Legal: what AI is really doing

What's actually changing

1. 1

   Legal AI vendors go from research assistant to agent that produces the deliverable

   How it works. Thomson Reuters rebuilt CoCounsel on Anthropic's Claude Agent SDK so it plans and executes across research, drafting and verification in one workflow instead of answering single questions; Ironclad and LinkSquares shipped agents that draft, redline and route contracts end to end rather than just flagging risk for a human to fix. The mechanism is the same across vendors: an agent chains retrieval, drafting and citation-checking steps and only stops for a human at defined checkpoints.

   Why it matters. Work that used to be billed as discrete associate or paralegal hours (first-draft research memos, first-pass redlines) is increasingly produced in one pass, which is what is putting pressure on hourly billing and junior staffing at the same time.

   What's overstated"Agentic" does not mean unsupervised. Every one of these launches, including CoCounsel's, still requires a lawyer to open and check the citations before filing; the product claim is speed to a defensible first draft, not a replacement for review.

   Thomson Reuters (CoCounsel)IroncladLinkSquaresHarveyLegora

   4 receipts

   • Thomson Reuters' next-generation CoCounsel Legal reached general availability in the US in August 2026, built on Anthropic's Claude Agent SDK. fully agentic AI experience designed to help legal professionals move from research and issue analysis to trusted work product within a single workflow Thomson Reuters, 20 Aug 2026 ↗
   • Thomson Reuters counts one million professionals in 107 countries and territories on CoCounsel overall; the next-generation version launched in the US only. One million professionals across 107 countries and territories now rely on CoCounsel Thomson Reuters, 20 Aug 2026 ↗
   • LinkSquares launched what it calls the first all-agentic CLM platform on 2026-05-05, rebuilt so AI executes work inside the system rather than just generating suggestions. We took a different approach by rebuilding the platform so that AI can operate across the entire contract lifecycle, planning and executing work inside of the system instead of generating outputs that someone has to act on. PR Newswire, 5 May 2026 ↗
   • Ironclad's August 2026 procurement agent release adds obligation extraction and precedent-based redlining across the whole contract lifecycle, not just review. AI Obligation Extraction makes renewal windows, discounts, credits, rebates, termination rights, and payment terms visible and trackable. PR Newswire, 5 Aug 2026 ↗
2. 2

   Legal-AI funding turned into a two-company arms race priced in the billions

   How it works. Harvey went from $11B in March 2026 to $15.5B in September 2026; Legora reached $8.8B in an October Series E, seven months after its March Series D, and says its annual recurring revenue now tops $200 million. Harvey says the money goes to expanding agent deployments and embedded legal engineering teams; Legora says it will expand across North America.

   Why it matters. This is now one of the most capital-dense corners of enterprise AI outside foundation-model labs themselves, which means law firms are being courted as design partners and data sources, and smaller vendors without a war chest will struggle to compete on model access or sales headcount.

   What's overstatedValuation is not the same as proven margin improvement for law firms: Legora's $8.8B price is up to 44 times the "over $200 million" ARR it reports, and Harvey's own materials describe the earlier capital as funding "embedded legal engineering teams", which suggests still-heavy services support behind the software.

   HarveyLegora

   3 receipts

   • Harvey raised $200M at an $11B valuation in March 2026 to expand agent deployments and embedded legal engineering teams. The investment will be used to expand the agents customers run on Harvey and grow the embedded legal engineering teams supporting them globally. Harvey, 25 Mar 2026 ↗
   • Harvey raised a further $550M at a $15.5B valuation in September 2026, co-led by Diffusion and Lightspeed. Today, Harvey announced a $550M funding round at a $15.5B valuation, co-led by Diffusion and Lightspeed Venture Partners. Harvey, 9 Sept 2026 ↗
   • Legora raised a $550M Series E at an $8.8B valuation in October 2026, seven months after its Series D. Legal AI startup Legora announced Wednesday that it raised a $550 million Series E round, valuing it at $8.8 billion. TechCrunch, 21 Oct 2026 ↗
3. 3

   AI hallucination sanctions became a routine, tracked category of court order

   How it works. A publicly maintained database has documented over 1,148 cases of lawyers submitting AI-hallucinated case citations or quotes to courts as of mid-2026, with penalties ranging from public reprimand to five-figure monetary sanctions to disciplinary referral and, in Nebraska, an interim law license suspension. Courts are explicit that verification duties are non-delegable: a filing cannot contain a citation the filing attorney has not personally read.

