← Insights home
The Intake — Weekly briefing

This week in legal AI

Ironclad's 2026 State of AI in Legal confirms the tipping point: 91.6% of lawyers now use AI, 42% have cut outside counsel spend — but 60% of in-house teams see none of those savings in their bills. California closes public comment on the first binding AI professional conduct rules. UC Berkeley Law bans AI for all student work, citing hallucinated citations in submissions. Lavern, a 67-agent open-source agentic law firm, goes viral on GitHub. And Docusign goes agentic — calling its partnerships with Harvey, Legora, and Thomson Reuters "co-opetition."

Week of 23 May – 29 May 2026
Category Market intelligence
Reading time 8 minutes
01 — The week at a glance

Five stories that matter

May 23 — Ethics
The State Bar of California's 45-day public comment period closed this week on proposed amendments to six professional conduct rules — covering competence, client communication, confidentiality, candour toward tribunals, and supervision of lawyers and staff. Unlike California's 2023 guidance, which carried no enforcement weight, these amendments write specific AI obligations into binding ethics rules for the first time. Every state bar in the country is watching.
May 26 — Architecture
Built by Helsinki-based legal tech founder Antti Innanen over six months, Lavern is now open-sourced under Apache 2.0 and gaining rapid traction. It coordinates 67 specialist AI agents through an evidence-backed debate protocol, with a 10-pass verification loop and mandatory human gates before any critical decision executes. It runs against Anthropic, Mistral, or locally via Ollama. It is not a real law firm. It is the clearest public demonstration yet of what supervised multi-agent legal architecture actually looks like.
May 27 — Research
Ironclad's third-annual State of AI in Legal surveyed more than 800 legal professionals. The headline number — 91.6% adoption, up from 69% in 2025 — confirms the tipping point has passed. The harder number: nearly 60% of in-house counsel report no noticeable cost savings despite their outside firms running AI at scale. The efficiency is real. The economics are not passing through to clients.
May 27 — Education
Berkeley Law announced a blanket prohibition on AI use in any work submitted for credit, in any exam, and in uploading course materials to AI systems. The trigger was specific: faculty observed a spike in "flawed legal analyses and hallucinated citations" in student submissions. The policy defaults to a ban; individual instructors may permit AI use in writing. It is the most aggressive response yet from a top law school to the training pipeline crisis.
May 28 — Enterprise
Docusign launched agentic contract workflows for its IAM platform and announced simultaneous partnerships with Harvey, Legora, and CoCounsel Legal by Thomson Reuters. Harvey and Legora handle legal reasoning and drafting; Docusign handles execution and post-signature obligation tracking across the enterprise. Law360 asked Docusign to describe the arrangement — the answer was "co-opetition." The largest e-signature platform on earth has decided that embedding its competitors is better than competing with them.
02 — The efficiency paradox

Lawyers are more AI-powered than ever. Clients aren't saving a penny.

Ironclad's 2026 survey is the clearest data yet on a structural disconnect the legal market has been circling for two years. AI is making lawyers faster. The savings are not reaching the people who pay for legal work.

The headline numbers are striking. AI adoption among lawyers jumped from 69% in 2025 to 91.6% in 2026 — a 33-point increase in twelve months. Ninety-nine percent now trust AI tools. Eighty-nine percent report spending more time on complex, strategic work since adopting AI. Forty-two percent of respondents say they have cut outside counsel spend.

Then there is the number that should matter more to enterprise legal teams: nearly 60% of in-house legal leaders say they have seen no noticeable reduction in what they pay their outside firms, even as those firms are publicly reporting AI-driven productivity gains. The efficiency is documented. The economics are not passing through.

What law firms report
The AI-powered lawyer
  • 91.6% of lawyers now use AI — up from 69% in 2025
  • 97% report at least one measurable business outcome from AI use
  • 89% spend more time on complex, strategic matters since adopting AI
  • 52% report faster response times to business stakeholders
  • 50% report faster contract turnaround times
  • 96% say their organisation expects more from the legal function than two years ago
What clients experience
The unchanged invoice
  • ~60% of in-house counsel: no noticeable cost reduction from outside firm AI use
  • Law firms face no structural obligation to pass AI efficiency savings to clients
  • Hourly and value-based billing both let the firm absorb productivity gains as margin
  • Enterprise clients pay the same per-matter rates while firms invest in AI infrastructure
  • 90% of GCs report resource constraints prevent delivery of strategic impact
  • Only 42% of in-house teams — those who insourced or renegotiated — have captured savings

Why the gap persists — and why it isn't closing

The disconnect is structural, not incidental. Law firms do not bill for time saved — they bill for time spent, or for outcomes under value-based arrangements. When an AI tool reduces a four-hour contract review to forty minutes, the firm can pass the saving to the client, absorb it as higher margin, or re-allocate the lawyer to the next matter. The first option requires a renegotiated engagement model that most firms are not proactively offering. The second and third are the rational choices under the existing billing structure.

