Anthropic and OpenAI moved into legal AI products as firms began metering AI costs

Thomson Reuters rebuilt CoCounsel on Anthropic's agent toolkit, OpenAI previewed a legal research model that it says Harvey and Legora will be able to build on, and Harvey raised $550 million. Inside law firms, AI spending is shifting from seats to usage meters, while in-house leaders report AI being cited to justify higher fees.

Editorial snapshot: 31 October 2026

The central finding

Legal AI now runs on a few model providers beneath competing legal brands, and firms are starting to meter its cost by usage. In Axiom's survey of in-house leaders, 59% had seen firms cite AI to raise fees, against just over 30% who had seen AI savings passed through.

01Key findings

What changed in legal AI in the third quarter of 2026

Capability

Legal AI products moved onto frontier models

Thomson Reuters made its next-generation CoCounsel Legal generally available in the U.S. on Aug. 20, built on Anthropic's Claude Agent SDK. On Sept. 17, OpenAI previewed Astra for Law for selected firms and said API customers including Harvey and Legora will be able to build on it.

Capital

Legal AI capital concentrated in Harvey

Harvey raised $550 million at a $15.5 billion valuation on Sept. 9, up from $11 billion in March, and bought Guardrails AI the same day. Most other rounds that legal-tech trackers listed for the quarter were Series A and seed deals of $3 million to $26 million.

Adoption and economics

Nearly all surveyed firms were using or exploring AI as clients reported higher fees

In the International Legal Technology Association's 2026 survey, 94% of firms were using or exploring generative AI, and 84% of firms with 700 or more lawyers were using or exploring Harvey, against 22% of firms under 50. In Axiom's survey of 510 in-house leaders, 59% had seen firms raise fees while citing AI as added value.

Accountability

Responsibility stayed with the lawyer who signs

Richards, Layton & Finger, one of Delaware's largest firms, asked the Court of Chancery not to sanction it for AI hallucinations in a brief, and a Los Angeles court set one sanction at $999.99, one cent below California's bar-notification threshold. Malpractice insurers reported no wave of AI claims; their policies generally exclude sanctions and fines.

Operating question

As AI moves from per-seat licenses to usage meters, who pays for each matter's AI use, and what evidence of savings or quality does that client receive?

02Capability

Legal AI products moved onto frontier and vendor-trained models

Three structural changes arrived in one quarter. Frontier labs moved inside legal products and alongside them; legal vendors ran more work on their own models, Harvey's trained to lower the cost of each task; and checking cited authority became a product feature. Together they make the model beneath a legal product a commercial choice that the vendor controls, with direct effects on cost and data handling.

Thomson Reuters · Generally available in the U.S., Aug. 20

CoCounsel Legal now runs on Anthropic's agent toolkit

The next-generation CoCounsel Legal is built on Anthropic's Claude Agent SDK and adds Westlaw Brief Builder, matter workspaces and a drafting agent in Word. Bulk extraction across up to 10,000 documents and 100 questions runs on Thomson Reuters' own legal model, called Thomson.

OpenAI · Preview, Sept. 17

OpenAI began offering grounded legal research to selected firms

Astra for Law pairs GPT-6 Astra with a legal search index of more than 230 million URLs and the Free Law Project's CourtListener, which covers more than 99.9% of published U.S. precedential case law. OpenAI describes the index as complementing content firms already license from providers such as Thomson Reuters. Harvey and Legora are named as future API customers.

Harvey · Research release, Aug. 20

Harvey trained its own model on open weights

Tenet, Harvey's first post-trained open-weight model, is built on Kimi K3 from China's Moonshot AI and was trained with reinforcement learning on about 1,750 agentic legal environments. Training rewarded the model for using fewer tokens, which is intended to lower the cost of each task.

Verification · Released in August and September

Checking cited authority became a product feature

Thomson Reuters' Deep Research Verify checks whether cited Westlaw and Practical Law authority supports specific legal assertions, and in August it reached Practical Law in the U.K. and Westlaw Canada. In OpenAI's release, Harvey's Niko Grupen and Sullivan & Cromwell's John Savva endorsed Astra for Law's grounding in cited sources.

Limitation: Thomson Reuters published no accuracy figure for the check.

Cost and consequence

The vendor now chooses the model beneath each legal task, and that choice sets much of its cost. On Artificial Analysis' legal benchmark, GLM-5.2 matched Claude Opus 4.8's score at about 15% of the cost. Buyers should ask which model each step of a product calls and what notice they receive when it changes.

