A new law firm backed by the former global senior partner of Allen & Overy announced itself last week with its prices attached. Up to six contract reviews a month for €1,000, each one back the same day with a lawyer's name on it. It was one of four new-model firms in the news in four days.
THE LEAD PLAY
Six Contracts, €1,000, Back the Same Day
Falcon went public in Belgium on October 1. It is two entities. Falcon AI is a technology company that raised €1.5 million in pre-seed money, and Falcon Law is the firm it powers, which former Freshfields lawyer Rémy Bonnaffé runs as sole partner. Wim Dejonghe, the former global senior partner of Allen & Overy, is an adviser in the company's own release and a co-founder in most of the trade coverage. The release lists "three published plans starting at EUR 1,000 a month for up to six reviews and rising to EUR 5,000 a month for high-volume teams, with no overage fees." The work is NDAs, data processing agreements, master service agreements, and customer and supplier contracts. A client drops the document into Slack or Teams, and "a named, qualified lawyer then reviews and signs off before anything goes back to the client, the same day." Falcon says the same review runs €3,000 to €4,000 at a traditional firm and takes days or weeks. That comparison is Falcon's own, and it is doing some marketing. The €1,000 is on a price list. Use all six reviews and it comes to about €167 each.
Dejonghe's line in the launch coverage: "The traditional legal sector is due its neo-bank moment. AI has changed how quickly legal work can be done, but the traditional model still rewards firms for the time that work takes."
The other three made news the same week. In San Francisco, Arceus announced a $17 million round led by Greycroft. It handles commercial contracts at a flat price and promises that "standard reviews are guaranteed within eight hours or the review is free." Pierson Ferdinand, which has no associates at all, said it had reached 300 partners "less than three years after launching in January 2024." Its co-chairman, Michael Pierson, explained the staffing in a Harvey customer story: "In place of junior staff, we use AI to largely replace the tasks that would have ordinarily been performed by more junior lawyers." In Amsterdam, Jelle Drok, a Freshfields partner until recently, and Arvid van Oorschot, who spent more than six years there, opened Driven Lawyers, a disputes and IP firm on fixed fees with drafting software they developed themselves. Drok: "We'd rather be among the people building AI, than among the people holding meetings about it."
Some discounting is in order. Falcon has one named customer, a Belgian earplug company. An NDA is the easiest contract in the building, and fixed-fee contract shops existed before anyone had heard of a language model. Richard Tromans at Artificial Lawyer says there are now dozens of these firms. Nobody should assume most of them last. The price is not the part worth copying. What they had to build before they could print one is.
Falcon's release describes it: "Falcon starts every client relationship with a 45-minute intake call to understand how the company does business and what its risk appetite looks like. That conversation becomes the client's legal 'playbook,' the set of rules that guides every review Falcon does for them afterwards." Falcon AI's CEO, Amaury Gérard, told Artificial Lawyer what the firm watches: "The most important metric for us, is the time that a human spends reviewing a contract." Arceus gets to the same place from the other direction, with software it calls CounselOS that pulls "context on a company's business, preferences, agreements, prior decisions and negotiation history" out of HubSpot, Salesforce, Gmail and Slack.
So both firms wrote the client's positions down before the first contract arrived, and both treat lawyer minutes as a cost to drive down. At an hourly firm those positions live in one partner's head and last year's redline, and the only record of review time is the entry on the bill, which measures something else. Fixing that takes neither a startup nor a model. It is an intake problem. You ask the questions once, at the start, and write the answers where the next person can use them. The model is the part you can license. The page with the positions on it has to be written by someone who holds them.
The Play this week: Write one playbook and start a clock. Pick the contract your firm or department reviews most often. For most readers that is the NDA; pull last quarter's matters if you are not sure. Sit down for 45 minutes, the length of Falcon's intake call, with the lawyer who handles the most of them, and get one page: your position on each term that actually gets negotiated, the fallback you will accept, and the point at which it goes to a partner. For an NDA that is a short list: mutual or one-way, the definition of confidential information, term, residuals, non-solicitation, governing law.
Then time the next ten. The reviewer writes down minutes of attention from open to send, kept apart from whatever gets billed, and works from the page on every one, with the firm's AI tool or without it. After ten you have a standard that a second-year, a paralegal or a tool can apply the same way twice, and you have your own number for what the work takes. Falcon's works out to about €167 a review. You do not have to match it, but you should know yours before a client's procurement team does the same division.
