Harvey pivots to legal workflow OS

This marked the end of legal reasoning as Harvey's core moat, and the start of workflow orchestration as the real product. Once Gemini 2.5 Pro, OpenAI o3, xAI Grok 3, and Anthropic Sonnet 3.7 beat Harvey's custom model on BigLaw Bench, owning a fine tuned legal model stopped mattering. Harvey shifted to chaining the best external models with legal tools, document systems, and implementation support inside firms.

• The product changed from one smart model to a legal work system. Harvey now routes different tasks to different models, then wraps them with pre built workflows and a custom agent builder for firms. That is a more durable wedge because it fits how lawyers actually work, across drafting, review, research, and approvals.
  1 sacra 2 sacra 6 bloomberg
• This also exposed Harvey's weak point in research. Its June 18, 2025 LexisNexis alliance gave Harvey citation backed answers and Shepard's inside the product, but through an integration rather than ownership of the corpus. That helps research quality, but leaves Harvey dependent on someone else's data moat.
  3 sacra 4 lexisnexis 5 harvey
• Competitors adapted faster by skipping the custom model step entirely or by buying data and workflow assets. Legora went straight to off the shelf frontier models for international firms. Clio bought vLex for $1B to combine practice management with a 1B plus document research corpus, turning legal AI into a distribution and data race.
  2 sacra 7 sacra 8 prnewswire

Going forward, the winners in legal AI will look less like model companies and more like operating systems for legal work. Harvey is already moving that way, with agent building, document system integrations, and service heavy rollout. The next battle is who can combine trusted legal sources, internal firm knowledge, and daily workflow entry points into one product lawyers use all day.