persistent or automated access to email, messaging platforms, document management systems, knowledge bases, client files and calendaring systems, raising acute confidentiality considerations — and unrestricted or poorly configured agentic systems may unintentionally disclose confidential information across different matters and even expose privileged material. The guidance then draws the hard line: a lawyer must not deploy an agentic system in a manner that permits autonomous external transmission of client information, including automated communications, filings or data transfers, without appropriate safeguards and human review.

One sentence in that document justifies this entire article, and it is worth pinning above a procurement desk: "Reasonable efforts require more than reliance on generalized marketing assurances."

There is a conflicts angle most rankings miss entirely. ABA 512 cites a joint Pennsylvania and Philadelphia bar opinion warning that because large language models continue to develop, some without safeguards similar to those already in use in law offices such as ethical walls, they may run afoul of Rules 1.7 and 1.9 by using information developed from one representation to inform another. That is precisely why Harvey's July 2026 Intapp integration for ethical-wall enforcement is a more interesting feature than most of the agent marketing around it.

The record of what happens when the verification duty is skipped is now large, and it needs to be quoted carefully. The AI Hallucination Cases database, maintained by Damien Charlotin, a research fellow at HEC Paris's Smart Law Hub, records decisions where a court or tribunal has explicitly found or implied that a party relied on hallucinated content. It expressly does not track the wider universe of all fake citations in filings. As of August 23, 2026 it holds 1,955 cases worldwide and 1,344 in the United States.

The finding that makes this relevant to a product ranking is that twenty-five cases in the database name a commercial legal-research AI product rather than a consumer chatbot. Before you read that table, read the two caveats that must travel with it, because a ranking that cites the tracker against a vendor without carrying the vendor's dispute is not citable. First, the database's own disclaimer: the mention of a specific tool does not necessarily mean that tool was responsible for the hallucinations in question — the tool named is the tool the sanctioned party said they used. Second, Thomson Reuters investigated and states that the errors did not originate in either Westlaw or CoCounsel, and that dispute is recorded against every one of its entries.

Selected US entries naming a commercial legal AI product, from the AI Hallucination Cases database as of August 23, 2026. Sanctioned, struck, disqualified, referred and warned are different outcomes; each is stated as the record shows it. Vendor dispute column records Thomson Reuters's stated position, which appears on twelve records in the database.

Even with both caveats fully carried, the operational point stands and is worth stating plainly to your partnership: buying a professional-grade legal AI product does not discharge the verification duty, and courts have sanctioned lawyers who assumed it did.That is not an argument against these tools. It is an argument for treating the verification step as billable, staffed work rather than as something the subscription bought you. And it is the reason ABA 512's footnote citing the Stanford study — the profession's ethics regulator relying on the independent evaluation — is the most quotable pairing in this whole field.

One procedural note with a status caveat, because standing orders are amended and withdrawn. The peer-reviewed Stanford paper records that as of May 2024, more than twenty-five federal judges had issued standing orders instructing attorneys to disclose or limit AI use in their courtrooms. We could not verify a current 2026 count and do not state one; nor could we confirm that one frequently cited certification order remains in force on its court's current requirements page. Both ABA 512 and the California guidance tell lawyers to check the applicable jurisdiction's rules and orders. That is the correct instruction, and it is a per-matter task rather than a firm-level one.

If your firm has EU exposure, two facts govern your planning, and the version circulating in most 2026 articles is out of date on both.

The first is the calendar. Regulation (EU) 2026/1744 of 8 July 2026 — the Digital Omnibus on AI, published in the Official Journal on 24 July 2026 and in force from 27 July 2026 — deferred the high-risk dates. Recital 40 sets the date of application of Sections 1, 2 and 3 of Chapter III to 2 December 2027 for systems classified as high-risk under Article 6(2) and Annex III, and to 2 August 2028 for those classified under Article 6(1) and Annex I. Anyone still writing that high-risk obligations apply from 2 August 2026 is repeating the pre-Omnibus timeline, and that is the single likeliest regulatory error in this subject area.

