[Note: Please welcome Laurent Wiesel, Principal at Justly Consulting. This article was originally published in LinkedIn. – GL]

Harvey AI bills its platform as providing a suite of products tailored to lawyers and law firms across all practice areas and workflows.

Harvey’s debut product video released on June 28, clocking in at 1:44, plays to a background of Mozart’s Piano Sonata No. 11 and no further audio. While light on explanation, the video introduces several features along with some interesting bells and whistles.

Harvey’s features, presented here in the order they appear in the left nav:

1. Client-Matter Number Integration

  • Critical feature for law firm operations
  • Helps enforce client policies and avoid data blending
  • Similar to established platforms like Lexis and Westlaw
  • Potential benefits for billing and ethical wall enforcement
  • Modal alert indicates ability to attach client-matter numbers to individual queries
Assistant feature with prompt loading and “save example”

2. Assistant

  • Offers chat and document Q&A capabilities
  • Currently provides single responses, no interactive chat yet
  • Prompt limit: 100,000 characters, which appears to reduce to 4,000 when even a single documents is added
  • Feature to save and load prompts, separated into Private, Team (collaboration), and Harvey (pre-built) varieties
  • Unexplained “Save Example” feature (briefly visible at the 0:18 mark)
  • Citations in assistant and research responses for quick verification

Research feature

3. Research

  • Expanded beyond Edgar data to include:
    • Tax Regulation
    • Case Law (EUR-Lex, US, French)
    • “Memos” (possibly internal)
  • Claims industry-leading datasets, but seems limited to specific regions for tax:
    • UK, Japan, Netherlands
    • PwC worldwide tax summaries
    • Case law coverage unclear and likely limited


Vault feature with file uploader

4. Vault

  • Bulk document review tool
  • Increased document limit for RAG from 50 in Assistant to 1000 in Vault
  • Bulk upload capability
    • Example: 283 documents uploaded, 404 files out of 1000 possible
    • Example processing time: 11 minutes for 121 documents
    • Project usage up to 1 GB
  • Can extract insights from uploaded data
    • Example shows contractual due diligence automation by extracting title, parties, dates, renewal terms, and termination rights from 70+ documents at once

5. Workflows

  • Described as orchestrating hundreds of specialized models for complex tasks and allowing collaboration to build bespoke use cases
  • Previously, workflows included language translation tool
  • Unclear if workflows entail actual model training, fine-tuning, or applications similar to OpenAI’s GPTs or Assistant API or other agentic framework

6. History

  • Query and response history (not shown)

7. Library

  • New feature “to help users maximize Harvey’s potential” (as stated in modal)
  • Details not provided

Quality Claims

  • Claims of industry-leading datasets may be overstated (currently there are rumors of interest by Harvey in acquiring vLex to bridge the case law and statute gap)
  • Document integration still limited compared to AI search capabilities offered by startups like DeepJudge or VecFlow.
  • No apparent integration with major document management systems (iManage, NetDocuments, OneDrive, Azure)
  • Claims superior performance on legal reasoning tasks, including outperforming GPT-4o, but lacks specific evidence
  • Security improvements with Microsoft Azure deployment (March 2024 press release)

Questions and Considerations

  • Timeline for interactive chat feature
  • Technical details of the “Save Example” feature
  • Data quality and relevance assurance across large document sets
  • Specific benchmarks for legal reasoning task performance
  • Comparison to established platforms in dataset completeness and accuracy
  • Extent of case law coverage
  • Remaining security concerns, if any, from potential clients

Based on personal experience with the product and updates gleaned from the video, Harvey has done a fine job going to market with an intuitive product poised to be legal’s ChatGPT.

Please let me know what I missed or got wrong.