In-house teams should write the same page for their own paper. It is what the intake call would get out of you anyway, and once it exists, outside counsel, an ALSP and a subscription firm are all bidding on the same instructions.
SECOND CHAIR
California's AI Statute Starts in 88 Days
Governor Newsom signed SB 574 on September 30, the last day before it would have become law without him. It is Chapter 858 and takes effect January 1, 2027. Harris Beach Murtha calls it the first law in the nation governing AI use by attorneys. New Business and Professions Code section 6068.1 says a lawyer "shall not delegate the practice of law to generative artificial intelligence" and may not put confidential or nonpublic information into a system unless access is "restricted to the attorney and persons authorized by the attorney under obligations to protect the confidentiality of the information." The lawyer must take reasonable steps to verify outputs, "including, but not limited to, the accuracy of all case and statutory citations." And the lawyer must disclose "the use of generative artificial intelligence to the court for all documents submitted to the court." The amendment to Code of Civil Procedure section 128.7 reaches past AI: a filing "shall not contain any citations that an attorney responsible for submitting the pleading has not personally verified, including any citation provided by generative artificial intelligence." Arbitrators get their own section. They "shall not delegate any part of their decisionmaking process to any generative artificial intelligence tool."
The read: The disclosure sentence is the one that needs work. The statute does not say what form the disclosure takes, where in the document it goes, or how much use counts as use. It defines generative AI as "an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio that emulates the structure and characteristics of the system's training data." That reads broadly enough to take in the drafting assistant in a word processor and the AI features in a research service along with the chatbots. A firm with California filings needs two things by January: a standard disclosure sentence in the pleading templates, and a written answer on what counts, so that forty lawyers do not make forty different calls. The confidentiality clause also bears on the personal-account question from last issue. It is hard to describe a free consumer login as a system restricted to the attorney and people the attorney has bound to confidentiality. Firms outside California should read section 6068.1 anyway. It is the first complete version of this statute, and it is the one the next legislature will have on the table.
OpenAI Sent Notices and Shipped Agents in the Same Week
On September 29, at its DevDay conference, OpenAI launched dots. The company's description: "always-on agents" that "have their own cloud computer" and "can readily connect to over 4,000 apps." They are rolling out now to Pro and Business Premium users. For Enterprise, the help page says the beta is "initially turned off by default" until a workspace admin enables it. Local computer access is a separate setting, also off by default, and a further permission governs adding dots to Slack and Microsoft Teams.
The surrounding days went like this. On September 28 Florida's attorney general asked a state court for a temporary injunction barring OpenAI from developing new models without independent oversight. On September 30, according to The Register, OpenAI updated its Hugging Face investigation and "said it has notified more than 100 organizations that 'misaligned models' may have accessed their systems." California's attorney general served an investigative subpoena the same day and announced it on October 1. Rob Bonta's statement: "Developers that fail to do so can and should be held legally accountable, and my office is committed to determining if that is the case here."
The read: There are two jobs here. Find out whether your firm or a client received one of those notices, and which inbox it would have reached. The Australian government's went to a public mailbox, as we covered last issue. If your website or client portal sits with a vendor, ask the vendor too. Then check which ChatGPT plan the firm is on. An Enterprise workspace has an admin switch and it starts off. The help page describes no equivalent default for Business Premium, where dots are simply available, so a small firm on that plan may already have a lawyer connecting an always-on agent to an inbox. Before anyone turns on the Enterprise beta, read two sentences from OpenAI's documentation that Rajesh Beri flagged on October 2: "During the Enterprise beta, dots do not support data residency or inference residency," and "Cloud orchestration events do not reach your existing OpenTelemetry collector." An agent that works around the clock inside firm systems, with a hole in the log, is a records problem on its first day.
The Bench Has a Reading Tool, and Clio Owns It
Clio has bought Learned Hand, which builds an AI workspace for judges and clerks. LawSites describes the product as tools "designed to help them manage crushing caseloads by organizing case materials, flagging when lawyers bend the truth, and drafting bench memos and orders." Artificial Lawyer reports it is already used by the Superior Court of Los Angeles County, the nation's largest trial court. When the Los Angeles Times described the Los Angeles pilot in February, the plan was for half a dozen civil judges to use it on summary judgment motions and class settlement approvals, and to "even draft tentative rulings." The same story said "Judges would not have to disclose whether they used the program." Clio began building its judiciary business in August. Terms were not disclosed.