The second fact is a classification point that is routinely conflated, and getting it right will save you an unnecessary compliance programme. Annex III point 8(a) covers "AI systems intended to be used by a judicial authority or on their behalf to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts, or to be used in a similar way in alternative dispute resolution."

Two further details for anyone drafting an internal position. Providers who placed systems on the market before 2 August 2026 have four months to adapt their marking practices — a December 2026 checkpoint for content marking on pre-existing systems. And amended Article 50(7) now frames the Commission's role as encouraging and facilitating codes of practice at Union level for detection, marking and labelling of artificially generated or manipulated content, rather than exercising a lost implementing-act empowerment.

On the US side, the relevant regulators are the state bars rather than a federal AI statute. California's 2026 guidance is the fullest treatment of agentic AI by any US regulator. The New York State Bar Association's AI Task Force, the State Bar of California in its earlier 2023 form and the Florida Bar had each published detailed guidance before it, and ABA 512 additionally cites opinions from Florida, a joint Pennsylvania and Philadelphia opinion, West Virginia and the District of Columbia. We could neither confirm nor exclude the existence of a 2025–26 ABA successor opinion to 512, so we assert neither — though 512 itself anticipates that updated guidance will follow as specific tools develop.

One evidence-rules caution, stated with status precision because this is an area where confident wrong answers circulate. A proposed Federal Rule of Evidence dealing with machine-generated evidence is not listed among pending amendments as of August 23, 2026, and we do not write that it is coming into force. What is currently published for public comment is amendments to Evidence Rules 104 and 902, published 14 August 2026 with a comment deadline of 15 February 2027 and a projected effective date of 1 December 2028 — and we could not confirm whether those amendments concern machine-generated or AI evidence, so we characterise their subject matter no further.

Consider a 24-lawyer litigation and commercial firm producing 40 research memos a month, each carrying roughly 12 citations. Assume, before any AI, that a memo takes a senior associate 3.0 hours end to end. That is 120 hours a month of memo production, and 480 citations a month leaving the building over a partner's signature.

Step one: the verification load, computed from the only independent measurement available. The Stanford study found hallucination rates of 17%–33% across the products it tested in April 2024, where a hallucination means a statement that is either incorrect or misgrounded — including a factually correct statement citing a source that does not support it. Applied to a 12-citation memo, that range implies roughly two defective citations per memo at the low end and four at the high end. Across 40 memos, that is 82 to 158 defective citations a month if nobody checks.

Two caveats have to travel with that arithmetic or it is dishonest. The measurement covered three products, one of which no longer exists, and it is a point-in-time result from April 2024 rather than a current performance figure for anything on sale today. And no successor evaluation has been published, so we cannot tell you the rate is lower now — only that we do not know. What the arithmetic legitimately demonstrates is the shape of the exposure: at any plausible rate, unverified AI-drafted memos put defective citations into your work product at a monthly volume, not an annual one. That is the mechanism behind the 293 US lawyer sanction cases recorded in 2026 through August 23.

Step two: the time budget. Assume drafting time falls from 3.0 hours to 1.0 hour with a research assistant in the workflow. Now add the verification step that ABA Formal Opinion 512 requires — pulling each cited authority and confirming it says what the draft claims. At four minutes a citation, twelve citations is 48 minutes, or 0.8 hours. Total per memo: 1.8 hours against a 3.0-hour baseline. Across 40 memos, 72 hours instead of 120. 48 hours a month released.

Note what the verification step does to the headline. Naive arithmetic on drafting alone would claim a 67% reduction. Honest arithmetic, with the ethics obligation priced in, gives 40%. If a vendor's business case does not contain a verification line, it is quoting the naive number — and 512 does not permit you to run the naive number.

Step three, and the part that surprises most partnerships: who captures the 48 hours. On hourly matters, nobody at the firm does. ABA 512 states that lawyers who bill an hourly rate must bill for their actual time, and that where a tool makes work much faster it may be unreasonable under Model Rule 1.5 to charge the same flat fee as before. So on hourly work those 48 hours simply leave the invoice. The firm captures value only where the work is flat-fee, contingency, or where released capacity is redeployed onto matters the firm would otherwise have declined. That is a real and often substantial gain — but it is a capacity argument, not a margin-per-matter argument, and the two get conflated constantly in vendor decks.

