On May 2, 2026, Will Chen—a former associate at Latham & Watkins—released Mike, an open-source legal AI platform he built in two weeks using Claude. The platform replicates the core feature set of Harvey and Legora: a chat interface that reads documents and cites verbatim, matter-scoped workspaces, tabular review for bulk document extraction, and multi-step workflows for contract drafting and analysis. Users supply their own API keys from Anthropic or Google. There is no subscription, no per-seat licensing, and no vendor lock-in. The GitHub repository passed a thousand stars and three hundred forks within seventy-two hours—among the highest engagement figures Legal IT Insider has reported for a legaltech project on the platform.
The reaction was immediate and, for the most part, predictable. On X, the framing was disruption: a single developer had replicated what two companies valued at a combined $16 billion or more sell to AmLaw firms for five- and six-figure annual contracts. On Hacker News, the conversation split between enthusiasm and skepticism about the repository’s maturity. The most widely quoted take came from James Harrison, former IT Director of Leigh Day, who told Legal IT Insider that Mike “doesn’t kill Harvey or Legora, but it absolutely changes the negotiation. Once a working open-source alternative is sitting on GitHub, the conversation in renewal meetings moves from ‘Is this magic?’ to ‘What exactly am I paying enterprise prices for?’”
That is a reasonable question about pricing. But the conversation it has produced—about moats, valuations, vendor margins, and feature parity—treats the profession’s AI challenge as a capability problem. It assumes the bottleneck is access to the tool: firms that can afford Harvey have AI; firms that cannot are locked out; Mike levels the field. The framing is clean, and the market dynamics are real. What the framing misses is that the profession’s AI problem has not been a capability problem for some time. The tools have been ahead of the lawyers using them for at least two years. Mike raises the ceiling further, but the difficulty has never been at the ceiling.
Ceiling and floor
For the past eighteen months, the legal AI conversation has been dominated by capability announcements. Thomson Reuters launched CoCounsel Legal with “fiduciary-grade AI” that performs “just as well as a senior associate.” Harvey processes more than 400,000 agentic queries per day. LexisNexis shipped Protégé with agentic drafting that promises “review-ready work product in minutes.” And now Mike demonstrates that the core feature set of these platforms—document analysis, contract review, citation-grounded chat, bulk extraction—can be reproduced by a single developer in two weeks. Each announcement raises the ceiling of what legal AI tools can do, and each one is a capability story.
The floor—what lawyers and law firms are doing with these tools in practice—has barely moved. As of 2024, Clio’s Legal Trends Report found that 53 percent of law firms had no AI use policy at all. The Q1 2026 sanctions data showed that none of the lawyers sanctioned for AI-generated citation errors had functioning verification practices. Sullivan & Cromwell, with everything the sanctioned lawyers lacked—mandatory training, tracked completion, written verification protocols—still filed a motion with roughly 40 corrupted citations. The floor of professional AI practice is not rising with the ceiling. It may not be rising at all.
The commentary about Mike has focused entirely on the ceiling. Can a free tool match a $16-billion-plus industry? How thin is the enterprise moat? What are firms paying for? These are interesting market questions. They are not the questions the profession should be asking, because the profession’s failures are not happening at the capability frontier. They are happening at the baseline, with the simplest tools, on the most basic obligation—checking whether the cases you cite exist.
Why more capability does not solve a floor problem
The assumption embedded in the Mike commentary—and in much of the legal AI discourse generally—is that better tools produce better outcomes. If Harvey is too expensive for small firms, give them Mike. If consumer chatbots lack document-grounding features, build platforms with verbatim citation. If verification is hard, ship tools that include page-level source references. The logic is intuitive and, at the level of product design, defensible. Tools with citation-grounding features are better than tools without them.
But the Q1 sanctions cases did not fail because the tools lacked features. They failed because the lawyers did not verify the output. Brigandi filed 23 fabricated citations across three briefs even after opposing counsel flagged the earlier errors. Lake submitted a brief with 57 defective citations. Seth used ChatGPT under time pressure and did not check a single citation. In none of these cases would a more capable tool have changed the outcome, because the failure was in the lawyer’s response to the output—or, more precisely, the absence of one.
Mike’s tabular review and verbatim citation features are better engineered than ChatGPT’s raw output. A lawyer using Mike is less likely to encounter a fabricated citation than a lawyer pasting a question into a consumer chatbot. But the gap between “less likely to encounter an error” and “will verify the output before filing it” is the gap between ceiling and floor, and no amount of product improvement closes it. The lawyer who does not verify ChatGPT output will not verify Mike output either, because the problem was the lawyer’s practice, not the tool’s reliability.
