Contract review AI and how to use it without missing what matters

A contract review done with AI fails in the same places a rushed human review fails. The reviewer reads the document without knowing which side the firm represents, without the firm's standard positions, and without a list of what this client cares about. The output then looks thorough and misses the one clause that matters, such as an uncapped indemnity buried in a definitions section or an auto-renewal with a ninety-day notice window.

The fix is not a better prompt. It is the same discipline a senior associate brings to a review: define the scope, set the standard you are measuring against, read every clause against that standard, and record why each change is proposed. AI makes each of those steps faster. It does not make any of them optional.

• Decide whose interests the review protects before anything is uploaded.
• Measure the contract against a written standard, not against the model's general sense of fairness.
• Keep a record of every proposed change and the reason for it.
• Treat everything the AI produces as a draft for attorney review.

Most of the quality of an AI-assisted review is decided before the first page is read. Spend five minutes on setup and the rest of the review gets shorter and more accurate.

Start with confidentiality. ABA Model Rule 1.6 requires reasonable efforts to prevent unauthorized disclosure of information relating to the representation, and ABA Formal Opinion 512 (July 2024) applies that duty directly to generative AI tools. Before a client contract goes into any AI system, you should know whether the provider keeps the text, whether it is used to train models, who inside your firm can see it, and whether your engagement letter or the client's outside counsel guidelines say anything about AI use. If you cannot answer those questions, the contract should not go into that tool.

Then collect the inputs that make the review specific to this deal.

• The side you represent (buyer or seller, licensor or licensee, landlord or tenant).
• The contract type and governing law, if already known.
• The client's business priorities, such as payment terms, exit rights or data use.
• The firm's playbook or standard positions for this contract type, if one exists.
• Any prior version, term sheet or letter of intent the contract should match.

The first pass is where AI earns its place. A good tool reads the full document, including schedules and exhibits, and gives you a structured map you can verify against the page.

Ask for outputs you can check, not opinions you have to trust. A useful first pass contains a short summary of the commercial deal, the parties and their roles, all dates and periods (effective date, term, renewal, notice periods, cure periods), payment obligations, defined terms with their definitions, and a clause-by-clause list of risks ranked by severity. Every item should point to the section and page where it comes from, so a reviewer can open the source in one click.

Watch for the things AI tools most often get wrong on a first pass.

• Cross-references that change meaning, such as a limitation of liability that excludes obligations defined elsewhere.
• Definitions that quietly expand scope, such as an affiliate definition that pulls in future entities.
• Exhibits and order forms that override the main agreement through an order of precedence clause.
• Scanned pages where OCR misread a number, a date or a percentage.

A generic risk list tells you what a contract could say. A playbook review tells you how this contract differs from what your firm or your client accepts. The second is what clients pay for.

If your firm has a written playbook, give the AI the positions for each clause type: preferred position, acceptable fallback, and walk-away point. Then ask it to classify each clause as matching, deviating within the fallback, or outside the acceptable range. If you do not have a playbook, build a short one from your last few negotiated deals of the same type. Even ten positions on the clauses that matter most (limitation of liability, indemnity, termination, assignment, confidentiality, IP ownership, governing law, dispute resolution, payment, auto-renewal) will improve the review more than any change to the tool.

• Limitation of liability, including carve-outs and whether the cap covers indemnity.
• Indemnification scope, defense control and notice requirements.
• Termination for convenience, termination for cause and cure periods.
• Assignment and change of control.
• Confidentiality term and return or destruction of information.
• Ownership of work product and license-back rights.
• Data protection obligations and breach notice timing.
• Governing law, venue and arbitration terms.

A risk list is not a markup. The deliverable most clients expect is a tracked-changes document with comments that explain each change in a sentence the other side can accept.

Ask the AI to propose specific replacement language for each deviation and to state the reason in plain words, tied to your playbook position. Then edit. You will often shorten the proposed language, soften the tone in comments for a relationship-sensitive counterparty, or drop a change that is not worth the negotiation capital. The attorney decides which battles to fight. The AI only makes sure none of them are missed.

Before the markup leaves the firm, read it the way opposing counsel will. Check that every defined term you introduce is defined, that section numbers in cross-references still match after edits, and that the comments do not reveal internal strategy or the client's priorities.

The ethics rules do not prohibit AI-assisted review. They set the conditions for it. ABA Model Rule 1.1 and its Comment 8 require competence, including an understanding of the benefits and risks of the technology you use. Rule 1.6 covers confidentiality, as described above. Rules 5.1 and 5.3 make supervising lawyers responsible for the work of other lawyers and nonlawyer assistance, and Formal Opinion 512 treats AI tools as a form of assistance that needs supervision, firm policies and training. The same opinion also addresses communication with clients about AI use and reasonable fees when AI reduces the time a task takes.

In practice, that means a few firm-level decisions.

• Write down which tools are approved for client documents and which are not.
• Require attorney review of every AI output before it reaches a client or a counterparty.
• Train paralegals and associates on what the tool does well and where it fails.
• Bill for the time actually spent, not the time the task used to take.

AI review also has limits no policy removes. It cannot know commercial context that is not in the document or your instructions. It can misread poor scans. It can be confident about a clause it misunderstood. None of that makes it unusable. It makes the attorney's read of the high-risk clauses the part of the review you never skip.