Step four: can you pass the subscription through?Ordinarily not. California's 2026 guidance says subscription fees for tools providing general office functionality — drafting assistance, research capabilities, document review — typically constitute overhead similar to library maintenance or general computer systems, and should be absorbed in the fee rather than charged separately. Where a charge is legitimate, it must reasonably reflect actual cost, with no markup or profit element absent the client's informed written consent. ABA 512 draws the same line differently: a grammar tool embedded in word processing is overhead, while a third-party service charging per use for one client's document review is ordinarily a billable out-of-pocket expense. The usage-based e-discovery platforms are the clearest case on the billable side of that line; a firmwide research seat is the clearest case on the overhead side.

The honest conclusion from that arithmetic is not that legal AI does not pay. It is that the payback is a capacity and quality story on a defined subset of your matters, and the verification step is a permanent line item rather than a transitional one. A firm that budgets for the verification and targets flat-fee work will do well. A firm that buys on a 67% number and bills clients for the software will have two problems.

In this market the first failure is almost never the model — it is the roster, the contract or the boundary. Here are the failure modes in rough order of likelihood, each with the signal that reveals it and the rollback that contains it.

Every row of this table is grounded in ABA Formal Opinion 512 or the State Bar of California's 2026 Practical Guidance, not in our opinion about good practice. It is written to be lifted straight into a firm AI policy, and the right-hand column is deliberately absolute.

The organising principle behind the table is the California guidance's scaling rule, which is more useful than any fixed checklist: the greater the level of system autonomy, the greater the lawyer's obligation to implement oversight mechanisms sufficient to ensure that professional judgment remains with the lawyer. In practice that means the supervision budget is a function of the product you bought. A Word add-in that surfaces defined terms needs almost none. An agent with persistent DMS access that runs overnight batches needs a named owner, an audit log somebody reads, and a documented sampling protocol.

Three further boundary rules from the same source belong in the policy. Training is a duty, not a perk: supervision includes providing appropriate training on the ethical and practical aspects, and pitfalls, of tools that perform tasks with limited or no real-time human direction. Juniors are not shielded by instruction: a subordinate lawyer must not use AI-enabled tools, including agentic systems, at a supervisor's direction in a manner that violates their own professional obligations. And competence is recurring: because these systems evolve through updates and model changes, the duty of competence includes periodic reassessment of capabilities and risks, including when deploying a system for a new legal task. That last one is why a firm AI policy needs a review date on it.

One bias note worth carrying into any multi-agent design: the California guidance observes that this can be of particular concern when using agentic AI, as multiple agents can compound any underlying bias. If your architecture chains agents, the compounding is a design property rather than an edge case.

For most small and midsize firms, the honest advice is to buy a product first and commission a build only where the workflow is specific to your practice and nobody sells it. The ranked products cover research, drafting, contract review and e-discovery competently. Where a custom build earns its keep is the connective tissue: intake triage against your own criteria, matter-specific extraction from a document type your practice sees constantly, or an integration between two systems neither vendor will build. We work through that decision in detail in our build-versus-buy analysis for AI agents, and the short version is that buying loses to building only when the workflow is a differentiator rather than a commodity.

When a build is the right answer, these are Frenchy Digital's bands. They are the same bands we quote across every industry, and we publish them because nobody else in this market publishes a number at all.

Senior-led work runs $150–$225 per hour; retainers run $2,500–$9,500 per month. Every engagement carries a 30-day post-launch warranty, and full source-code and IP ownership transfers to you — which for a law firm is not a nicety but a conflicts and confidentiality requirement. We return a fixed-price phased proposal within 5 business days. Frenchy Digital is a senior-led, Black-owned agency in Los Angeles, and you can book a discovery call at calendly.com/frenchydigital/discovery-call or reach us on +1 (424) 272-5601.

