Executive Summary
Google Cloud launched Gemini Enterprise for Legal in preview on 25 August 2026, with Cleary Gottlieb, Freshfields, Weil, and Williams & Connolly named as partners. It joins Harvey, Legora, Thomson Reuters’ CoCounsel and Anthropic in a market that did not meaningfully exist three years ago.
The more consequential number sits on the other side of the table. The share of general counsel reporting generative AI use inside their own teams has gone from 20 percent in 2023 to 87 percent in 2026. Corporate legal departments have already automated a good deal of their own contract review and research. They know what that work now takes.
The argument of this report. Because clients moved first, the repricing of legal work is already underway and is not contingent on what any individual firm does. A firm that delays adoption does not avoid the compression — it experiences the compression without the productivity gain. Adoption is not what causes the squeeze. It is what determines whether the efficiency stays with the firm or passes to the client.
Disclosure: the author works in technology leadership at a law firm. This analysis reflects publicly reported figures and the author’s own views, not those of any employer or client.
What Google Actually Announced
Gemini Enterprise for Legal entered preview alongside a financial-services equivalent, with healthcare and life sciences flagged as next. It comprises legal-specific skills spanning contract review, playbook creation, regulatory scanning, legal research and data-subject-access-request fulfilment; secure connectors that inherit existing document permissions; pre-built agents that execute work autonomously; and a partner ecosystem.
The connector list is the useful part for anyone evaluating it. It reaches into iManage and NetDocuments for document management, Everlaw and RelativityOne for e-discovery, Docusign for contract lifecycle, Thomson Reuters HighQ and the Free Law Project’s CourtListener for research. Critically, it also integrates Harvey, Legora and Solve Intelligence rather than competing with them. For a firm that has already invested in one of those platforms, this is additive rather than a rip-and-replace decision.
No pricing, performance metrics or ROI claims accompanied the announcement — worth noting in a sector where most vendor claims remain marketing artefacts rather than measurements. Firms evaluating any of these products should expect to generate their own baselines.
The Clients Moved First
The FTI Consulting and Relativity General Counsel Report puts generative AI use inside corporate legal teams at 87 percent in 2026, against 44 percent the prior year and 20 percent in 2023. That is close to full adoption in three years, and it happened on the buying side.
This sequence is the single most important fact for firm strategy, and it is frequently read backwards. The concern one hears most often is that adopting AI will cannibalise billable hours. But a general counsel who has automated first-pass contract review in her own department already knows roughly what that work costs. Her expectations have moved whether or not her outside counsel has bought anything.
Put plainly: the price pressure is exogenous. It arrives from the client relationship, not from the firm’s own technology decisions. Declining to adopt does not hold the old pricing in place; it simply means meeting new expectations with old costs.
Two Responses to the Same Client Behaviour
A law firm’s revenue is approximately hours billed, multiplied by realised rate, multiplied by leverage — the ratio of fee-earners to partners. Every term is a quantity of human time. That is genuinely unusual. A manufacturer that halves labour per unit improves gross margin because it sells the unit; a software company that ships faster keeps the same subscription at lower cost. In both, efficiency accrues automatically to the seller.
In hourly professional services it does not accrue automatically to anyone. It has to be captured deliberately — through pricing, scope or capacity — or it dissipates. That is the real lesson of the structure, and it is a reason to move early rather than a reason to hesitate.
Consider the two columns above. In both, the client already knows the task can be done in a fraction of the time and negotiates accordingly. The firm that has not adopted meets the new price with the old cost base and absorbs the difference. The firm that has adopted meets the same price with a matching cost base, and keeps the released capacity to deploy elsewhere. The compression is identical. Only the margin outcome differs.
What Adoption Actually Buys
The defensive case is the weaker half of the argument. The positive case is more interesting, and it has four parts.
Fixed fees stop being a gamble. Alternative fee arrangements have been discussed for two decades and resisted for a simple reason: quoting a fixed price on work of uncertain duration transfers risk to the firm. When the variance in delivery time collapses, that risk collapses with it. A firm confident it can complete a matter in a known window can price it profitably — and a fixed fee quoted with an accurate cost base is a higher-margin instrument than an hourly bill, not a lower one.
Released capacity is the point, not the saving. Thirty-six hours recovered from a diligence exercise is thirty-six hours available for matters a firm previously could not staff, clients it could not take, or depth on premium work it had to ration. Firms that treat AI as a cost-reduction exercise will book a small saving. Firms that treat it as capacity creation have a growth instrument.
It is becoming a procurement requirement. Panel reviews and RFPs increasingly ask what technology a firm uses and how it affects delivery. This is shifting from differentiator to table stakes, and the firms with a substantive answer — deployed tools, measured outcomes, trained lawyers — are better placed than those with an intention.