The 42% of respondents who report cutting outside counsel spend are almost certainly the in-house teams that renegotiated fees, insourced more work, or moved volume to alternative providers. They are not the majority. The majority is still paying the same rates — for faster lawyers.

The Flank read

The Ironclad data confirms the structural problem that Flank is built to solve. Inexpensive legal work is still being done by expensive people — just faster. The productivity gains from AI tools accumulate on the law firm balance sheet, not in the enterprise legal budget. The architectural shift that changes this equation is not a faster lawyer. It is an agent that does the inexpensive work without the billing structure around it. When the NDA goes to the agent instead of to outside counsel, 100% of the efficiency is in the enterprise budget — not renegotiated, not passed through: simply never charged.

91.6%
Lawyer AI adoption in 2026 — up from 69% in 2025, a 33-point jump in twelve months
~60%
In-house teams seeing no cost reduction from outside firms' AI use — despite widespread adoption
42%
In-house counsel who have cut outside counsel spend — the minority who found an alternative
97%
Lawyers reporting measurable business outcomes from AI — the productivity is real; the billing pass-through is not
03 — Professional conduct

California's AI ethics rules move from guidance to enforceable law. Six rules. Every bar association watching.

In 2023, the California State Bar published practical guidance on AI for lawyers. It was a living document with no enforcement authority. The proposed amendments now through public comment are categorically different: six changes to binding professional conduct rules, with the full disciplinary infrastructure of the existing rules behind them.

The California bar has proposed amendments to six rules of professional conduct. Each amendment writes AI-specific obligations into the existing rule rather than creating a standalone AI policy — meaning they carry the enforcement weight of the parent rule. Non-compliance is a matter of professional conduct, not just a guideline violation.

Rule 1.1 — Competence
AI fluency is now a competence requirement
Lawyers must independently review, verify, and exercise professional judgment on any AI output used in client representation. Understanding model limitations — including hallucination risk — becomes part of what it means to be competent.
Rule 1.4 — Communication
Material AI use must be disclosed to clients
Where AI use materially affects the representation — costs, scope, or the nature of work product — lawyers must inform clients. The "material" threshold gives firms discretion but creates a genuine, enforceable disclosure obligation for the first time.
Rule 1.6 — Confidentiality
Client data must be protected when sent to AI
Lawyers may not disclose confidential client information to AI systems without informed consent or a recognised exception. This directly governs use of public LLM endpoints with client data — and creates liability for choices about which AI vendor to use.
Rule 3.3 — Candour
No AI hallucinations in court filings — ever
The candour obligation now explicitly reaches AI-generated content. Submitting fabricated citations — even unknowingly, if the lawyer failed to verify — becomes an ethics violation, not just grounds for sanction by the court.
Rule 5.1 — Supervising lawyers
Partners are responsible for junior lawyers' AI use
Supervisory lawyers must ensure that subordinates comply with AI-related professional conduct obligations. The supervision liability from AI errors travels up the chain — to the partner who signed off on the work product.
Rule 5.3 — Non-lawyer staff
Staff AI use is a professional responsibility
The same supervision obligations extend to paralegals, legal operations staff, and contract reviewers. Any non-lawyer using AI tools in a matter must be appropriately supervised under the amended rules.

What California does, others follow

California's professional conduct rules have historically been bellwethers for national standards. At least 25 federal district courts have already adopted standing orders requiring AI disclosure in filings. Florida, Texas, and New York have issued ethics opinions that will likely need updating once California's binding rules take effect. The professional conduct framework for legal AI is moving from guidance to enforcement — and this time, at a national pace.

What it means for enterprise legal teams

For enterprise legal teams, the most consequential rules are 1.6 (confidentiality) and 5.3 (non-lawyer staff supervision). If your legal operations team, paralegals, or contract reviewers are using AI tools that handle counterparty or client data — which describes almost all current deployments — those tools are now subject to professional conduct scrutiny, not just data security policy. The question your GC and CLO will face when these rules are adopted: can you demonstrate that your AI vendor's governance architecture satisfies the supervision and confidentiality obligations? "We use enterprise data handling" is not the same as "we can show how every AI output was reviewed before it affected a client relationship."