A sequencing note specific to law firms. Start with the discovery and workflow audit, because the deliverable you need before anything else is a vendor diligence pack mapped to Rule 5.3 and a written human-in-the-loop boundary your ethics partner will sign. That artefact is useful whether you end up building or buying, it is the evidence of reasonable supervision the rules require, and it costs a fraction of the first phase of a build. Then take one workflow — not three — into a single-workflow agent, with the baseline measured before you start so the result is arguable rather than anecdotal.

Every item below is a pattern we observed in this specific roster while researching this article, not a generic warning list. None of them disqualifies a vendor on its own. All of them are questions you are entitled to ask in a procurement call.

• A regulation listed as if it were a certification: One vendor's security page lists “EU AI Act Compliant” alongside SOC 2 and GDPR. No certification scheme issues that. Ask which body audited it and what the scope statement says.
• An awards programme presented as independent testing: A submission-based industry award is marketing, not measurement. One ranked vendor's only third-party recognition is an award. That is worth knowing, and it is not a benchmark.
• User-review platforms cited as benchmarks: Peer-review sites aggregate opinion from self-selecting users. They tell you about sales motion and support quality. They tell you nothing about accuracy.
• A vendor-sponsored podcast or panel offered as validation: One frequently cited “independent validation” in e-discovery is a discussion with no dataset, no metrics, no methodology, speakers from a firm's own technology subsidiary, and no disclosure of a commercial relationship.
• Marketing claiming a certification the security page does not: One vendor's homepage claims SOC 2 Type II certification while its security page says only that its programme follows the SOC 2 framework. Following criteria is not holding an attestation. Ask for the report and the observation period.
• A no-training promise that lives only on a marketing page: Ask where it appears in the contract. In this roster, only two vendors put it in the binding document unqualified, one qualifies it, one permits first-party training on opt-in, and one published agreement reserves the opposite right entirely.
• A price quoted to you from a third-party blog: No vendor in this market publishes pricing. If a comparison site shows you a per-seat figure, ask for its source — one widely circulated figure is contradicted by the vendor's own page, and a competing vendor is among the most-cited sources for another.
• “Hallucination-free” or any absolute accuracy claim: The claim that leading tools deliver hallucination-free citations is the specific claim the one peer-reviewed independent study was designed to test, and contradicted. Treat any absolute as a reason to ask for the evaluation protocol.
• Agent language the product documentation does not support: Several vendors market agents while their own product pages describe augmentation, plan-then-approve workflows, or explicitly disclaim autonomy. Read the documentation, then ask the rep to point at the autonomous execution loop.
• A certification badge doing the work of a confidentiality commitment: One vendor holds ISO/IEC 42001 and publishes no no-training commitment anywhere. A management-system certification says how the vendor governs AI, not what it will do with your client's documents.
• A trust centre or documentation URL that does not resolve: A de-provisioned trust subdomain, a 404 pricing page and a stale copyright footer were the earliest public signals in the one vendor we refused. They cost thirty seconds to check.
• No answer on data residency or retention: Six of the vendors here publish nothing on retention. For a firm with EU clients or outside-counsel guidelines, that is a contract negotiation, not a footnote.

The one question that does more work than all of the above: ask the vendor to name any independent evaluation of its product and to send you the protocol. Four vendors in this entire market can answer that at all, and two of those answers point to a benchmark the vendor chose which tasks to enter. Everyone else will send you a case study. That is not a reason to walk away — it is the state of the market — but it tells you exactly how much weight the performance conversation can carry, which is very little.

A ranking that reports only what it found is not a ranking you can audit. Here is what we chased and could not establish, stated plainly so you can weigh everything above accordingly.