   Why it matters. This is now a malpractice and bar-discipline risk category firms have to build process around, not a one-off embarrassment; it is also reshaping bar guidance on what "supervision" of an AI tool has to look like.

   What's overstatedThe sanctions are for lawyers who did not verify AI output, not proof that legal AI tools are unusable; several cases (e.g. the Seventh Circuit's Dec v. Mullin) resulted in no further sanction once the court judged the error unintentional and quickly corrected.

   State and federal courtsState bar associations

   3 receipts

   • As of mid-2026 a tracking database has documented more than 1,148 cases of lawyer AI-hallucination citations in US courts. a researcher who maintains a database tracking Gen AI-related court orders in the United States has documented over 1,148 cases of hallucinations by lawyers. Norton Rose Fulbright, 1 June 2026 ↗
   • In Whiting v. City of Athens, Tennessee (March 2026), the Sixth Circuit imposed punitive sanctions after finding over two dozen fake citations and misrepresentations of fact. $15,000 each in punitive sanctions Norton Rose Fulbright, 1 June 2026 ↗
   • In the Seventh Circuit's Dec v. Mullin (decided March 2026), the court declined further sanctions after an attorney's AI-hallucinated citations were deemed unintentional and quickly corrected, but stressed lawyers must verify citations themselves. petitioner's counsel cited two non-existent cases and included a false quotation in a "standard of review" section of an appellate brief National Law Review, 2 June 2026 ↗
4. 4

   Courts and bars propose to write AI verification into the rules, not just the guidance

   How it works. California's Supreme Court asked the State Bar in August 2025 to consider writing its 2023 generative-AI guidance into the Rules of Professional Conduct, and to consider further guidance on agentic AI. The Bar's ethics committee (COPRAC) put amendments to Rules 1.1, 1.4, 1.6, 3.3, 5.1 and 5.3 out for public comment in March 2026 and, after revising several of them, approved a second comment round in June. The draft adds a duty to verify that no cited authority is fabricated, misstated or taken out of context before it reaches a tribunal. Courts are moving the same way: the Tenth Circuit proposed a rule that would make anyone filing with generative AI certify that a human checked the citations, quotations and analysis first.

   Why it matters. The obligation moves from "disclose AI use" to a written, signed duty to verify AI output before it goes anywhere near a court, which changes what a firm's internal AI policy and its filing checklist have to cover.

   What's overstatedBoth are still proposals, not black-letter requirements, and the Fifth Circuit dropped a similar certification rule in 2024 after lawyers argued existing rules already covered it; none of this treats AI use itself as prohibited, the entire thrust is process.

   State Bar of CaliforniaU.S. Court of Appeals for the Tenth Circuit

   7 receipts

   • The California Supreme Court directed the State Bar to consider additional guidance in light of agentic AI tools that can autonomously perform tasks or workflows without human prompting. The Court also directed the State Bar to consider any additional guidance in light of agentic artificial intelligence tools, which can enable systems to autonomously perform tasks or workflows without human prompting. State Bar of California, 13 Mar 2026 ↗
   • The proposed amendments add a duty of candor requiring lawyers to verify that no AI-assisted citation is fabricated before it is submitted to a court. a lawyer’s duty of candor toward the tribunal includes the obligation to verify the accuracy and existence of cited authorities, including ensuring no cited authority is fabricated, misstated, or taken out of context, before submission to a tribunal State Bar of California, 13 Mar 2026 ↗
   • COPRAC approved the proposed amendments to Rules 1.1, 1.4, 1.6, 3.3, 5.1 and 5.3 for a 45-day public comment period on 13 March 2026. At its March 13, 2026, meeting, COPRAC approved the proposed amendments to 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 for a 45-day public comment period. State Bar of California, 13 Mar 2026 ↗
   • COPRAC revised the proposals after the first comment round and on 12 June 2026 approved the modified versions for a second round of public comment. COPRAC made several changes to the proposed rules in response to public comments received during the first public comment period, which closed on May 4, 2026. At its June 12, 2026, meeting, COPRAC approved the modified rule proposals for a second round of public comment. State Bar of California, 12 June 2026 ↗
   • The second comment round on the revised proposals closed on 6 August 2026. Deadline: August 6, 2026, 11:59 p.m. (45 days) State Bar of California, 12 June 2026 ↗
   • The Board of Trustees approved COPRAC's changes to the State Bar's practical guidance on generative AI in May 2026, separately from the proposed rule amendments. COPRAC also proposed changes to the above-mentioned Practical Guidance, which were approved by the Board of Trustees at their May 2026 meeting. State Bar of California, 12 June 2026 ↗
   • The Tenth Circuit proposed a rule requiring a certification that a human reviewed any filing prepared with generative AI. The U.S. Court of Appeals for the Tenth Circuit proposed a new rule September 18 that would require lawyers and self-represented litigants to certify that a human reviewed filings prepared with generative AI before they are submitted to the court. Law Commentary, 23 Sept 2026 ↗
5. 5

