Google launched a legal product on Tuesday, and the interesting part is not the product. It is the list of companies that agreed to plug into it. Two of them are legal AI platforms your firm may already have in its evaluation pipeline.
THE LEAD PLAY
Your Legal AI Vendor Is Now a Connector
Google Cloud put Gemini Enterprise for Legal into preview on Tuesday. The launch customers are Cleary, Freshfields, Weil, and Williams and Connolly. There are pre-built agents from Deloitte and Eudia, an integrator bench that includes Accenture, KPMG, and Factor Law, and a set of legal skills covering contract review, playbook creation, regulatory horizon scanning, legal research, and DSAR fulfillment.
None of that is the story. The connector list is the story.
Gemini Enterprise for Legal reaches your systems through MCP connectors, and Google published which ones. Google Workspace and Microsoft 365, including Word, Outlook, and SharePoint. iManage and NetDocuments. RelativityOne and Everlaw. Docusign. Thomson Reuters HighQ. CourtListener and Courtroom5. Then, at the bottom of the list, filed under specialized legal AI: Harvey, Legora, and Solve Intelligence.
Read that again. The two legal AI platforms most likely to be sitting in your procurement queue this fall appear on Google's launch announcement as connectors.
Legora put out its own release the same day, and it is worth reading closely, because it describes the arrangement in Legora's own words. Through a beta MCP connection, approved mutual customers “will be able to ask Legora legal questions from within Gemini Enterprise and receive text-based answers with links back to Legora for source review, citations, thread history, and continuity.” From there, “users can continue their work in Legora, where citations, source grounding, permissions, and legal context remain available for review.”
Strip the press release grammar off that sentence. Inside Gemini, Legora returns some text and a link. The actual product is one click away, somewhere else.
To be fair to Legora, this is a rational trade and they are not hiding it. Distribution inside a platform that a general counsel's whole company already runs is worth more than most marketing budgets. As someone who has spent years in legal case management systems, the trade is familiar. Building on someone else's platform buys reach you could not otherwise afford. What it costs you is the ability to answer one question: what happens at renewal when the platform ships your feature natively? Every company in that position will tell you it is the depth layer and the platform is just the front door. Sometimes that holds for years. Sometimes the front door grows a room.
For a 150-lawyer firm, none of this means you should do anything about Gemini specifically this week. It is in preview. Google has not published a price. The four launch firms have technology budgets that do not resemble yours, and their names are on the release partly because their names are on the release.
What it changes is the contract you are about to sign with somebody else.
Legal AI is still being procured the way software was procured in 2015. Seats, term, uptime, a security review, a data processing addendum with real teeth if your risk committee is any good. Three-year term for the discount. What almost nobody is negotiating is the case where the capability you just bought turns into a checkbox inside a platform the firm already licenses.
That case stopped being hypothetical this month. iManage shipped an MCP server in May and announced a strategic partnership with Thomson Reuters on August 20. DeepJudge published an open handoff protocol on August 13 with Harvey and Thomson Reuters signed on as first adopters. Bloomberg Law demonstrated an MCP integration with Anthropic's Claude at ILTACON. Google published a legal connector map with the specialists on it. The plumbing that lets a general platform absorb a specialist product is being installed in public, and the specialists are helping install it.
The Play this week: Pull every legal AI contract with a renewal or a signature date in the next two quarters. Before you sign any of them, put three terms on the table.
First, portability, in writing: your prompts, playbooks, and configured workflows come out in a usable format if you leave. Second, a repricing trigger: if the core capability ships natively in a platform the firm already licenses, you get a renegotiation right, not just a renewal notice. Third, a shorter term. Twelve months at a worse rate beats thirty-six months of being wrong.
If a vendor will not discuss any of the three, you have learned something useful about how that vendor rates its own position.
SECOND CHAIR
The bill got softer on citations and harder on disclosure
California's legislature adjourns at midnight on Monday. SB 574, Senator Tom Umberg's bill covering attorneys, arbitrators, judicial officers, and alternative resolution providers, was read a third time and amended on August 21 and is sitting on the Assembly floor. If it does not pass by Monday, it dies.
The headline provision is real, and it is blunt. New Business and Professions Code section 6068.1 opens with this: “An attorney shall not delegate the practice of law to generative artificial intelligence.” Arbitrators get their own version. “An arbitrator shall not delegate any part of their decisionmaking process to any generative artificial intelligence tool.”
What went almost entirely uncovered is what the August 21 amendment did, and it moved in two directions at once.
It got softer on citations. The August 13 version barred any filing containing citations the attorney had not “personally read and verified.” The August 21 text strikes “read and.” As amended, 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.” The August 13 version also carried duties to strip biased, offensive, or harmful output and to avoid disparate impact on protected classes. Those are gone from the current text entirely.
