The Myth of Legal AI: Why No One is Doing It Right (...Except Us)

In a widely circulated analysis, Irys founder Sabih Siddiqi argues that the legal AI market has consistently made the same set of architectural errors, producing tools that perform impressively in controlled settings but fall short in actual legal practice.

The analysis identifies three compounding failures: general-purpose language models that lack a reliable internal hierarchy of legal authority; retrieval-augmented generation pipelines that were designed for enterprise document search rather than legal citation structures; and output formats that make it impossible for supervising lawyers to verify what was retrieved versus what was generated.

Siddiqi's argument has attracted significant attention among legal technology practitioners, with several large-firm technology partners publicly agreeing that the gap between demo performance and practice-ready performance remains the central problem in the category.

"Doing legal AI right means treating the professional responsibility standard as a design specification, not a compliance afterthought," Siddiqi writes. "That requires harder engineering decisions than the market has been willing to make — and it is why we built Irys the way we did."