Crosby raises $60M to scale the AI law firm model. A San Diego attorney receives the largest AI hallucination sanction in US history. Nineteen new state AI laws are signed. Harvey publishes its approach to training legal agents. And LexisNexis launches a practice area dedicated to AI risk.
When Sequoia's Jess Lee wrote in March that "a copilot sells the tool, an autopilot sells the work," she was describing a thesis. Crosby is now the best-funded company trying to prove it.
The numbers are striking. Less than a year out of stealth, Crosby has negotiated contracts worth over $1 billion for clients and claims contracts are completed up to 80% faster than traditional processes. The round was led by Lux Capital and Index Ventures, with Sequoia, Bain Capital Ventures, Elad Gil, and Patrick Collison (Stripe's CEO) all participating. Total funding now exceeds $85 million.
What makes Crosby worth watching isn't the funding. It's the model. Crosby is a law firm. It employs licensed attorneys. But the first pass on every contract is handled by proprietary AI agents. Attorneys review the output, not the intake. The client pays per contract, not per hour. The median turnaround is 58 minutes, with a guaranteed four-hour SLA.
The roadmap disclosed alongside the raise is more revealing than the headline numbers. Crosby is building counterparty simulation (predicting how the other side will respond to specific redlines), voice agents that can negotiate on behalf of clients, and collaborative platforms that give clients real-time visibility into legal work in progress. These are not incremental improvements to a copilot. They are the infrastructure of a service business that happens to use AI as its production layer.
There's a convergence happening in legal AI that I think is worth naming explicitly. Service companies are building platforms. Platform companies are building services. Crosby started as a law firm and just launched a Client Console. Harvey started as a platform and just launched Agent Builder, which lets firms create their own autonomous workflows. The end state looks the same from both directions: AI does the work, humans supervise, the client pays for outcomes.
Crosby's client list is overwhelmingly high-growth startups without GCs. The enterprise buyer looks different: bigger volumes, stricter governance, more complex playbooks, multiple jurisdictions. The question is whether the AI law firm model can move upmarket without losing the speed and simplicity that make it compelling in the first place. The firms that figure out enterprise-grade supervision at startup-speed turnaround times will own the next phase of this market.
The Brigandi decision isn't just the largest monetary sanction for AI-generated legal fiction. It marks a shift from embarrassment to enforcement in how courts handle AI misuse.
Stephen Brigandi, a San Diego attorney serving as pro bono counsel in a Valley View Winery dispute in Oregon, submitted three filings containing 23 fabricated legal citations and 8 false quotations generated entirely by artificial intelligence. Judge Clarke's characterisation was blunt: "a notorious outlier in both degree and volume" in the growing universe of AI sanction cases. The $96,000 in direct sanctions, combined with additional penalties against co-counsel, pushed the total past $110,000. The client's case was dismissed with prejudice.
The acceleration is staggering. Damien Charlotin's global database of AI hallucination cases has catalogued 1,227 incidents, up from 660 in December 2025 and roughly 120 total between April 2023 and May 2025. That's approximately five to six new documented cases per day. In Q1 2026, US courts imposed at least $145,000 in sanctions related to AI-generated filings.
On a single day, 31 March, seventeen US court decisions referenced suspected AI hallucinations in filings. That's not a trend. It's an institutional problem.
The courts are solving this problem the way courts solve problems: with rules, disclosure requirements, and penalties. But rules address the symptom. The root cause is a workflow that has no supervision layer between AI generation and human filing. Every one of these cases would have been prevented by a system where AI output is reviewed against verified sources before it reaches a downstream process. That's not a technology gap. It's an architecture gap.
Nineteen new state AI laws in a matter of weeks. 1,561 bills introduced across 45 states. The federal preemption debate is becoming academic as states build their own regulatory infrastructure.
The pace of state AI legislation in 2026 has no precedent in technology regulation. Between mid-March and early April, 19 new AI laws were signed, bringing the year's total to 25. Governors in Oregon, Idaho, and Tennessee each signed AI-focused legislation during the period. The bills span chatbot safety, healthcare AI disclosure, algorithmic accountability, and deepfake protections.
The White House published its preemption framework on 20 March, urging Congress to prevent states from regulating AI model development. In the three weeks since, states have passed more AI laws than the federal government has enacted in the past two years combined. Congressional preemption has already been rejected twice this session.
For enterprise legal teams evaluating AI tools, the compliance surface area is expanding rapidly. Any AI system used in legal services will need to demonstrate conformity with the EU AI Act's high-risk requirements by August. Colorado's impact assessment obligations follow weeks later. Governance isn't a feature request any more. It's a procurement filter. The vendors whose AI supervision and audit trail capabilities are native to the product have a structural advantage over those bolting compliance on after the fact.
Two of the largest companies in legal AI made moves this week that reveal different theories of where the value accrues in an AI-transformed legal market.
Harvey published two pieces of research that, taken together, suggest a significant strategic shift. The first, from Head of Applied Research Niko Grupen, details "harness engineering," a methodology for improving legal agent performance through structured feedback loops rather than larger models or more data. The second introduces the Spectre agent and a concept Harvey calls a "law firm world model," which models not just documents but the operational patterns of an entire firm.
This matters because it signals Harvey's ambition beyond document-level AI. A "law firm world model" implies agents that understand billing structures, staffing patterns, client relationship dynamics, and workflow interdependencies. Harvey processes 400,000+ queries daily across 1,300 organisations, giving it a data advantage that is difficult to replicate. The recent $200M raise at $11 billion and the Agent Builder launch suggest this is where the investment is going.
LexisNexis launched Practical Guidance AI & Technology on 9 April, a task-based practice area for attorneys advising on AI deployment, procurement, and governance. The timing is deliberate: with the EU AI Act high-risk provisions four months away and state-level obligations multiplying, every corporate legal team needs guidance on AI compliance.
The new practice area covers AI-specific contract drafting (SaaS, cloud, licensing, outsourcing), regulatory compliance assessment, and dispute management involving AI systems. The content integrates with Lexis+ with Protégé, LexisNexis's workflow AI assistant launched in February.
This week crystallised something that's been building for months. The legal AI market is separating into two distinct economies: one that sells tools to lawyers, and one that sells the work lawyers used to do. The Crosby raise and the sanctions crisis are two sides of the same structural shift.
The sanctions crisis and the Crosby raise are mirror images of the same underlying problem. Unsupervised AI output is entering the legal system because the current workflow has no structural layer between generation and delivery. Crosby's answer is to be the supervision layer: AI generates, their attorneys review, the client gets the finished work. The courts' answer is to mandate disclosure and penalise negligence. Both responses point to the same conclusion: the value in legal AI is not in generating output. It's in supervising it.
This maps directly to the structural question in enterprise legal. In-house teams spend the majority of their capacity on routine, rules-based work: NDAs, DPAs, vendor agreements, procurement contracts. Inexpensive work, done by expensive resources. The copilot model makes those expensive resources slightly faster. The agentic model takes the work off their desk entirely, under their supervision, against their playbooks. That's not a technology budget conversation. It's a services budget conversation. For every pound spent on legal software, ten to fifty are spent on legal services. The companies that compete for the services budget, not the technology budget, are the ones rewriting the economics of enterprise legal.