The profession is already ahead of most industries. Anthropic’s associate general counsel Mark Pike told Bloomberg that outside software developers, lawyers are the single most active professional group on Claude — at, in his description, basically the highest rate of any other profession. A company webcast on legal applications drew more than 20,000 registrations. Whatever hesitancy exists at institutional level, individual lawyers have already decided. Firms are choosing between sanctioned, supervised deployment and unsanctioned use they cannot see.
Five Competitors, One Customer Base
Harvey raised $200 million in March 2026 co-led by GIC and Sequoia at an $11 billion valuation, on roughly $190 million of annual recurring revenue reported in January, up from $100 million in August 2025. It reports more than 100,000 lawyers across 1,300 organisations, including a majority of the AmLaw 100. It has since been reported in talks at around $15 billion.
Legora, founded in Stockholm in 2023, is growing fastest. It reported $150 million of ARR in the second quarter of 2026, up 50 percent, with its customer base rising roughly a quarter in three months to 1,500 law firms and in-house teams across more than 50 markets — Linklaters, White & Case, Dentons, Goodwin, Deloitte and Cleary Gottlieb among them. Valued at $5.6 billion in the spring, it was reported in August to be in early talks at $10 billion or more.
Thomson Reuters’ CoCounsel reached one million professional users across 107 countries by February 2026, with an agentic version grounded in Westlaw and Practical Law entering beta. Its advantage is proprietary content — the corpus a general model cannot lawfully reproduce.
Anthropic released legal tooling and a Westlaw integration in May 2026, with Freshfields, Quinn Emanuel, Holland & Knight and Crosby Legal reported as using Claude on live matters. Claude Opus 4.7 scored 90.9 percent on Harvey’s BigLaw Bench.
For buyers, the competition is good news. Two specialists roughly doubled their paper valuations within four months, which tells you how contested this is — and contested markets discipline pricing. Note also that Cleary Gottlieb appears twice in this landscape, as a Google launch partner and a Legora customer. Large firms are not selecting a single vendor; they are running several and letting practice groups find what works. That is a sensible posture while the layer that ultimately captures the margin — model, application, content or infrastructure — remains genuinely unresolved.
What Has to Be Managed
Encouraging adoption is not the same as encouraging carelessness. Three risks are real, and all three are manageable with deliberate design.
Verification is a professional obligation, not a preference. Fabricated citations have appeared in filed court documents, with sanctions following. In a profession where an invented case is a conduct matter, review protocols are not optional overhead — they are the cost of using the tool properly. Firms that build verification into workflow from the start avoid the incidents that set programmes back a year.
Protect the training pipeline deliberately. Junior work is how senior judgment is manufactured. A firm that automates the bottom of the pyramid without redesigning how associates learn may find in a decade that it has no partners. There is useful evidence on exactly this. In a randomised trial published in PNAS, students given an unguarded chatbot during practice performed 48 percent better while it was available and 17 percent worse on an unaided exam afterwards. Students given the same model configured to provide hints rather than answers gained more during practice and lost nothing. This publication examined that study in detail. The implication for firms is concrete: deploy tutor-mode configurations to trainees and answer-mode to experienced lawyers who can evaluate the output. The difference is a prompt design choice, and it is worth making explicitly.
Capture the efficiency rather than discovering it was given away. Reporting on the 2026 procurement cycle describes AI-related discounts appearing in tenders and clients scrutinising time entries against the automation firms describe. Whatever one thinks of that, it argues for firms getting ahead of the conversation with their own pricing proposals rather than defending hourly totals after the fact.
A Claim Worth Correcting
It is widely repeated that Anthropic has published research finding law firms uniquely exposed because of their linear structure. That does not survive checking, and firms making strategy decisions should know it.
Anthropic’s Economic Index — its published research programme on AI and work — contains no legal-industry finding of that kind. Its January 2026 report is dominated by computer and mathematical occupations, at 34 to 52 percent of observed usage depending on platform. Legal does not feature among the leading categories. Its only prominent legal reference is methodological: the task “review legal publications and perform database searches to identify laws and court decisions relevant to pending cases” is modelled as requiring 17.7 years of education, because it resembles work done by lawyers rather than the legal secretaries who often perform it.
The accurate position is the usage observation cited earlier — lawyers as the most active profession on Claude after developers — combined with a long-standing and well-understood body of law-firm economics. The argument holds. Its provenance has been misattributed, and a firm citing a non-existent study in a partnership meeting is unnecessarily exposed.
