Michael and Peyton Marcotte built their startup around a single conviction: law firms will pay to keep AI inside their own walls. The brothers, accepted into Y Combinator's program, created what they call a "grounding engine"—software that blocks any AI output unless it can trace the claim back to a document the firm already owns. No shared training pools, no data crossing the firewall, no invented case citations slipping into a brief at three in the morning.
The pitch arrives at a peculiar moment. Personal adoption of AI among lawyers has accelerated faster than anyone predicted, yet institutional guardrails remain scarce. Law firms find themselves caught between curiosity and caution, experimenting with tools they don't entirely trust.
The Trust Gap
Bloomberg Law's State of Practice survey, published in late June 2026, reported that 83 percent of lawyers had used AI in some capacity. The 8am Legal Industry Report tracked an even sharper climb: 69 percent of respondents said they personally deployed general-purpose AI for work, up from 31 percent the prior year. Firm-level governance, however, lagged badly. The DC Bar noted in coverage that summer that official policies trailed individual experimentation by a ratio of two to one.
Meanwhile, the big platforms were racing to dominate. Thomson Reuters announced that CoCounsel had reached one million users by late February 2026. LexisNexis launched Lexis+ with Protégé the same day. Harvey, the startup that serves a majority of the AmLaw 100, raised $200 million in March at an $11 billion valuation—up from $8 billion in December 2025 and $3 billion in February 2025, according to TechCrunch. Forbes reported that clients include A&O Shearman, Latham & Watkins, and O'Melveny.
Still, doubt persists. Damien Charlotin's AI Hallucination Cases database had cataloged more than 1,700 incidents by September 8, Axios noted. The Fifth Circuit cited that tracker in a published opinion, warning that fabricated citations posed a growing threat to the integrity of legal work. Who controls the data? What happens when a model invents a case number that looks plausible but doesn't exist?
Three Pressures
Privilege remains the first worry. In United States v. Heppner, decided in February 2026 by the Southern District of New York, a court found that using consumer AI tools waived attorney-client privilege. The Washington Legal Foundation analyzed the ruling in late April, cautioning that outcomes might differ when firms operate under enterprise contracts—but the risk is real.
Ethics codes are tightening. The ABA's Formal Opinion 512, issued in July 2024, applied Model Rules to generative AI: competence requires understanding a tool's limits, confidentiality demands scrutiny of data handling, and supervision applies when associates or paralegals use these systems. California's Rule of Court 10.430, effective September 2025, requires courts to adopt AI policies and may compel disclosure of entirely AI-generated filings. Trackers maintained by Ropes & Gray and Orthodoxy Legal show disclosure orders proliferating across benches.
Then there's utilization. Clio's Legal Trends Report, covering a decade of lawyer productivity through 2026, found that attorneys had improved from billing 28 percent of an eight-hour day to roughly 37 percent. Five hours still went unbilled or unrealized, the firm wrote in a June blog post. Bloomberg Law's 2026 Attorney Workload & Hours Survey put the gap differently: lawyers work around 49 hours weekly but bill only 37. Firms see AI as a way to recapture time lost to timekeeping, docket monitoring, and research verification.
The Build-or-Buy Question
Kirkland & Ellis made the loudest bet. In May 2026, Bloomberg Law reported, the firm committed $500 million to build a proprietary AI platform. Davis Wright Tremaine announced in August that it had deployed Harvey firmwide alongside Microsoft Copilot, aiming for 90 percent adoption. Managing Partner Jaime Drozd said in the announcement that combining legal-specific AI with enterprise productivity tools would let lawyers "apply their judgment to the most complex challenges."
Smaller vendors occupy the same territory. Venance markets "private legal AI" with on-premises deployment and matter-level isolation, according to its website. TerraPilot, from Lawgorithm, advertises hosting options that keep everything on private infrastructure. An open-source project called Enclave offers self-hosted legal AI inside a firm's own virtual private cloud, the GitHub readme explains.
Perceptron ML enters this niche at seed scale. Michael Marcotte previously led AI research at NVIDIA, building production agents for hardware debugging and formal verification; he holds degrees in mathematics and computer science from Stanford, Y Combinator's directory notes. Peyton Marcotte, still a mechanical engineering student at Brown, founded PMARC, which built astronaut exercise equipment with backing from NASA and the 1517 Fund. The company's case studies describe a system that verifies citations against established precedent—Zubulake, Residential Funding, FRCP 37(e)—and refuses to output a claim unless it traces to a document the firm uploaded.
In their YC launch, the brothers wrote that they "make law firms AI-native" by building systems "around how your firm already works, where no fact is used until it traces to a real document." The company declined to disclose funding beyond Y Combinator. Crunchbase lists the startup but locks the details.
What Comes Next
Thomson Reuters and LexisNexis are both moving toward what they call "agentic" workflows: end-to-end task automation grounded in authoritative content plus firm data. Thomson Reuters rebuilt CoCounsel around the Model Context Protocol and firm-level connectors, the company explained in blog posts published between March and August. LexisNexis announced in late August that Protégé now delivers "frictionless agentic productivity from first idea to review-ready legal work product."
The choice for firms breaks three ways: build, buy, or assemble. Kirkland's half-billion-dollar investment suggests some will construct governed stacks tailored to their practices. Others will layer Harvey, CoCounsel, or Protégé atop Microsoft 365 and Google Workspace, connecting via MCP. K&L Gates and Willkie both earned ISO/IEC 42001 AI governance certification early in the year, the firms announced in spring, signaling that formal standards may soon become table stakes.
Perceptron ML is wagering on a third path: bespoke systems trained on a single firm's matters and deployed behind the firm's firewall, with verification engines that refuse to guess. Whether managing partners write checks for private models or settle for vendor assurances of zero data retention, the direction seems fixed. Every vendor now emphasizes control, auditability, and citations over confidence.
The era of uploading client files to a shared cloud and hoping the terms of service hold may be ending. Perhaps faster than the platforms expected.
