Document Processing · Enterprise Knowledge Search
Harvey needed a legal AI system that could reason over vast case-law and contract corpora with higher accuracy and fewer hallucinations than generic fine-tuning or basic retrieval could provide for professional legal work.
Reported impact
Harvey improved legal-research quality and citation grounding, won strong attorney preference over baseline GPT-4 output, and created a foundation for broader drafting, contract analysis, and agent-assisted professional-services workflows.
Harvey partnered with OpenAI to build a custom-trained model for legal professionals after finding that standard fine-tuning and retrieval-augmented generation could not reliably handle open-ended case-law research. The company combined its secure professional-services platform with a custom case-law model that injects large volumes of legal knowledge and domain-specific reasoning into the base model, first with Delaware case law and then across all U.S. case law. Harvey added the equivalent of 10 billion tokens of data and evaluated the system side by side against GPT-4 with attorneys from 10 major law firms. The resulting workflow helps lawyers ask complex research questions, draft documents, compare large contract sets, and receive longer answers grounded in citations, while Harvey continues exploring agent patterns that combine multiple model calls into more complete legal work products.
Lawyer questions and document-review tasks enter the Harvey platform, where the custom-trained case law model reasons over the added legal corpus and returns answers with supporting citations. Side-by-side evaluations and attorney feedback then inform further tuning of the model and future agent workflows.
"97% of the time, the lawyers preferred the output from the case law model"
"Harvey partnered with OpenAI to create a custom-trained case law model."
"Not only does the case law model not make up cases, but every sentence is actually supported with the case it’s citing"
"They added the equivalent of 10 billion tokens worth of data to power the custom-trained case law model."
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
The company profile states 'Operating headquarters' in 'San Francisco, CA, US'.
Related
OpenAI Sample Aug 20, 2026 Stampli Stampli needed to launch complex new products and keep product knowledge current without waiting on scarce design, contractor, and operations bandwidth. 243 modeled active role-hours compressed to about 77 for the Deep Finance launch ChatGPT Work · Codex · Deep Finance OpenAI Sample Mar 7, 2025 Hebbia Hebbia needed a way to help finance and legal professionals search, interpret, and synthesize huge volumes of private offline documents where standard RAG systems miss implicit answers and manual review takes days or… $2,000 per hour in legal fees GPT-4o · Matrix · OpenAI o1 OpenAI Sample Dec 11, 2024 Rogo By fine-tuning OpenAI’s models and integrating a broad set of financial datasets such as S&P Global, Crunchbase, and FactSet, Rogo delivers real-time financial intelligence to thousands of financial professionals,… Automated diligence : Rogo integrates private data rooms, generates tailored question lists, offers writing… GPT-4o · OpenAI API · o1 