04 — The education response

Berkeley Law bans AI. The question it raises is bigger than the policy.

Last week we covered how AI is eliminating the associate training ground inside law firms. This week, one of the country's top law schools drew the obvious institutional conclusion: if AI is removing the reps that build legal judgment, law schools need to protect the spaces where those reps still happen.

UC Berkeley School of Law announced a blanket AI ban for student work submitted for credit, effective summer 2026. The prohibition covers: drafting, outlining, revising, translating, or editing submitted work; all exam situations; and uploading course materials to AI systems. Students may still use AI to identify cases or statutes for research — but they must verify those sources manually.

The trigger was specific and measurable. Faculty observed a spike in flawed legal analyses and hallucinated citations in student submissions. This is not a philosophical position about AI in education. It is an institutional response to evidence that students using AI are producing worse legal work than students who are not.

The problem Berkeley is solving
Protecting the formation of judgment
The cognitive skills Berkeley is defending — legal analysis, case reading, argument structure, identifying what matters in a mass of text — are built through repetition and error. When AI handles the first draft, the student never develops the judgment to know when the draft is wrong. The ability to supervise AI output depends entirely on having judgment the AI lacks. You cannot acquire that judgment by watching AI perform the task.
The problem Berkeley is not solving
What graduates will walk into
Law firms are deploying AI at scale. Enterprise legal teams are building agentic workflows. The lawyers graduating in 2027 and 2028 will be expected to operate alongside AI from day one. The Berkeley ban protects the foundational skills required to supervise AI outputs — but defers the question of when and how lawyers learn to work with them in practice.

The tension may be irresolvable by policy alone. The skills that make someone a good lawyer — precision under pressure, analytical judgment, knowing which detail changes everything — are the same skills that make someone a good supervisor of AI outputs. You cannot train the latter without the former. Berkeley is defending the foundation. The question of how the superstructure gets built is still open.

The structural question for enterprise legal teams

The legal talent pipeline is under pressure from both ends. Inside firms, AI is taking the volume work that trained junior associates. Inside law schools, the best institutions are restricting AI to protect foundational reasoning skills. For enterprise legal teams still routing volume work to outside counsel — where it lands with associates in an increasingly hollowed pipeline — this creates a practical quality risk that isn't priced into the hourly rate. The teams insulating themselves are the ones that have moved routine work off lawyers' plates entirely, so the pipeline question stops being their problem.

05 — The open-source moment

Lavern: 67 agents, human gates, Apache 2.0. The architecture is no longer proprietary.

Lavern, released publicly this week on GitHub under an Apache 2.0 licence, is the most complete open-source implementation of supervised multi-agent legal work published to date. What it reveals is not just a tool — it is the architecture that the legal AI market has been circling for two years, now visible in 155,000 lines of code.

Built by Antti Innanen — a Finnish legal tech founder and practitioner — over six months, Lavern coordinates 67 specialist AI agents through an evidence-backed debate protocol. Each agent posts findings with cited evidence. Multiple agents with different analytical perspectives debate before a conclusion is reached. A 10-pass verification loop runs on all outputs. A mandatory human gate pauses the system before any critical decision executes. The system runs against Anthropic, Mistral, or fully local via Ollama.

What is instructive about Lavern is not what it can do in enterprise production today — it is what the four structural layers of its architecture tell us about what supervised legal AI needs to look like, regardless of whether it is open-source or commercial.

Layer 1 — Specialisation
67 agents, each optimised for a specific task
Rather than routing all legal work to a single generalist model, Lavern deploys specialist agents for each task type. The bet is that specialisation produces more reliable outputs in a given domain. The same architecture — specialist agents over generalist models — underlies most enterprise agentic legal services.
Layer 2 — Grounding
Evidence-backed debate before conclusions
Agents must cite evidence for their positions. Multiple agents debate before a conclusion is reached. The design reduces hallucination risk structurally — by requiring grounded reasoning, not just confident-sounding output — rather than relying on prompt engineering alone.
Layer 3 — Verification
10-pass verification loop on all outputs
Every output runs through a multi-pass verification process before it reaches the human gate. The loop catches internal inconsistencies, missing citations, and factual errors that a single-pass review misses. Quality is an architectural constraint, not a post-hoc check.
Layer 4 — Supervision
Mandatory human gates — not optional review steps
No critical decision executes without a human review point. The human gate is not a setting or a flag — it is a structural constraint. Supervision is baked into the workflow before anything reaches a counterparty. This is the architecture the California Bar rules now require.