03Capital

Harvey's $550 million round led a concentrated quarter for legal AI funding

Crunchbase counted $159 billion of global venture funding in the third quarter, 64% of it for AI companies. For legal technology, it counted more than $2.2 billion raised in 2026 through late August, including at least 12 rounds of $50 million or more, after a record $4.6 billion in 2025. Within the quarter, Harvey's round stood far above the other disclosed deals, most of them early-stage, and the two AI-native leaders spent on acquisitions.

Disclosed legal-tech rounds

Harvey's round against the quarter's other disclosed rounds

Disclosed amounts in USD millions for rounds announced July 7 to Sept. 22, 2026.

Harvey

Equity round — $550M

$15.5 billion valuation; co-led by Diffusion and Lightspeed

Norm Ai

Series C — $120M

Compliance and legal agents; $1.2 billion valuation; led by Khosla

Chamelio

Series A — $26M

Luminary

Series A — $22M

Blee

Series A — $20M

Jusfy

Series A — $15M

Newcode

Series A — $13.5M

Discern

Series A — $10M

Magnar

Series A — $8M

Aloi

Seed — $7M

Concorda

Seed — $3.8M

CopySight

Seed — $3M

Company announcements and legal-tech funding trackers. A compiled selection, not a complete count; it excludes undisclosed amounts, rounds in other currencies, including Noxtua's euro-denominated Series C, and some smaller seed rounds.

>$400M

Harvey and Legora combined annual recurring revenue, as cited by an analyst

Cited by a BNP Paribas analyst on RELX's July 23 results call, summarizing the companies' own statements.

$772M

Thomson Reuters Legal Professionals revenue, April to June alone

Up 10% organically. LexisNexis Legal reported £959 million for the first half, up 10% on an underlying basis.

Harvey · Sept. 9

Harvey's valuation rose from $11 billion to $15.5 billion

Diffusion and Lightspeed co-led the $550 million round, with Sapphire and Whale Rock joining as new investors. It followed an undisclosed July investment from Goldman Sachs Alternatives and J.P. Morgan Growth Equity Partners.

Acquisitions · July 29 and Sept. 9

The AI-native leaders bought capability

Legora agreed to buy Wexler, a London company that builds fact-intelligence tools for disputes, in its fifth acquisition of 2026; Legora said Wexler's team of 18 grew revenue eightfold in a year. On the day of its round, Harvey bought Guardrails AI, an AI security platform for agents, its fourth acquisition of 2026, Artificial Lawyer reported. Legora said more than 100,000 lawyers at 1,500 firms and in-house teams in more than 50 markets use its platform.

Strategic investors · July to September

Publishers and platforms invested in smaller players

Thomson Reuters Ventures joined Jusfy's $15 million Series A, and Relativity's Rel Labs backed Newcode's $13.5 million Series A. In Germany, Noxtua raised a Series C of more than €100 million, with publisher C.H. Beck as majority shareholder and MANZ as a minority holder.

What it signals

By revenue, the AI-native leaders are still smaller than the incumbent publishers. Their valuations assume growth into legal workflow software, which RELX's chief executive put at about $25 billion today, against about $5 billion for U.S. legal information, citing expectations of $75 billion to $100 billion within a decade. He said RELX will not lead that software market. Buyers signing multiyear contracts with AI-native vendors are buying into that expansion.

04Products

Competing legal AI products now share models, content and connectors

The quarter's releases show who supplies whom. Thomson Reuters' generally available product runs on Anthropic's toolkit; OpenAI plans to offer its legal model through its API to Thomson Reuters' rivals Harvey and Legora, while Thomson Reuters builds connectors into ChatGPT; Harvey moved capability onto an open-weight model. Connecting these tools to a firm's own documents became the main deployment work.

Release status of the quarter's main legal AI products

Status as announced by each company in 2026. No list prices were published for CoCounsel, Lexis+ with Protégé, Harvey or Astra for Law.

