The Lab Notebook: No. 6
Welcome to The Lab Notebook, VAILL's field log for tracking how AI is reshaping legal work. Our goal: fewer hot takes, more meaningful signals. We're designing each Notebook entry to reflect an industry in motion: Version Control highlights recent meaningful updates, Open Query surfaces the unknowns we should be debating, Network Effects highlights what's compounding through shared practice, and System Status checks the guardrails and institutional infrastructure. We close with The Signal, one high-fidelity takeaway that suggests where the future is heading.
🔁 Version Control
Frontier models open-source the legal stack. On May 12, Anthropic released Claude for Legal under Apache 2.0 — 12 practice-area plugins and 20+ MCP connectors, all on GitHub. Days later, legal-tech founder Antti Innanen open-sourced Lavern, a 67-agent legal system with an evidence-backed debate protocol and a 10-pass verification loop, also Apache 2.0. Anthropic · Lavern (GitHub) · Artificial Lawyer
The dominant legal incumbent rebuilds on the same foundation. On the same day as the Claude for Legal release, Thomson Reuters announced that the next generation of CoCounsel Legal is rebuilt on Anthropic’s Claude Agent SDK and connected to Claude via MCP. Thomson Reuters
VAILL take: Last issue’s question about the vendor layer being bypassed has a new wrinkle. Capability is now open at one end of the stack and embedded in the dominant incumbent at the other — both running on the same Anthropic substrate, announced the same week.
🧐 Open Query
The supervision rule meets the agentic workflow. Baker McKenzie’s Danielle Benecke introduced the “supervision gap” framework at Legalweek 2026, noted this month in Sean Harrington’s structural analysis of barriers to AI adoption in law for Diffuse AI. Her argument: when agentic systems handle entire workflows, complete attorney review becomes economically irrational, and liability begins shifting from lawyer to vendor — what she calls the “outcome economy.” Diffuse AI
Three federal courts, three answers on AI privilege. In US v. Heppner, the S.D.N.Y. ruled that the use of AI by consumers waives attorney-client privilege. The E.D. Mich. reached the opposite result in Warner v. Gilbarco, treating AI as “tools, not persons.” And the D. Colo. in Morgan v. V2X granted Rule 26(b)(3) work-product protection to a pro se litigant’s AI use, prompting a comparative synthesis from Kirkland & Ellis this month. Kirkland & Ellis
VAILL take: Benecke names the structural problem. The privilege split is the same problem surfacing one case at a time. Neither will resolve in the comment period of a single state bar’s rules.
🌐 Network Effects
A new working paper makes a structural claim about the docket. In "Access to Justice in the Age of AI," MIT's Anand V. Shah and USC's Joshua Y. Levy analyze 4.5 million federal civil cases and 46 million PACER docket entries from FY2005–FY2026, documenting that the non-prisoner pro se filing share rose from a long-term steady state of approximately 11 percent to 16.8 percent in FY2025 — with the share of complaints flagging positive for AI-generated text rising from essentially zero in the pre-AI period to over 18 percent by early 2026. They report that pro se cases are not resolving any faster, but intra-case docket activity in the first 180 days is up 158 percent from pre-AI means — concentrated in case types characterized by formulaic document production. SSRN
The innovation-leader job rewrites itself. Bloomberg Law columnist Roy Strom reports that Big Law innovation leaders describe their roles transforming from solving discrete problems to driving firmwide AI adoption while firm leadership demands roadmaps they don’t yet have. Fried Frank’s Conan Hines, on LinkedIn: “I am amazed at what’s happening as much as the next soul. I’m also exhausted.” Bloomberg Law
VAILL take: Entry No. 5 covered how courts have changed who answers for AI mistakes. These items track what AI is doing before any mistake reaches a courtroom — to the docket itself, and to the jobs of the people overseeing AI adoption.
🚦 System Status
California’s two-track build-out continues. Entry No. 5 covered the COPRAC-proposed Rule amendments, with the comment period closing May 4. On May 14, the State Bar’s Board of Trustees approved an updated Practical Guidance for Generative AI, explicitly addressing agentic AI for the first time — including the requirement that “a lawyer must not deploy an agentic AI system in a manner that permits autonomous external transmission of client information.” State Bar of California
Fixed-fee surfaces inside Big Law’s flagship AI offering. Linklaters launched Applied Intelligence in early May — a practice combining lawyers and data scientists to build bespoke AI workflows for matters where off-the-shelf tools fall short — explicitly on a fixed-fee basis. Co-founded by partner Tom Quoroll and AI delivery director Sarah Barnard, the practice is explicitly fixed-fee. Quoroll, to The American Lawyer: “It would be hypocritical for me to do anything else but to say this is a fixed fee model.” The American Lawyer
VAILL take: Two parts of the institutional layer are being rewritten at once. California is writing agentic AI into the rules that govern legal work. Linklaters is pricing its newest AI practice by the matter rather than the hour.
📡 The Signal
The supervision rule is the only thing that hasn’t moved.
Claude for Legal and Lavern open-sourced the legal stack. Thomson Reuters rebuilt its flagship on the same Anthropic SDK. Linklaters priced its newest AI practice by the matter, not the hour. California named “agentic AI” in its guidance. A working paper put a number on what AI is doing to the federal civil docket.
The lawyer still signs the brief. Danielle Benecke’s name for the distance between those two facts — the supervision gap — got wider this month.
The Lab Notebook is a collaboration between VAILL researchers and Claude (Anthropic). Humans and AI work together to curate sources, set direction, analyze and compile entries, and edit.