ZetrixwebScope this correctly from the start
"AI for contract review" tends to get pitched as replacing legal judgment, which is the wrong frame and the reason a lot of these projects stall on trust. The realistic, valuable version is a first-pass assistant: it reads a contract, flags the clauses that deviate from your standard positions or carry real risk, and hands a lawyer a prioritized starting point instead of a blank document. The lawyer still makes every actual call. What changes is how much of the document they have to read cold before getting to the parts that need their judgment.
Why citations are the non-negotiable piece
The single most important design requirement for a document-review assistant is that every claim about what a contract says is checkable against the actual source text. When a document is uploaded with citations enabled, the response splits into blocks that each carry a reference back to the exact passage and location - a page number for a PDF, a specific character range for plain text. That means a flag like "the indemnification clause has no cap" isn't just an assertion - it points directly at the clause, so a lawyer can verify it in seconds instead of re-reading the whole document to confirm.
Without citations, a document assistant is trading one problem (reading the whole contract) for a worse one (trusting an unverifiable summary of it). Citations are what make the tradeoff actually favorable.
What goes in the Skill vs. what goes in the upload
The document: this specific contract
Uploaded fresh for every review - the actual text being checked, with citations enabled so every flag traces back to a real passage.
The Skill: your standing redline rules
What counts as an unacceptable indemnification cap, which jurisdictions need extra scrutiny, your standard liability limits - the same policy applied consistently to every contract, not re-explained each time.
Want to scope a first-pass review assistant against your actual contract templates? Talk to us about a legal-ops pilot.
Where to start - and where not to
Start with the highest-volume, most routine contract type your team reviews - usually vendor agreements or standard NDAs, not the bespoke, high-value deals where the stakes of a missed flag are highest. Prove the assistant's flagging accuracy against real historical contracts your legal team already reviewed manually, comparing its output to what a lawyer actually caught. Only extend it to higher-stakes document types once that accuracy has a real track record on the routine ones.
The workflow that actually saves time
| Step | What happens |
|---|---|
| 1. Upload | Contract uploaded with citations enabled, alongside the organization's redline Skill |
| 2. First-pass flag | Claude identifies clauses that deviate from standard positions, each with a citation to the source text |
| 3. Lawyer review | Counsel reviews the flagged clauses first, verifying each citation - not re-reading the entire document cold |
| 4. Judgment | The lawyer makes the actual call on every flag - accept, negotiate, escalate |
Key takeaways
- Scope this as first-pass triage, not a replacement for legal judgment - the lawyer still makes every real decision.
- Citations are non-negotiable: every flag should trace back to the exact source passage, so it's verifiable in seconds, not an unverifiable summary.
- The Skill carries your standing redline standards; the document upload is the specific contract - keep the two separated, the same way policy and live data stay separated in any Claude business assistant.
- Start with high-volume, routine contract types and measure flagging accuracy against real historical reviews before extending to higher-stakes documents.