Release status of the quarter's main legal AI products
Product StatusDateComponents named by the vendor
CoCounsel Legal, next generation (Thomson Reuters) Status Generally available in the U.S. Date Aug. 20 Components named by the vendor Claude Agent SDK; Thomson model; Westlaw; Practical Law
CoCounsel Legal in Canada, the U.K. and Australia Status Expected Date Later in 2026 Components named by the vendor — No comparable figure
Syncly connectors (Thomson Reuters) Status Launched in the U.S., U.K. and Canada Date August Components named by the vendor iManage on premises and cloud; SharePoint
Astra for Law (OpenAI) Status Preview for selected firms Date Sept. 17 Components named by the vendor GPT-6 Astra; OpenAI legal search index; CourtListener
Astra for Law API (OpenAI) Status Announced as forthcoming Date — No comparable figure Components named by the vendor Harvey and Legora named as customers
ChatGPT for Word (OpenAI) Status Generally available Date Sept. 17 Components named by the vendor — No comparable figure
Tenet (Harvey) Status Research model; to reach products over time Date Aug. 20 Components named by the vendor Kimi K3 open weights, post-trained by Harvey

— No comparable figure.

Integration

Document management connectors became the integration work

Thomson Reuters released Syncly connectors to iManage, on premises and in the cloud, and to SharePoint, and expanded its CoCounsel connector for Claude; OpenAI launched 26 partner-built plugins, including Thomson Reuters, Relativity, Clio, iManage, Intapp and DeepJudge. Greenberg Traurig's Aug. 26 rollout of CoCounsel to about 3,200 lawyers in 51 locations stressed integration with DeepJudge, Westlaw and Practical Law.

Risk: Malicious servers using the Model Context Protocol, the connector standard Thomson Reuters uses with Claude, were reported in August.

Confidentiality

Data controls became product features

OpenAI offers eligible firms zero data retention on the Astra for Law API, and Latham & Watkins is working with OpenAI on information permissions and ethical walls. Firms in ILTA's survey ranked generative AI their second-biggest security challenge, the first time it appeared on the list.

Large firms

Elite firms are building tools on OpenAI's legal model

Astra for Law's launch firms each named a use: Sullivan & Cromwell, an agreement analyzer; Ropes & Gray, due diligence; Cooley, IPO preparation; and Skadden, regulatory risk. Wachtell is exploring uses.

LexisNexis

LexisNexis sold most new business on Protégé

RELX said on July 23 that about 90% of new sales value and about three-quarters of renewal value were on Lexis+ with Protégé, and that usage was growing faster than the number of users.

Contract terms

A single matter can now pass through a model lab, a content owner and an application vendor, each with its own retention, training and access terms. A firm's contract with the application vendor may not cover the layers beneath it, so those terms need to be named in the agreement.

05Adoption and economics

AI use by firm size, and what clients report about fees

Among the more than 500 firms in ILTA's 2026 survey, 94% were using or exploring generative AI, up from 80% a year earlier; first drafts, meeting summaries, legal research and long-document analysis were each cited by 73% to 79%. The difference between segments now lies in which products firms use. Clients report rising fees, and firms are moving AI costs onto usage meters.

Share of firms using or exploring each AI product, by firm size

ILTA 2026 Technology Survey, as reported by LawSites on Sept. 14: more than 500 firms with about 140,000 lawyers in about a dozen countries, including 137 firms under 50 lawyers and 59 with 700 or more.

Share of firms using or exploring each AI product, by firm size
Product All firmsUnder 50 lawyers700 or more lawyers
Microsoft 365 Copilot All firms 76%
Under 50 lawyers 59%
700 or more lawyers 88%
Claude All firms 44%
Under 50 lawyers 42%
700 or more lawyers 64%
CoCounsel All firms 44%
Under 50 lawyers 40%
700 or more lawyers 43%
Harvey All firms 43%
Under 50 lawyers 22%
700 or more lawyers 84%
Legora All firms 16%
Under 50 lawyers 6%
700 or more lawyers 47%
Lexis+ AI All firms 26%
Under 50 lawyers — No comparable figure 700 or more lawyers — No comparable figure

Using or exploring includes firms still evaluating a product. Size-band figures for Lexis+ AI are not available.

Larger firms · April to June data, published in August

Larger firms spent more on technology and earned more per lawyer

In Thomson Reuters Institute's index of Am Law 100, Second Hundred and midsize firms, technology and knowledge-management spending rose 11.6% overall and 8.7% per lawyer, which Thomson Reuters said reflects a renewed emphasis on AI and other technologies. For the Am Law 100, profit per full-time-equivalent lawyer grew faster than revenue.

+7.1%Worked rates, all index firms

+11.6%Technology and knowledge-management spending

Midsize firms · April to June data

Midsize firms' productivity was still contracting

Midsize firms in the same index were still seeing productivity contract in the second quarter, while Am Law 100 and Second Hundred productivity improved. For firms in that position, metered AI charges that cannot be tied to a client matter become overhead.