The parallel nobody wants to draw
The structural pattern here resembles the one I described in an earlier post about APIs: switching to a better technical posture solves one problem while leaving the compliance architecture untouched. The version with Mike is more uncomfortable because it challenges the democratization narrative the profession has embraced.
Mike gives every firm access to a capable legal AI platform. The firms best positioned to use it well are the ones that have already built the institutional infrastructure: written AI use policies, verification workflows, supervisory protocols, trained lawyers who understand both what the tools can do and where they fail. Those firms are, overwhelmingly, the large firms that could afford Harvey in the first place. The firms where Mike could provide the most value—small and midsize practices priced out of the enterprise market—are also the firms least likely to have any of that infrastructure in place. For those firms, Mike offers a more capable tool and the same absence of institutional support that produced the Q1 sanctions.
This does not mean Mike is a bad project or that Chen has done something irresponsible. The software is well-built, the open-source model is sound, and the pricing transparency is a genuine service to a market that has been opaque about what firms are paying for. But the celebration of Mike as democratization assumes that access to the tool was the constraint. For firms without AI governance, the tool was never the constraint. The constraint is the set of practices—verification, supervision, judgment about which tasks to delegate and which to retain—that sit between the tool and the work product. Those practices cannot be forked from a repository.
What Mike unbundles
The most useful way to read the Mike OSS moment is as a decomposition of the legal AI value proposition. Mike reveals which components of what Harvey and Legora sell are commodity software and which are something else.
The interface layer—chat, document upload, tabular review, workflow orchestration—is commodity software, as Chen demonstrated by building it in two weeks. The model layer was always a commodity from the firm’s perspective; Harvey and Legora license the same frontier models that Mike users access through their own API keys. Neither of these layers justifies enterprise pricing, and firms negotiating renewals should say so, because the gap between the free tool and the six-figure platform is not in what the AI can do.
What remains is the integrations and the institutional infrastructure. Harvey’s 25,000 custom workflows, its DMS connectors, its ethical-wall enforcement, its audit trails, its enterprise authentication—these are the features that cannot be reproduced in two weeks, because they require integration with the firm’s existing systems, policies, and workflows. Whether Harvey and Legora deliver on these features well enough to justify their pricing is a question individual firms should press hard. But the existence of these features reflects a recognition—at least among the firms that buy them—that using legal AI responsibly requires more than the AI itself.
The question Mike poses for large firms is whether what they are paying for beyond the commodity layers functions as working infrastructure or as an expensive way to avoid building their own. The question Mike poses for smaller firms is where the institutional competence will come from if not from a vendor—and what happens to the clients whose work passes through the tool before that question is answered. Both questions are about the floor, and neither can be resolved by raising the ceiling.
The floor is the story
Chen told Artificial Lawyer that “thin wrappers without unique value will be replaced by platform providers with distribution advantages.” He is probably right about the market dynamic. But the market dynamic is a ceiling story—which vendors survive, which products win, where the margins settle. The profession’s problem is a floor story, and the two have remarkably little to do with each other.
California’s proposed amendments to the Rules of Professional Conduct would require independent verification of all AI output and functioning AI governance under Rules 5.1 and 5.3. Those obligations attach to the lawyer and the firm, not to the vendor or the tool. They apply whether the firm uses Harvey at six figures, Mike at zero, or Claude through a direct API integration. The tool’s price does not change the obligation. The tool’s capability does not satisfy it.
The legal AI ceiling will keep rising. Vendors will ship new features, open-source projects will replicate them, and the capability gap between the most expensive platform and the cheapest alternative will continue to narrow. None of that changes the fact that in Q1 2026, courts sanctioned lawyers $145,000 for failing to do the one thing every tool already makes possible: checking whether the cited cases exist. The floor of professional practice has not kept pace with any of them—the six-figure platform, the free repository, or the consumer chatbot that costs nothing at all.
This post draws on coverage of Mike OSS from Artificial Lawyer, Legal IT Insider, Lucio AI, and the Mike OSS GitHub repository; social media commentary on X and Hacker News; Clio’s 2024 Legal Trends Report; and product announcements from Thomson Reuters, Harvey, and LexisNexis. It extends arguments from prior posts on data-handling compliance, judgment delegation, verification practices, sycophancy, and agentic AI supervision.