AI contract clause review with source citations.

What a cited clause review gives the lawyer

A cited clause review is a table, not an essay. Each row names an issue, the firm’s preferred position, the clause number, the quoted words that create the issue, how far they depart from the playbook, and a suggested comment or markup. A finding that cannot quote the contract is either dropped or shown separately as unsupported. Absence findings (no limitation of liability, no termination for convenience) cite the playbook entry and list the headings that were searched.

The lawyer’s job shifts from reading 30 pages cold to verifying 15 to 25 specific claims, then reading the agreement once for context. It suits firms that see many agreements of a few known types, and it is a document processing build with a playbook at its centre.

How the review runs

Inputs

The contract as Word or a text PDF, the playbook for that contract type (preferred position, acceptable fallbacks and red lines for each clause), the client’s instructions and commercial context, and earlier negotiated versions where they exist.

Systems

The firm’s document management system, a model deployment under terms that exclude training on inputs, a quote checker written in ordinary code, and Word for comments and tracked changes.

Steps

Convert the contract keeping clause numbering; resolve defined terms and cross-references; classify each clause; compare it with the playbook; write findings with quotes; check every quote verbatim against the cited clause; retry once on a failed check, then mark it unverified; rank by severity.

Human review

The responsible lawyer reads every finding against its clause and reads the whole agreement, including schedules. A partner signs off any red-line issue and any advice going to the client.

Outputs

A review table, draft Word comments, and a list of clauses the system could not classify.

Measures

Share of quotes passing the verbatim check, recall against a lawyer’s issues list on a test set of past agreements, lawyer minutes per contract, and findings dismissed as noise.

Risks

Misread definitions, missed schedules, quotes that are accurate but misinterpreted, confidentiality breaches, and lawyers trusting the table instead of the contract.

Illustrative example: a nine-lawyer Nelson commercial team

Illustrative example, synthetic numbers: the team reviews 60 supply and services agreements a month, averaging 28 pages. A first review by a junior lawyer takes about 150 minutes, then a partner spends 30 minutes. With a cited draft, the junior verifies findings and reads for context; the firm assumes 95 minutes after a pilot.

The fee line assumes 31,000 input tokens per pass (the agreement, playbook extract and instructions) and 3,000 output tokens, on Claude Opus 5.5 at US$4 per million input and US$20 per million output tokens, as listed on 19 October 2026. That is about US$0.18 a pass; a missed indemnity costs far more than the gap to a cheaper model.

Why the citation check is code, not a prompt

Telling a model to “only quote real text” reduces errors without removing them. The dependable control is ordinary software that searches the cited clause for each quoted string, normalising whitespace and curly quotes, and rejects anything that does not match. Our guide to structured outputs and validation shows the pattern. Lucent, a Levy Street venture in investment research, is built on the same principle: outputs grounded in citations a reader can open. For a manual version of the habit, see summarising documents with citations.

Professional rules that shape the design

• Responsibility stays with the lawyer. The Law Society’s generative AI guidance says a lawyer is not absolved from responsibility for defects in an end product, such as a contract, because it came from generative AI. It points to rules 3 (competence) and 10.9 (misleading conduct), and to rules 11 and 11.1 on supervising staff use.
• Confidentiality. Chapter 8 of the conduct rules applies to what is sent to any AI provider. The courts’ guidelines for lawyers advise against entering confidential or privileged information into public chatbots.
• Client information. Rule 3.4 requires written client care information in advance; the guidance suggests revisiting it, and billing, when AI is used.
• Unfair contract terms. Since 16 August 2022 the Fair Trading Act 1986 extends unfair contract terms rules to standard form small trade contracts under a $250,000 annual value threshold. A playbook for clients using standard terms should flag clauses at risk.

What stays with the lawyer

The lawyer decides whether a deviation matters to this client, what to concede, how to word the advice and what goes to the other side. The system never sends comments to a counterparty, never edits the master document, and never reviews contract types without a playbook.

Where clause review goes wrong

• Definitions change the meaning. A liability cap looks standard until “Losses” is defined to exclude most claims. Resolve defined terms into each clause before comparing.
• The operative term is elsewhere. It sits in Schedule 4 or in standard terms incorporated by reference that nobody supplied. List every incorporated document and flag those missing.
• Accurate quote, wrong conclusion. The verbatim check proves the words exist, not that the reasoning holds. That is the lawyer’s read.
• Silence is hard to see. Models notice what is present more readily than what is absent. Run absence checks from the playbook list, clause by clause.
• Scanned agreements. Optical character recognition can drop a “not”. Prefer Word originals and route low-quality scans to manual review.
• A stale playbook. The system enforces last year’s position if nobody updated it. Give each playbook an owner and a review date.

Clause review is one of three legal workflows we map. The others are intake and matter triage and chronology extraction from case documents; the legal sector guide sets out how they connect.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.