Firms under 50 lawyers · ILTA, 137 firms

Smaller firms rely on general assistants and research vendors

In ILTA's smallest size band, Microsoft's Copilot, Anthropic's Claude and the CoCounsel research product lead, and AI-native platforms are far less common. For these firms, the data and pricing terms of a general assistant and a research subscription matter most. The band does not separate solo practices.

In-house legal departments · Surveys published July and September

In-house teams expect gains and lack AI budgets

Corporate legal respondents to Thomson Reuters Institute were far more likely to expect productivity gains from their firms' AI than to say most firms deliver them, and almost one-third are reconsidering firms that cannot show AI-enabled value within 12 months. In Axiom's survey, AI mandates arrived without dedicated funding 92% of the time, up from 78% a year earlier.

70%Expect productivity gains from firms' AI (203 respondents)

6%Say most firms deliver them

Alternative providers

Alternative providers gained from budget pressure

More than 6 in 10 in-house leaders in Axiom's survey increased their use of alternative legal service providers in the past year, and 83% of chief financial officers encouraged moving work to alternative providers; 2% preferred established firms. Axiom is itself an alternative provider.

What in-house leaders saw in outside-counsel fees

Axiom survey of 510 in-house legal leaders and CFOs in six countries, released Sept. 8; 64% work at companies with $1 billion or more in revenue.

What in-house leaders saw in outside-counsel fees
Over the past year Share of respondents
Outside-counsel rates rose Share of respondents 85%
Firms raised fees, citing AI as added value Share of respondents 59%
Firms cut fees by passing AI savings through Share of respondents Just over 30%
Fees fell through fewer hours billed, which respondents linked to AI Share of respondents 4%

Axiom is an alternative legal service provider. The just-over-30% and 4% figures come from Law360 Pulse's report on the survey.

How firms pay · ILTACON, August

Firms' AI costs moved from seats to meters

Firm technologists at ILTACON in Nashville described a shift from per-seat subscriptions to usage-based charges metered by tokens, Thomson Reuters Institute reported. Speakers from Littler and Ogletree Deakins gave two reference points.

$2One PowerPoint in consumption charges (Littler)

About 1%Of revenue, where firms' AI spending is settling (Ogletree Deakins)

How firms charge

Metering turns AI into a billable line

One speaker argued that metering ties AI cost to specific clients and produces a line item that can be billed, and the conference report asked whether token spend could become a pass-through cost like e-discovery hosting. In a Thomson Reuters Institute survey published in July, almost two-thirds of law firm professionals said their pricing structure had not changed.

How vendors charge

Publishers kept subscription pricing

LexisNexis' legal revenue base is 85% subscription. RELX said token costs are under 1% of its overall cost base and it has no current plan to move to consumption pricing.

Who pays

If token costs become disbursements while hours are billed as before, the client pays on two lines; under a fixed fee, the firm absorbs the token cost as margin risk. Either way, the firm needs AI usage recorded against client-matter codes before it can decide.

06Outcomes

Drafting improved in a field trial; complete deliverables mostly failed a benchmark

The quarter's two strongest measurements describe different work. A randomized field trial measured assisted drafting by practicing patent lawyers, and an independent benchmark measured whether models could finish complete legal deliverables on their own. Firm rollouts announced in the quarter, including Greenberg Traurig's, published no outcome measures.

Field experiment · NBER working paper, September

AI assistance raised patent-drafting quality

About two-thirds of 133 lawyers at 11 U.S. IP firms that do business with Google were randomly given an unreleased Google patent-drafting assistant for three months; the rest received AI training without the tool. Quality rose because there were fewer poor drafts and more good ones; excellent drafts were no more frequent.

Conditions: Google's tool, Google-employed co-authors and a small sample.

+0.34Drafting quality at 10 days, standard deviations

+0.38Drafting quality at 90 days, standard deviations

18 minutesSaved by junior lawyers, against the control group's 124-minute average

Field experiment

Only senior lawyers gained unassisted skill

In an unassisted redlining test at 90 days, lawyers who had used the tool scored higher overall, with the gain coming entirely from those with seven or more years of experience. Junior lawyers showed no average gain, and their scores spread out.

+0.32Unassisted redlining, all lawyers, standard deviations

+0.45Unassisted redlining, senior lawyers

No average gainUnassisted redlining, junior lawyers

Vendor-reported

Vendors reported gains on their own tests

OpenAI reported that Astra for Law scored 54.0% on a 200-question validation set of the Vals Legal Research Bench at its highest effort setting, against 38.7% for GPT-6 Astra with web search. Harvey reported that Tenet passed almost twice as many held-out tasks on its benchmark as its base model, with no significant change on PRBench-hard, from 36.0% to 36.8%.

