Harvey is a legal AI built on foundation models for law-firm and in-house GC work. It answers research questions, drafts memos, summarizes contracts, and runs comparisons across a matter file. The workflow that compounds is a loop: the agent pulls from authoritative sources, the attorney reviews and cites, the work product lives where the next matter can reuse it. Without the last step, every matter starts cold.
The workflow
1. Frame the question against authoritative sources. Harvey, Westlaw Precision, and Lexis+ AI each ground answers in case law and statutes. Decide which corpus is authoritative before prompting. A securities question routes to Westlaw. A commercial dispute routes to Harvey plus the client's contract file.
2. Draft the memo with a tool that shows its work. Harvey produces a memo skeleton with citations. Hebbia runs the same query across the firm's document set and returns a matrix of evidence per sub-question. Both expose the source chain.
3. Cross-check. Run it through a second model. ChatGPT or Claude catches logical gaps the legal-tuned model glosses over. Spellbook handles contract language if the memo recommends a redline; see AI redlining with Spellbook.
4. Score the risk surface. Attach a risk score to any concluded position. Lexion and LinkSquares score against playbook deviation; see AI risk scoring.
5. File the work product where the next matter sees it. Tag by issue, jurisdiction, client industry, and outcome. The next associate lands on the prior memo, not a blank prompt.
Worked example: vendor indemnification question
A senior associate gets a question about indemnification caps in a vendor MSA. Harvey returns a 3-page memo with citations to recent Delaware decisions. Hebbia pulls every prior memo the firm wrote on indemnification caps in four years and surfaces three close analogs. The associate finds one inconsistency, resolves it by reading the underlying case, and runs the result through Claude for clarity. They file it tagged indemnification / Delaware / SaaS vendor. Six weeks later, a junior on a different deal searches the same tag and starts at draft 4.
The persistent-state pain
The memo lives in the DMS. The Harvey transcript lives in Harvey. The Hebbia matrix lives in Hebbia. The risk score lives in Lexion. The partner's approval lives in an email. Six months later the associate finds the memo but not the reasoning, not the cases the agent rejected, not the partner's comment on why the firm took a narrower position than the case law supports.
One way to solve this is a workspace like Dock that holds the agent-side interpretation: the path Harvey walked, the comparison Hebbia produced, the attorney's correction, the partner's approval. The DMS still owns the memo. Harvey still owns the prompt history. Dock rows carry pointers (harvey_session_id, dms_document_id) back to those systems. Irreversible steps gate through two-key handshakes. Scoped credentials for the agent are covered in agent identity lifecycle. The cross-use-case pattern is in Dock for legal.
Why it matters
Legal work compounds when prior reasoning is reachable. Harvey gives the firm leverage on the input side. The output, the memo and the why behind it, has to live somewhere the next matter looks. Otherwise the firm pays Harvey to redo work it already did.
Read the full pillar: how to do legal review with AI.
FAQ
Is Harvey a replacement for Westlaw or Lexis? No. Harvey runs on top of foundation models and firm content. Westlaw Precision and Lexis+ AI remain primary for case law and statutory research. Most firms use Harvey alongside, not instead.
How do I evaluate Harvey against general models like Claude or ChatGPT? Run the same prompt through both on a representative matter. LegalBench (legalbench.org) is the public reference for comparing model performance on legal reasoning.
Where does Hebbia fit in the stack? Hebbia runs a structured question across a large document set and returns a matrix of answers. Use it for due diligence and matter file review where Harvey's memo format is too linear.
How do I keep work product reusable across matters? Tag every output by issue, jurisdiction, client type, and outcome. Store the agent rationale next to the memo. Stanford CodeX (law.stanford.edu/codex) tracks how computational legal practice is evolving.