What This Analysis Does Not Establish
No firm has disclosed the effect on its accounts. Law firms are overwhelmingly private partnerships and publish no financials. There is no equivalent of the SEC filings this publication used to test similar claims about enterprise AI spending, where we found returns widely asserted and nowhere demonstrable. Every economic claim here rests on survey data, vendor disclosure and trade reporting.
Vendor figures are self-reported. Harvey’s and Legora’s ARR, CoCounsel’s user count and benchmark scores come from parties with an interest in them. ARR is not audited revenue, “users” is not a defined term, and BigLaw Bench is owned by a market participant. Both valuation discussions are reported as in progress and may not close at the figures cited.
The 40-to-4 ratio is illustrative. It reflects compression described in industry reporting and is used in the schematic to make the arithmetic legible. It is not a measured sector average, and real variation across practice areas is wide — litigation strategy and novel transactional work compress far less than diligence and first-draft production.
Adoption is not impact. Eighty-seven percent of general counsel reporting generative AI somewhere in their team is a low bar. It says nothing about how much work is displaced or what it saves.
Demand may expand. If cheaper legal services bring in matters that were never worth pursuing and clients who never retained counsel, firms could bill fewer hours per matter across many more matters. This is plausible and unproven, and it would make the outlook considerably better than the arithmetic alone suggests.
What to Watch
Whether fixed fees actually arrive. Alternative fee arrangements have been imminent for twenty years. What is new is that both sides can now estimate delivery time. Watch the share of matters let on fixed fee, not the number of panels discussing it.
Associate hiring. Entry-class sizes are published and are the cleanest observable signal of whether firms are restructuring the pyramid or simply adding tools to it.
Whether Harvey and Legora grow into their valuations. Roughly $190 million and $150 million of ARR against $11 billion and a mooted $10 billion assumes the application layer captures the margin. If Google’s substrate strategy works, or firms build directly on frontier models, that assumption is wrong — and buyers benefit either way.
The first firm to publish real numbers. A partnership disclosing realisation rates, leverage ratios and technology spend against outcomes would advance this debate more than any vendor benchmark. Whichever firm does it first will also own the narrative.
Conclusion. Google’s arrival confirms what Harvey, Legora, Thomson Reuters and Anthropic already concluded: legal work is the most attractive professional-services target available. For firms, the useful reading is not that this is a threat to be delayed. Clients adopted first, so the repricing is already in motion and no firm’s caution will stop it. What remains within a firm’s control is whether it meets new client expectations with a matching cost base and released capacity, or with the old one. The evidence also suggests the profession is further along than its institutions sometimes acknowledge — lawyers are already among the heaviest users of these systems anywhere. The opportunity is to make that use deliberate, supervised and properly priced. Firms that do will find this a good decade. The risk is not moving too fast; it is arriving late to a conversation clients have already started.
Sources
Google Cloud. Introducing Gemini Enterprise for Legal, 25 August 2026 — components, named firms and connector ecosystem. https://cloud.google.com/blog/products/ai-machine-learning/introducing-gemini-enterprise-for-legal
Harvey. Harvey Raises at $11 Billion Valuation, March 2026, and CNBC coverage of the round. https://www.cnbc.com/2026/03/25/legal-ai-startup-harvey-raises-200-million-at-11-billion-valuation.html
Financial Times reporting, August 2026, on Legora’s $150 million ARR, 1,500 customers and talks at a $10 billion-plus valuation, and on Harvey’s reported $15 billion discussions.
Bloomberg and Fortune, May 2026 — Anthropic’s legal product release, Mark Pike’s comments on lawyer usage of Claude, BigLaw Bench score and named firms. https://fortune.com/2026/05/12/anthropic-legal-plug-in-release-claude-cowork-big-law/
Anthropic. Economic Index report: Economic primitives, January 2026 — cited for what it does and does not contain regarding legal occupations. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report
Bastani, H., Bastani, O., Sungu, A. et al. Generative AI without guardrails can harm learning. PNAS, 2025 — the guarded-versus-unguarded finding applied here to trainee development. https://www.pnas.org/doi/10.1073/pnas.2422633122
FTI Consulting and Relativity. General Counsel Report 2026 — generative AI adoption among corporate legal teams.
Thomson Reuters — CoCounsel user figures and the agentic CoCounsel Legal beta grounded in Westlaw and Practical Law.
Law.com and Thomson Reuters Institute, 2026 — alternative fee arrangements and the 2026 procurement cycle. https://www.law.com/americanlawyer/2025/12/18/its-real-now-with-law-firm-ai-use-on-the-rise-expect-alternative-fee-arrangements-to-pick-up-steam-in-2026/