Complete legal deliverables fully passed, by model

Artificial Analysis' July 7 launch run of LAB-AA, a reimplementation of Harvey's Legal Agent Benchmark: 120 private tasks across 24 practice areas, each graded all-or-nothing against a binary rubric.

Complete legal deliverables fully passed, by model
Model Tasks fully passedReported cost or effort
Claude Fable 5 (max effort, Opus 4.8 fallback) Tasks fully passed 14.2%
Reported cost or effort About $18.9 per task
Claude Opus 4.8 Tasks fully passed 7.5%
Reported cost or effort About $8.2 per task
GLM-5.2 Tasks fully passed 7.5%
Reported cost or effort About $1.3 per task, 15% of Opus 4.8's
MiniMax-M3 Tasks fully passed 6.7%
Reported cost or effort — No comparable figure
Claude Sonnet 5 Tasks fully passed 5.0%
Reported cost or effort About 161 turns per task
13 of the 28 models tested Tasks fully passed 0%
Reported cost or effort — No comparable figure

A single AI judge, Gemini 3.1 Pro, graded the tasks. Harvey's own tools were not tested, and models released later in the quarter, including Claude Opus 5 and GPT-6 Astra, were not in the launch run.

Where the results hold

AI assistance improved drafts and helped experienced lawyers' own work, while complete deliverables still need a lawyer to finish them. In the trial, juniors who drafted with AI did not become better unassisted, so firms that staff matters on junior leverage need another way to build that skill.

07Corrections

Four claims from the quarter, corrected

Four statements a reader could take from the quarter's coverage, and the distinctions that change them.

Law firm demand · Not supported by second-quarter data

Second-quarter data showed law firm demand on pace for its strongest year since 2021.

Claim under review: AI efficiency is already shrinking demand for law firm work.

Thomson Reuters Institute's Law Firm Financial Index for April to June, published Aug. 10, showed associate demand up 4.3% and non-equity partner demand up 6.0%, while equity partner demand fell 1.2%. Leverage reached its highest level since late 2018. The index covers larger U.S. firms and the three months before the quarter began.

Copilot deployment · Survey label misread

The 52% full-deployment figure is for a product ILTA labeled Microsoft Copilot (formerly Bing Chat).

Claim under review: Half of law firms have fully deployed Microsoft 365 Copilot.

In ILTA's 2026 survey, 76% of firms were using or exploring Microsoft 365 Copilot. The 52% figure for full deployment across lawyers refers to the option labeled "Microsoft Copilot (formerly Bing Chat)," a distinction LawSites flagged, and does not measure deployment of the paid Microsoft 365 product.

Malpractice insurance · Wrong line of coverage

Berkley Select's lawyers' professional liability coverage has no AI exclusion; Berkley's absolute AI exclusion is on other lines.

Claim under review: Berkley has excluded AI from lawyers' malpractice coverage.

Lockton's July market update said Berkley's absolute AI exclusion could apply to its D&O, employment practices and fiduciary coverages. The Daily Record reported Aug. 25 that Berkley Select, which writes lawyers' professional liability, has no AI exclusionary language and asks no supplemental AI underwriting questions. One surplus-lines carrier does use an AI-use exclusion endorsement, and ALPS expects to ask more detailed AI questions.

Vendor revenue · Unconfirmed

Harvey's funding announcement gave no revenue figure, and press estimates differ.

Claim under review: Harvey's annual recurring revenue is more than $350 million.

Non-Billable reported annualized revenue of about $300 million on July 24. SiliconANGLE, citing The Information, reported more than $350 million on Aug. 7, up from $190 million in January. Harvey's Sept. 9 announcement said 80% of the Am Law 100 and five of the Fortune 10 use its product.

08Accountability

Sanctions and insurers kept responsibility with the lawyer who signs

Courts kept applying the duty that attaches to a lawyer's signature. What changed was who was caught, an elite Delaware firm, and how one sanction was set. Inside firms, the failures were as much economic as legal: unapproved tools, AI services that produced no revenue and expensive models used for simple work.

Incident · Reported July 15

AI hallucinations reached a leading Delaware firm

Richards, Layton & Finger, one of Delaware's oldest and largest firms, asked the Court of Chancery not to sanction it for AI hallucinations in a Jan. 22 brief and for an incomplete correction. A firm director said a human-review process was in place. The outcome was pending when Bloomberg Law reported the filing.

