Harvey and LexisNexis shipped a workflow last Wednesday that turns a complaint into a cited motion to dismiss. Harvey says the job "commonly takes 40 to 50+ hours of work" and that its version takes a matter of hours. Before the software researches anything, it stops and waits for a lawyer to approve a ranked list of arguments.
THE LEAD PLAY
Harvey Drafts the Motion. A Lawyer Picks the Arguments.
The workflow went live October 7. It is the first of two that Harvey built with LexisNexis, and Harvey's post says "a Motion for Summary Judgment workflow is next." It is available to Harvey customers who have Ask LexisNexis or who also subscribe to Lexis+ with Protégé.
It runs in four stages. A planning agent reads the complaint, identifies the forum and the governing law, runs a first search of LexisNexis primary law, and comes back with "up to eight candidate grounds for dismissal." It checks each one against the elements of the pleaded claims, rates its strength, and proposes a ranked strategy. Then it waits. Harvey's sentence for the second stage is seven words long: "The lawyer reviews and approves the plan." The lawyer can take the plan as written or add and drop arguments.
After that the machine does the part that used to fill the timesheet. Research agents work each approved ground in parallel and write a memo on it, with the standard, the three to five strongest cases, the plaintiff's likely counterarguments, and the adverse authority with a way to distinguish it. Every case is run through Shepard's. A drafting agent assembles the motion as a Word document with a table of contents and a table of authorities. In state court it picks the vehicle on its own: a demurrer in California, a CPLR 3211 motion in New York, preliminary objections in Pennsylvania.
Some discounting is in order. The 40 to 50 hours is Harvey's number, and the post offers nothing measured behind it. There is no price on the page and no error rate. A Shepard's signal says a case is still good law. It does not say the case stands for the proposition in the memo, which was the problem Issue #18 was about. And plenty of firms reading this do not have Harvey.
They do have associates, and the stage worth studying needs no license. For as long as anyone can remember, a junior lawyer learned to rank grounds for dismissal by spending a week inside the research, getting the ranking wrong, and having a partner fix it in the draft. The workflow moves the ranking to the first hour and hands it to whoever is sitting at the checkpoint. The forty hours that taught the skill are the forty hours that went away.
The Play this week: Have a junior associate pick the arguments first. On your next motion to dismiss, give them the complaint and an hour to list the grounds they would raise, strongest first, one line each, before a partner weighs in or anyone opens a tool. Then set their list beside the partner's. Where the two disagree is what to teach. Plaintiff-side firms can run the same hour on their own complaint before it is filed, from the defense side of the table. In-house teams can ask outside counsel who ranked the grounds on the next motion.
SECOND CHAIR
AT&T Calls Its Legal Department an Agency
On October 8 AT&T said it is working with OpenAI as "a flagship in-house legal design partner." The work sits in AT&T LegalEdge, which the company calls its "in-house legal agency," a team of attorneys, paralegals, technologists and legal operations staff. The release says "LegalEdge has handled nearly 100 matters in its first six months." The aim is to handle routine legal work and make the department's own knowledge easier to reach. General counsel David McAtee: "AI gives legal departments an opportunity to do more than improve individual productivity." Chief compliance officer Bill Ryan: "Success will not be measured by how much technology we deploy." The same day, Legal IT Insider reported that BMW Group is rolling out Legora across its legal department for research, review and first drafts.
The read: Nearly 100 matters in six months is a pilot. The label is the news: an agency takes work in. AT&T's release does not say which matter types LegalEdge keeps.
Sorting Ten Thousand Documents Now Costs Under Two Dollars
OpenAI opened a public beta last week of something it calls the Decisions API. It does one narrow job. It reads text or images and returns the probability that a condition is true, a choice from a fixed list, or a score against a rubric. The price is $0.10 per million input tokens, and OpenAI's documentation says only input tokens are billed. At an assumed 2,000 tokens a document, a first pass over 10,000 documents comes to about two dollars before the instructions are counted. Three days later Microsoft listed its own version, Decision-1, at $0.042 per million input tokens with free output. The same pass comes to about 84 cents.
Ivo, the contract review company, went a step further on October 1 and gave a model away. Sage is a version of DeepSeek V4 Flash trained for contract work, posted under an MIT license. Ivo's release says it meets 91 percent of the pass criteria on a contracts benchmark, up from 70. The model card carries a second number. On the 55 contract tasks, Sage met every criterion on 16.4 percent of them, at $1.24 a task. GPT-6 Astra scored the same 16.4 at $74.29. The best model on the card, Fable 5.1, reached 29.1.
The read: Sampling a document set was a budget decision, and the budget reason is mostly gone. The cost that remains is checking the classifier against documents a person has already coded.
Vorys Has 36 Partners in Software. Clients Are Next.
Law.com reported on October 6 that Vorys has built AI versions of 36 senior partners with Stanford Law School's liftlab. In the firm's own summary of the story, lawyers and staff use the personas to review documents, suggest edits to drafts and flag contract risks. In May, Reuters put the count at 19. Managing partner Michael Martz said the firm plans to build personas of clients' general counsel and in-house lawyers and add them to its drafting assistant. His line: "We see AI as a way to not only improve efficiency, but really improve the work product."
The read: Two questions come with this, and the coverage answers neither. What happens to a partner's persona when the partner leaves? Does a general counsel get a say before a firm builds a model of her?
STILL WATCHING
The Copilot for Word prompt injection. Two hundred twenty days since Håkon Måløy reported it to Microsoft, seventy-six since he published the bypass.
The California rules amendments. Sixty-seven days since the second comment period closed. The State Bar's page still shows the package at the comment stage.
SB 574. Eighty-one days until the disclosure requirement takes effect in California.
Two default dates. 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."
QUICK HITS
Thomson Reuters has been sued over the Casetext deal. A Carlsbad, California firm filed a proposed class action in the Northern District of California on October 1, calling the $650 million purchase a "killer acquisition." Per the ABA Journal, the complaint says the firm paid $32.50 a month for Casetext and $111.48 for Westlaw Classic, on a three-year term with 4 percent annual increases. It asks for treble damages and for Casetext to be re-established as an independent company. Thomson Reuters says it has not been served.
Ivo also tested how models negotiate. With the data lab micro1 it built a contract benchmark, and its summary is that leading models "concede easily and push back poorly." When the right move was to counter or reject a provision, they got it right 46 percent of the time. They met 23 percent of the criteria for when to escalate. It is a vendor's benchmark, and Ivo says the full results come later.
Legora's agents now take instructions from a file you can keep. Legora released more than 250 prebuilt "Skills," which it describes, per Artificial Lawyer, as natural-language instructions saved as a Markdown file. Users can read, edit and export them. The one-page playbook from last issue now has a file format.
Cooley is helping Google build a redaction agent. It reviews court filings for confidential material and proposes redactions, and lawyers make the final call. Ian Shapiro, who chairs Cooley's litigation department, in Legal IT Insider: "Cooley has long collaborated with Google as litigation counsel." The client is now also the vendor.
Harvey's post says a lawyer reviews and approves the plan. It does not say which lawyer.
See you in the next one.