It got harder everywhere else. The surviving duties on an attorney “who uses generative artificial intelligence to assist in the practice of law” are to keep confidential, personal identifying, and other nonpublic information out of generative AI systems, to take reasonable steps to verify accuracy “including, but not limited to, the accuracy of all case and statutory citations” and to correct erroneous or hallucinated output, and then this: “Disclose the use of generative artificial intelligence to the court for all documents submitted to the court.”
The read: That last line is the one to walk down the hall to your litigation partners, and it is getting almost no attention. Not disclose on request. Not disclose when a standing order asks. Disclose, for all documents submitted to the court. If it passes in this form, every California filing your firm touches carries an AI-use determination, which means somebody has to know, per document, whether a tool was used and where. That is a docketing and workflow problem wearing the costume of a policy problem, and a firm that has only written a policy has not solved it. One more thing for anyone tracking the amendment trail: this bill has changed materially twice in two weeks, and at least one widely read outlet described the current version using language that was struck on August 21. Pull the text before you brief anyone on it.
Nashville spent the week talking about taxonomy
ILTACON filled Nashville this week, and the sessions were mostly not about tools.
Heath Harris of NetDocuments argued that well-structured data makes AI cheaper to run, because clean structure cuts how much context the model needs, and then warned against overbuilding it. “More is not better.” Ivy Grey of iManage pointed firms at the people who already do this work, meaning their librarians: “They understand how information is supposed to work together.” Harriet Joubert-Vaklyes of Michael Best put the change management point plainly: “You need to start with questions long before you decide that a new tool is necessary.”
Running alongside the conference, Neil Smith of LawVu published numbers that explain the mood. Forty-three percent of legal teams use AI contracting tools. Fewer than thirty percent are using any form of dedicated legal technology platform. Fifty-three percent can reach their data but cannot compile it across systems. His line is the one worth keeping: AI deployed against fragmented systems “does not just underperform. It produces fragmented answers dressed up as confident ones.”
The read: Put this next to the Lead Play and you have the week's real argument. Google is selling a layer that reaches into your DMS, your e-discovery platform, and your mailbox. That layer inherits whatever organization it finds there. A connector pointed at a badly structured document store returns confident answers from a badly structured document store. The firms that get value out of any of this next year are the ones spending this year on the boring part.
STILL WATCHING
Thomson Reuters v. ROSS Intelligence, No. 25-2153, argued in the Third Circuit on June 11. Seventy-six days, no opinion. The fair use question underneath every legal AI training claim is still sitting with three judges.
The handoff layer. DeepJudge released an open Agent Handoff Protocol on August 13, with Harvey signed on and Thomson Reuters saying it will support the protocol in CoCounsel Legal, implementation and timing still to come. iManage shipped an MCP server in May. Google published a connector map on Tuesday. What still has not happened is any firm describing a working handoff running in production between two vendors who did not build it together. Announcements are a long way ahead of deployments.
The California rules amendments to RPC 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 have not moved. Public comment closed May 4. One hundred fourteen days without a Board of Trustees vote.
The Copilot for Word prompt injection. One hundred seventy-four days since Hakon Maloy reported it, thirty since he published. Microsoft's next Patch Tuesday is September 8.
The prompt privilege split. Assini v. Hayward in New York and Tate Group Automotive in the Texas Business Court both treat AI prompts as work product. Heppner in the Southern District of New York does not. Nothing moved this week, and the split is still the most consequential unresolved question for anyone whose associates are typing case strategy into a text box.
QUICK HITS
Newcode raised $13.5 million in a Series A, with Relativity's investment arm Rel Labs participating, bringing it to $20 million raised this year. It sells firms a configurable platform for building their own AI rather than buying a finished one. CEO Maged Helmy: “The future of legal AI will not be defined by point solutions.” Read that next to the Lead Play and decide for yourself whether it is a thesis or a hedge.
Bloomberg Law showed a rebuilt AI experience at ILTACON. BLAW AI is generally available. Workspaces, AI Agents, and Watchlists are all in preview. There is an MCP integration with Anthropic's Claude. No price published.
iManage and Thomson Reuters announced a strategic partnership on August 20 to deliver governed AI-powered legal workflows across their products. Two of the systems on Google's connector list, wiring themselves to each other in the same week they both turned up on somebody else's map.
The reproducibility problem got a good framing, in a sponsored post. Writing a paid thought leadership piece in Artificial Lawyer, Neota Logic's Shaz Aziz posed the question almost no legal AI deployment can answer. The model has been updated twice, the person who wrote the prompt has left, and someone asks whether the same facts would produce the same decision today. It is a vendor arguing for the category of product that vendor sells. It is also the right question to put to your own stack.
Somewhere in your firm is a document store nobody has cleaned up since 2019, and a connector list published this week that includes it.
See you in the next one.