Sanction · Reported Aug. 21

A filer with 46 hallucinations was fined and referred to the bar

A Florida state appellate court fined an attorney whose filings contained at least 46 hallucinations and referred the attorney to The Florida Bar. Law360's headline called the AI use unapologetic.

Sanction · Reported Sept. 14

A $999.99 sanction fell one cent short of bar notification

A Los Angeles court sanctioned a State Farm attorney who cited nonexistent cases and fabricated quotations. The $999.99 amount sat one cent below the State Bar of California's notification threshold.

Insurance · Reported Aug. 25

Malpractice insurers see no claims wave and exclude sanctions

Lawyers' professional liability carriers told The Daily Record that AI appears in claims about citations and verification but has not produced a wave of claims. Policies generally exclude sanctions and fines. ALPS expects to ask more detailed AI questions, subject to state approval, and broker Lockton tells insureds to expect underwriters to ask which AI tools are approved and who may use them.

Governance · Reported July 21

More than a third of law firm professionals use unapproved AI

Thomson Reuters Institute's survey of 736 law firm professionals found more than one-third using AI tools their firms had not authorized. ALPS' chief operating officer said adoption might be moving faster than governance.

Cost control · ILTACON, August

Premium-model overuse and hidden usage raised AI costs

At ILTACON, Barnes & Thornburg's Mark Babcock described lawyers using premium models for simple tasks, a speaker from HIKE2 said major companies had switched AI services off because usage produced no revenue, and Lawfront's James Grice asked whether models were running up bills in the background that no one could see. Legaltech Hub now tracks more than 1,100 legal tool companies, a count growing more than 30% a year, and Lawfront inherited seven technology stacks from seven acquired firms.

Who is accountable

Courts kept applying the duty attached to a lawyer's signature, insurers place responsibility on the licensed lawyer, and malpractice policies generally exclude sanctions. A firm needs a record of which tools are approved, for whom and what verification ran before filing; Lockton tells insureds to expect underwriters to ask for the first two.

09Regulation

Colorado's lead-generation ban took effect; AI-specific rules were mostly deferred

The main binding rule to take effect in the quarter governed how consumers reach lawyers: Colorado's ban on paid lead generation, in force from Aug. 12. AI-specific measures were mostly deferrals and scheduled study, apart from the EU transparency duties that applied from Aug. 2. Two earlier 2026 measures also bear on the quarter: Arizona's March limits on referral-only ABS firms, which investors are now avoiding, and Alabama's ethics opinion on AI due diligence and billing.

Colorado · SB26-174 · In force Aug. 12

Colorado's ban on paid legal lead generation took effect

The law bars paying for, engaging in or selling leads from paid third-party lead generation legal marketing, whether priced per lead, per case, by subscription or through affiliates. Firm-identified search, pay-per-click, broadcast, billboard and directory advertising remains lawful, with exceptions for licensees, people working for them with the firm clearly identified, and nonprofits.

Colorado · Enforcement

Firms that buy leads carry the liability

Violations are deceptive trade practices under the Colorado Consumer Protection Act, with a private action for $10,000 per violation plus fees, injunctions and attorney general rulemaking. Greenspoon Marder reads the first prohibition as placing liability on lead purchasers and contrasts the remedy with the Telephone Consumer Protection Act's $500 to $1,500.

Arizona · Alternative business structures · Rules amended March 18; reported Sept. 16

Investors chose back-office deals over Arizona ownership

In March, Arizona amended its rules for alternative business structures, which permit nonlawyer ownership, to require ABS firms to provide legal services rather than referrals alone and to serve Arizona residents at least in part. In September, Holland & Knight's Pippa Balestrieri told Bloomberg Law that the vast majority of investors and firms now choose management services organizations over nonlawyer ownership.

Alabama · Formal Opinion 2026-01 · Issued before July; reported Aug. 4

Alabama's earlier opinion treats tool due diligence as part of competence

The State Bar's opinion, issued before the quarter and summarized by Jackson Lewis on Aug. 4, finds that AI creates no new duties but that deploying a tool without basic due diligence does not satisfy Rule 1.1. Before entering client data, lawyers should know whether prompts are retained or used for training, who can access them and what the vendor commits to on security. When AI shortens a task, lawyers may bill only for time actually spent.

