Implementation beats model training in legal AI

The winning moat in legal AI is moving from who trains the best model to who can get embedded in how lawyers already work every day. Harvey’s growth from $50M ARR at the end of 2024 to $150M by November 2025 came after it dropped its custom model, went multi model, and leaned harder into implementation, with ex lawyers helping firms wire the product into Word, document systems, and repeat drafting workflows that actually get used.

• In practice, implementation means fitting into the exact places legal work happens. Harvey added integrations with iManage and NetDocuments and competes inside Microsoft Word, where recurring drafting and redlining tasks can see time savings of up to 70%, making adoption visible to partners and associates.
  2 sacra
• This is why off the shelf model users can still win. Legora skipped the costly fine tuning path, built strong Word and document workflow products, sells large enterprise contracts around $280,000 annually, and expands from pilots into firm wide licenses by integrating with the systems firms already trust.
  3 sacra 4 sacra
• The strongest incumbents also win on distribution, not just AI quality. Clio grew to about $400M ARR by October 2025 by owning practice management, billing, payments, and documents, while Thomson Reuters and LexisNexis still benefit from large installed bases, data rights, and long standing enterprise contracts.
  2 sacra 5 sacra 6 sacra

From here, legal AI will look more like an implementation heavy enterprise software market than a model race. The platforms that control document access, Word drafting, search, security, and customer rollout will keep compounding, while standalone legal models get absorbed into broader workflow products and incumbent distribution channels.