The Second Wave of Legal AI: From Research to Reasoning
The first wave of legal AI automated search. Lexis and Westlaw pioneered Boolean and keyword search over case law databases decades ago, and the AI enhancement of that search — better relevance ranking, semantic search, conceptual retrieval — produced the legal research tools that have been deployed in law firms for the past decade. This wave produced significant efficiency gains and is now essentially commoditized.
The second wave is attempting to automate reasoning — the application of legal principles to specific fact patterns, the identification of risk in novel contractual arrangements, the analysis of regulatory agency intent from enforcement actions and guidance documents. This is a qualitatively different capability, and understanding the difference — and the liability implications of operating at this layer — matters enormously for anyone building legal AI products in 2026.
The distinction is not merely technical. Legal search tools produce outputs that attorneys recognize as information retrieval — analogous to what a law clerk does when tasked with finding cases on a specific point. The attorney then applies legal judgment to the retrieved information. Legal reasoning tools produce outputs that are structurally similar to legal conclusions — an analysis of whether a specific contract clause creates material liability, a prediction of how a regulator will characterize a specific practice. This output is what the attorney is paid to produce, and its quality directly affects the client's legal position.
The companies succeeding at the second wave are solving this accountability problem in two ways. First, by maintaining citation chains — every reasoning output is grounded in specific statutory text, case law, or regulatory guidance that the attorney can verify. Second, by making the uncertainty visible — reasoning outputs that include explicit confidence assessments and the specific scenarios under which the analysis would change are more useful to attorneys than outputs that present conclusions without qualification. Legal reasoning AI that acknowledges what it does not know earns attorney trust faster than AI that overstates its certainty.
We are actively evaluating several companies building at the legal reasoning layer in our Fund II pipeline. The market is large, the capability is real, and the compliance architecture requirements — citation integrity, professional responsibility alignment, data confidentiality — are exactly the kind of moat we look for. The companies that get this right will define the legal AI market for the next decade.