Open-source architecture, proprietary execution

Lavern's Apache 2.0 release removes one argument that enterprise legal AI vendors could make: that the architecture for supervised multi-agent legal work is proprietary and hard to replicate. It is now public. What the licence cannot commoditise is everything that makes the architecture work at enterprise scale: the legal workflow expertise encoded in the agent instructions, the institutional knowledge about what acceptable output looks like, the integration layer with CLM, DMS, and email systems, the escalation logic that routes edge cases to the right human, and the trust framework that makes the supervision real rather than nominal.

Open-source also does not solve the governance obligation. Under the California Bar rules that just closed for public comment, lawyers bear professional conduct responsibility for AI outputs. That responsibility does not ship with the repository. It comes with the service wrapper — and the humans — around it.

The Flank read

Lavern is the open-source proof-of-concept for something the market already knows: the architecture for supervised multi-agent legal work is not the hard part. The hard part is the legal expertise encoded in the agent instructions, the institutional knowledge about what "good" looks like for a specific enterprise's contracts, the integrations that put agents into existing workflows, and the governance layer that makes supervision real. When Lavern's 67 agents output a contract review, someone still has to have the judgment to know whether the output is right. That judgment is not in the repository. It is in the legal team and the service layer that sits between the agents and the counterparty.

06 — So what

What this week tells us

Five stories from one week — an efficiency report, an ethics ruling, a law school ban, an open-source release, and an enterprise co-opetition deal — and they converge on the same structural question: who bears responsibility for AI outputs, and who captures the value of AI efficiency?

The Ironclad data puts the economic question in numbers for the first time: AI adoption has tipped above 90%, productivity gains are documented, and the savings are not reaching clients. The law firm is absorbing AI efficiency as margin. The gap between what AI makes possible for enterprise legal teams and what their outside law firms are delivering to their budgets is now quantified.

The California Bar story puts the governance question into enforceable code: lawyers will soon be professionally obligated — under binding conduct rules — to verify AI outputs, protect client data in AI interactions, and supervise subordinates' AI use. The tools that survive the compliance environment will be the ones with supervision built into their architecture as a structural constraint, not a checkbox in the settings panel.

Berkeley Law's AI ban and Lavern's open-source release describe the same architectural truth from opposite directions. Berkeley is protecting the human judgment that makes AI supervision meaningful — recognising that supervision requires judgment, not just access. Lavern is showing what a system that respects that judgment looks like in practice, with mandatory human gates and a 10-pass verification loop. The premise of both is identical: AI outputs require genuine human review, and the human doing the reviewing needs real judgment, not just the ability to click approve.

And Docusign's "co-opetition" strategy shows where the enterprise contract market is settling. The largest execution platform in the world has decided that embedding AI legal reasoning is more valuable than competing with it. The question for enterprise legal teams is which side of that integration they want to be on: the customer of the combined Harvey-Docusign stack, paying for both layers separately, or the team that moved the routine work to a supervised agentic service before either layer got to price it.

The Flank read

This week's stories describe a market in the middle of a structural repricing — but one where the repricing is not yet reaching the enterprise legal budget. The old model (inexpensive work done by expensive people) is generating data that confirms its inefficiency, regulatory pressure that raises its compliance cost, and institutional responses that acknowledge its pipeline problems. The new model — supervised agents doing the inexpensive work, with experienced people in the loop where judgment is required — is now visible in open-source form and validated by the enterprise contract management leaders. The architecture is not in question. The question is who deploys it for enterprise legal teams — and who captures the economics when they do. The 60% of in-house teams not seeing savings from their outside firms' AI use are the teams that have not yet answered that question for themselves.

Subscribe

The Intake

Weekly briefings on what's actually changing in legal AI — the market shifts, regulatory moves, and structural questions that matter for enterprise legal teams.

Subscribe on Substack
Flank

Outsource legal work to supervised agents

Enterprise legal teams use Flank to handle high-volume contracting end-to-end — NDAs, MSA redlines, procurement, triage. Agents that know your templates, terms, and escalation rules. Every output reviewed before it leaves the system.

Learn more at flank.ai