• Independence for two vendors rests on absence of news: Everlaw's and Ironclad's private status is supported by recently refreshed first-party documents and no acquisition reporting — not by a registry, because no filing exists for a private US company. We write “no acquisition found as of August 23, 2026”, never a flat statement of independence.
• Whether an ABA successor opinion to Formal Opinion 512 exists: americanbar.org blocks programmatic fetch. We could neither confirm nor exclude a 2025–26 successor, so we assert neither. 512 itself anticipates that updated guidance will follow.
• The exact 2026 issue date of the California COPRAC guidance: We read the document in full; only its day and month are missing, because the bar's own index paths were unreachable. We cite it as the 2026 Practical Guidance without a precise date.
• Thomson Reuters's compliance posture for CoCounsel: No trust page evidencing SOC 2 Type II or ISO certification for the product could be located. That is a disclosure gap, not evidence of absence — and it is a question for your procurement call, not a mark against the product.
• Clio's no-training commitment and current pricing: clio.com returns 403 to programmatic fetch, so both cells are unverified rather than negative. Secondary pricing figures conflict badly and we print none.
• StrongSuit's compliance posture, pricing and legal entity name: strongsuit.com returns 403. We can confirm the November 2025 rename and the July 2026 product launch from press releases, and nothing further.
• The date and terms of the Silver Lake–Relativity transaction: Relativity's own announcement page is undated and Silver Lake's page returned a server error. The figures in circulation are secondary only, so we print no date and no valuation — only the board composition, which is verifiable.
• Whether Luminance's published Master Hosted Agreement is still current: It is dated July 3, 2024. Clause 10.1.3 is quoted as published, and any buyer should confirm the current MSA before relying on it either way. We also could not establish whether the August 2026 share allotments represent a new round.
• Definely's cloud provider and no-training position: Its trust centre says AWS and its privacy policy says Azure, so we state neither. And we found no no-training commitment in any published document despite its ISO 42001 certification.
• General availability of DISCO's agentic Cecilia: Announced February 9, 2026 for “later in 2026”, with no GA confirmation found. We do not write that it is shipping.
• Any DMS connector claim for Relativity: Documentation paths could not be reached, so we print no iManage or NetDocuments claim for Relativity at all, in either direction.
• A current count of federal judges with AI standing orders: The only figure we can source is “more than 25”, as of May 2024, via the peer-reviewed Stanford paper citing a legal-press tracker. We do not present a 2024 count as a 2026 count, and we could not confirm that one widely cited certification order remains in force.
• Whether the pending amendments to Federal Rules of Evidence 104 and 902 concern AI evidence: Published for comment on August 14, 2026 with a February 15, 2027 deadline and a projected December 1, 2028 effective date. We did not read the preliminary draft and characterise their subject matter no further. A proposed rule on machine-generated evidence is not on the pending list.
• Any utilisation, realisation or hours-recovered figure: Every one we encountered was published by a seller, including one attributed to a vendor's own General Counsel. We could not obtain the methodology behind the most-cited utilisation series, whose publisher is itself ranked in this article, and we print none of them.
• Independent evaluation for eleven of the fifteen verified vendors: None exists at all for Legora, Luminance, Spellbook, Eve, StrongSuit, Everlaw, Relativity, DISCO, Ironclad, Definely or Wordsmith. Only Harvey, CoCounsel, vLex Vincent AI and the retired Lexis+ AI have ever been evaluated by anyone outside the vendor.
• Any neutral accuracy corpus for e-discovery AI: There is none. NIST's TREC Legal Track ran 2006–2011 and has no successor, and the one benchmark-shaped exercise we found in 2026 was run by an e-discovery vendor with undisclosed funding. Every accuracy figure in that market is self-reported on a private dataset.

None of this argues against adopting AI in a law firm. It argues for adopting it the way the constraints actually permit: one workflow at a time, with the diligence file assembled before the pilot rather than after, with the verification step budgeted as permanent work, and with the boundary written down in the regulator's language rather than the vendor's.

And it argues for holding the line that makes the whole exercise defensible. An agent may research, draft, extract, summarise, chase and queue — which is a great deal of real work, honestly done. What it may never do is make the substantive legal determination, communicate the advice, or put a document in front of a court. A lawyer does that, personally, every time. That is not a limitation of the current generation of tools. It is the job.