U.S. federal courts · Expert panels scheduled Oct. 15

Federal AI evidence rules remained under study

The Advisory Committee on Evidence Rules withdrew proposed Rule 707 on machine-generated evidence in May for further study and did not approve publication of Rule 901(c) on deepfakes. Its agenda book for the Oct. 15 meeting schedules expert panels on both; no rule was adopted in the quarter.

EU and U.K. · EU duties applied from Aug. 2

EU transparency duties began; high-risk duties moved to 2027

AI Act transparency duties and enforcement of general-purpose AI model rules applied from Aug. 2, so client-facing legal chatbots in the EU must tell users they are dealing with AI. The omnibus amendment, in force since July 27, moved Annex III high-risk obligations to Dec. 2, 2027. Lewis Silkin does not expect a U.K. AI bill before the next King's Speech, late 2027 at the earliest.

For consumer legal services

The quarter's new binding obligation fell on intake businesses. In Colorado, an AI intake service paid per lead or by subscription is lawful only if it works for the firm and identifies that firm in its marketing, and the firm buying the leads carries the liability. Arizona's March changes already bar referral-only ABS firms, and investors are turning to management services organizations instead.

10Voices

Where suppliers, firms, clients and researchers agree and differ

The voices below agree on three things: the lawyer who signs remains responsible, AI becomes useful only when connected to a firm's own knowledge, and AI costs are rising and need control. They divide on who keeps the efficiency. Ballard Spahr's Lisa Mayo Haynes argues metered AI can be billed to clients, Bauer Media's Adrian Goss sees no AI efficiency in law firm costs yet, and Baker McKenzie's Sunny Mann wants the industry to make alternative fees work. They also divide on pricing, with RELX keeping subscriptions while firms that buy model access pay by the token, and on capability, where vendors' research scores run well ahead of the independent result for complete deliverables.

Information publishers

Thomson Reuters and RELX each grew legal revenue about 10% by selling AI inside research subscriptions. Thomson Reuters also builds connectors into OpenAI's ChatGPT while competing with OpenAI in legal research.

Raghu Ramanathan

President, Legal Professionals, Thomson Reuters

“The legal industry is moving beyond AI that simply generates answers”

From the Aug. 20 release launching next-generation CoCounsel Legal, built on Anthropic's Claude Agent SDK.

Incumbent vendor · product launch

Joel Hron

Chief Technology Officer, Thomson Reuters

“Legal professionals need more than access to information.”

Quoted in OpenAI's Sept. 17 Astra for Law announcement, in which Thomson Reuters brought HighQ into ChatGPT and previewed a CoCounsel Legal connector.

Incumbent vendor · quoted in OpenAI's release

Erik Engstrom

Chief Executive, RELX

RELX will keep growing its legal business by building AI into LexisNexis research and content, and does not expect to lead the much larger workflow and software market.

On the July 23 call he sized U.S. legal information at about $5 billion and legal software at about $25 billion, a market he said many expect to reach $75 billion to $100 billion within a decade.

Incumbent vendor · results call

Nick Luff

Chief Financial Officer, RELX

Protégé's low token use is a competitive advantage.

RELX's position contrasts with law firms that buy model access directly and pay by the token.

Incumbent vendor · results call

AI-native vendors

Harvey and Legora each made acquisitions in the quarter and position themselves as platforms for whole legal workflows.

Winston Weinberg and Gabe Pereyra

Co-founders, Harvey

Harvey wants to be the global partner legal teams turn to, and framed its $550 million round around helping them build and own their intelligence.

The announcement disclosed no revenue.

AI-native vendor · quoted in funding announcement

Max Junestrand

Chief Executive, Legora

“Every dispute and every investigation comes down to the facts”

His explanation for buying Wexler, Legora's fifth acquisition of 2026.

AI-native vendor · acquisition announcement

Law firm leaders and technologists

Firm technologists agree that AI costs are rising and hard to see. They differ on whether clients should see AI as a separate charge.

Lisa Mayo Haynes

Ballard Spahr; ILTACON 2026 speaker

Usage-based metering ties AI cost to specific clients and creates a traceable line item that can be billed.

Reported by Thomson Reuters Institute, which asked whether token spend could become a pass-through cost like e-discovery hosting.

Law firm · conference remarks, as reported

Johnny Wong

KM Counsel, Littler

One PowerPoint cost $2 in consumption charges, and firms need to ask whether they are getting that money back.

Law firm · conference remarks, as reported

Tim Fox

Chief Data and AI Officer, Ogletree Deakins

Most firms are settling on AI spending of about 1% of revenue, and Ogletree is not there yet.