   Brussels quietly gave courts and law firms an extra 16 months before high-risk AI rules bite

   How it works. The EU's Digital Omnibus, in force from 2026-07-27, pushed the AI Act's high-risk obligations (the category covering AI used by judicial authorities to research and interpret facts and law) from an August 2026 start date to December 2027 for stand-alone systems. Internal law-firm research and drafting tools generally fall outside the high-risk category altogether; the classification is aimed at AI courts themselves deploy.

   Why it matters. Vendors selling into courts and judicial bodies now have over a year longer before conformity assessments, logging and registration duties apply, while law firms buying commercial research/drafting tools were never squarely in that high-risk bucket to begin with, softening one of the loudest regulatory deadlines legal-tech had been bracing for.

   What's overstated"The EU AI Act regulates legal AI" is an overstatement for most of what law firms actually buy; the high-risk trigger is specifically judicial-authority use, and even that got delayed.

   European CommissionEU judicial authorities

   1 receipt

   • The Digital Omnibus moved the AI Act's high-risk start date for stand-alone systems from August 2026 to December 2027. The high-risk rules that were due on 2 August 2026 now start on 2 December 2027 for stand-alone systems and on 2 August 2028 for AI built into regulated products. Legal Desire, 23 Sept 2026 ↗
6. 6

   Clients take routine work in-house while the billable hour holds

   How it works. The predicted switch from hourly billing to flat fees for AI-accelerated work has not shown up in the invoices: small and mid-size firms sent a smaller share of flat-fee bills this year than last, and Above the Law's read is that checking the AI draft, and the extra work AI turns up, absorb the time saved. The pressure is arriving from the client side instead: in-house legal teams use generative AI to do routine work themselves and send less of it to outside counsel at all.

   Why it matters. A firm can keep its hourly rate and still lose the work: 54% of in-house teams that use outside counsel say AI has already cut what they send out (an Everlaw survey), and SignalFire says AI most compresses junior-associate work.

   What's overstatedThe in-house numbers come from a survey commissioned by Everlaw, a vendor that sells AI to those same teams, and its headline that outcome-based pricing will eclipse the billable hour within five years is a forecast, not a billing record.

   In-house legal departmentsEverlaw8am (MyCase, LawPay)AmLaw firms

   3 receipts

   • Flat fees fell to 15 percent of bills at small and mid-size firms, from 16 percent a year earlier. Flat fees made up 15 percent of all bills sent, which is down from 16 percent a year ago. Above the Law, 23 Oct 2026 ↗
   • Most in-house teams that use outside counsel say AI has already cut the work they send out. Among respondents whose departments use outside counsel, 54% say AI has already reduced the volume of work sent to outside counsel Business Wire via Yahoo Finance, 20 Oct 2026 ↗
   • AI's efficiency gains disproportionately compress work historically performed by junior associates. AI today disproportionately compresses the work historically performed by junior associates. SignalFire, 9 Mar 2026 ↗

Who's doing it

11 companies to know, 8 with a round we can source. Newest first where dated.

Search these companies in the startup database ↗ · Explore all research tools

Your move

Our read, built on the receipts above.

What's next

• 2 Dec 2027 EU AI Act high-risk obligations for stand-alone systems (including judicial-authority AI) take effect legaldesire.com ↗
• Dec 2026 Next-generation CoCounsel Legal, launched in the US in August, is expected in Canada, the UK and Australia later in 2026, Thomson Reuters said at launch. It named no month: the December date is our own year-end checkpoint, not a launch date Thomson Reuters announced thomsonreuters.com ↗
• 30 Nov 2026 Tenth Circuit publishes its final 2027 rules, including whether the AI human-review certification survives comment lawcommentary.com ↗