The read: Set this beside SB 574. From January a California lawyer has to tell the court when generative AI was used on a filing. The court has no matching duty. The bill's only word on judges is an instruction that the Judicial Council "publicly revisit, and revise as necessary" its standard on generative AI. So at the country's largest trial court, a summary judgment brief can be read first by a tool that, in LawSites' words, flags when lawyers bend the truth, and that tool can draft the tentative. A brief's account of the cases and the record gets checked by machine whether or not every citation is real, which is the problem Issue #18 was about. And a firm on Clio now buys its practice management from the company that sells the bench its reading tool. There is nothing wrong with that today. It belongs in the vendor file.
STILL WATCHING
Thomson Reuters v. ROSS Intelligence is decided. The Third Circuit affirmed for Thomson Reuters on September 29, 110 days after argument, in a precedential opinion by Judge Montgomery-Reeves. It was issued under seal and made public the next day. The court held that Westlaw's headnotes are original enough for copyright and that ROSS's use of 2,243 of them was not fair, with three of the four factors against ROSS, market harm among them. It calls the dispute "no more than an ordinary copyright case," and a footnote separates it from the generative cases: "Unlike the AI models in Bartz and In re: OpenAI, ROSS's AI platform cannot generate original expression." ROSS, now shuttered, says it will ask the Supreme Court to take the case.
The California rules amendments. A correction to our count. We have been measuring from the first comment period, which closed May 4. There was a second round on revised proposals for Rules 1.1, 1.4, 1.6, 3.3, 5.1 and 5.3, and it closed August 6, sixty days ago. The State Bar's page still shows the package at the comment stage.
Lisandrillo v. Palozzi. Counsel has answered the Fourth District's show-cause order and asked that it be discharged. Per Eugene Volokh's October 2 update, her response reviews about 237 citations across four filings and reports two errors, which it calls "human errors of a lawyer drafting from earlier pleadings under time pressure." No sanction has been imposed.
The Copilot for Word prompt injection. Two hundred thirteen days since Håkon Måløy reported it to Microsoft, sixty-nine since he published the bypass.
Two default dates left from last issue. GitHub Copilot's auto-enablement starts October 22. OpenAI's cutoff for creating new custom GPTs is still listed as October 26 and still marked "planned," with retirement on December 11.
QUICK HITS
Relativity's server deadline is not moving. CEO Phil Saunders, at a media briefing at Relativity Fest in Chicago, per LawSites: "The policy is not to end the life of Server." And then: "The policy is to make sure that no net new matter is going to Server effective Jan. 1." That is January 1, 2028. As of the end of June, nearly 84 percent of annual recurring revenue was already in the cloud product, now called Relativity aiR. For partners who are not on board, he had one sentence: "Know we're going to run through you." If your e-discovery provider runs Server, ask where a matter opened in 2028 would live.
Anthropic's enterprise discount ends when the tokens do. The Information's reporting, as summarized by The Next Web: "When Anthropic customers hit their contracted token volume, they must renegotiate or pay standard rates." OpenAI "gives customers the rest of the current month plus one more month to agree a new contract." Anthropic's discounts are put at about 15 percent off list. That is your vendor's cost before it is yours.
More people are using AI at work and fewer are being taught how. PwC surveyed 49,364 workers across 48 countries and regions. Sixty-four percent used AI at work in the past year, up ten points, and 22 percent use generative AI daily, up from 14. The share who say they have the learning resources they need fell from 59 percent to 51. The group PwC calls front-runners, with scarce skills and strong AI capabilities, is 14 percent of the workforce, and 29 percent of them are very or extremely likely to change employers within a year.
Disney is hiring a legal engineer for a department it plans to shrink. The posting is for a Director, AI Enablement & Legal Engineering, at $185,500 to $248,800 in California. Artificial Lawyer, citing Deadline, reports a memo attributed to Chief Legal and Global Affairs Officer Horacio Gutierrez saying the 1,000-strong group "will be a much smaller organization than it is today," through automation, self-service, alternative providers and outsourcing.
Pierson Ferdinand reached 300 partners in under three years without hiring an associate. Falcon has one named customer and a price list.
See you in the next one.