Law firm · conference remarks, as reported

Sunny Mann

Global Chair, Baker McKenzie

The industry has to make alternative fee arrangements work collectively.

Said on a Bloomberg Law podcast in July, as reported by Above the Law.

Firm leader · podcast remarks, as reported

Clients

Bloomberg Law's Roy Strom reported Sept. 3 that a growing chorus of major Big Law clients, citing AI, say the billable hour is no longer sustainable. The in-house leaders below describe the budget side of that pressure.

Adrian Goss

Vice President, Legal and Governance, Bauer Media Group

He is not yet seeing AI efficiencies flow through from law firm costs.

Quoted in Law360 Pulse's coverage of Axiom's survey.

Buyer · reported by Law360 Pulse

Jimmy Yun

Director and Associate Counsel, Meta

Legal departments have to govern and evaluate AI out of budgets that are already stretched.

Axiom found AI mandates arrive without dedicated funding 92% of the time.

Buyer · reported by Law360 Pulse

Researchers

The quarter's strongest task study came with commercial ties to the tool it tested.

David Autor and co-authors

Autor: MIT and visiting fellow at Google; NBER Working Paper 35720

AI assistance improved patent-drafting quality, but only senior lawyers became better at the work without AI.

The trial used an unreleased Google tool and included Google-employed co-authors.

Researchers · randomized field experiment

11Decisions

Three decisions for firm and legal department leaders

Each decision follows from the quarter's findings on metered cost, unfunded client mandates and a supply chain with several layers.

Managing partner

Decide how metered AI costs reach the client

Consumption charges now tie AI cost to individual matters. In Axiom's survey, more in-house leaders had seen firms cite AI to raise fees than had seen AI savings passed through, and Alabama's bar opinion lets lawyers bill only for time actually spent when AI shortens a task. Choose, by client and matter type, among absorbing the cost, passing it through as a disbursement and pricing it into fixed fees.

Which clients will see AI cost on their invoices, on what basis, and what will they receive for it?

General counsel

Set what outside counsel must show before AI changes a fee

Most AI mandates reach legal departments without dedicated funding, and finance chiefs are pushing work toward alternative providers. Outside-counsel guidelines for 2027 can require firms to disclose AI pass-through charges and report time saved on matters where they cite AI as added value.

What evidence must a firm provide before AI appears on our bills as a cost or a premium?

Chief information officer

Decide which layers of the AI supply chain to depend on

A workflow can now run on one company's model, another's content and a third's application, and vendors can change the model underneath. Usage-based charges also mean cost follows each call. Map each workflow's layers and require usage data that can be tied to client-matter codes.

For each AI workflow, which companies handle client data, and can its cost be attributed to a matter?

12Actions

Four actions for the rest of 2026

Each action has one owner and a result that can be checked by year-end.

  1. 1

    Attribute AI usage to client matters

    Require AI vendors to report usage by user and client-matter code, and reconcile it with billing each month. The result is a per-matter AI cost the firm can price, absorb or pass through deliberately, measured against the roughly 1% of revenue that Ogletree Deakins' Tim Fox said most firms are settling on.

    Finance

  2. 2

    Audit paid lead sources against Colorado's law

    List every per-lead, per-case, subscription and affiliate source, including AI intake services, that delivers Colorado prospects without clearly identifying the firm. The result is a list of sources to end or restructure before they expose the firm to $10,000-per-violation private actions.

    Marketing and intake

  3. 3

    Build unassisted work into junior training

    Where juniors draft with AI, schedule periodic drafting or redlining without it and review the results. The result is evidence of whether juniors are building the unassisted skill that, in the patent-drafting trial, only senior lawyers gained.

    Practice group leaders

  4. 4

    Prepare AI answers for insurance renewal

    Document approved tools, who may use them and the verification step required before filing, and address unapproved use, which more than one-third of law firm professionals report. The result is ready answers to underwriters' AI questions and a record to show a court, since malpractice policies generally exclude sanctions.

    Firm general counsel

13References

References

Covers July 1 to Sept. 30, 2026, as of Oct. 31, 2026. Earnings, financial-index and survey data published in the quarter may describe earlier periods; the text gives each period.

Products and vendors

Research and benchmarks

Thoughtful engineering for what’s next.

Have a question about what this means for your team? Let’s talk it through.

Start a conversation

Have a project in mind, or just exploring an idea? Let’s talk.
Fill out the form or give us a call.

413-570-3299

What can we help you with?