Practical guide

How to create a contract review with AI

Use AI to accelerate issue spotting and evidence organization, not to decide what a contract means or whether it should be accepted. A defensible review keeps the signed source, the reviewer’s instructions, every extracted clause, and the accountable professional’s final judgment connected in one record.

By Two Prune

Use AI to accelerate issue spotting and evidence organization, not to decide what a contract means or whether it should be accepted. A defensible review keeps the signed source, the reviewer’s instructions, every extracted clause, and the accountable professional’s final judgment connected in one record.

Start with the decision and authority boundary

A contract review is useful only when it is tied to a named decision: sign, reject, negotiate, escalate, or request missing information. Before opening an AI tool, record the agreement version, parties, governing-law assumption, business purpose, deadline, approved playbook, reviewer, and person authorized to accept risk.

The American Bar Association’s Formal Opinion 512 says lawyers using generative AI must consider competence, confidentiality, communication, supervision, accuracy, and reasonable fees. It also emphasizes understanding a tool’s capabilities and limitations and reviewing its output. Those duties depend on jurisdiction and engagement; this page is an operational artifact guide, not legal advice. ABA Formal Opinion 512

Do not send confidential terms to an unapproved service. Record the tool, account or tenant, data-handling setting, permitted input class, retention rule, and approval reference. If those facts are unknown, remove sensitive material or hold the AI-assisted step until an authorized owner resolves them.

Sources: [1]

Build a RAISE review record

RAISE separates extraction from judgment so a reviewer can reproduce every material finding.

NIST’s AI Risk Management Framework describes governance as a cross-cutting function with documented roles, monitoring, risk priorities, and accountability structures. RAISE adapts those general principles to a review artifact; it does not establish that a particular legal control is effective. NIST AI RMF Core

  • R — Reference set: freeze the exact agreement, schedules, amendments, incorporated policies, commercial proposal, and comparison playbook. Give every file a version and date.
  • A — Authority and acceptance: name who may interpret the clause, approve a deviation, accept residual risk, and communicate with the counterparty.
  • I — Issues with anchors: record the clause, page, quoted text, playbook position, variance, operational consequence, and confidence. Never keep an issue that cannot be traced back to the source.
  • S — Source verification: reopen the original files and verify quotations, defined terms, cross-references, calculations, dates, and external law or policy against authoritative sources.
  • E — Escalation and evidence: assign a disposition, owner, due date, proposed language, unresolved question, and final approval. Preserve both accepted and rejected AI suggestions.

Sources: [2]

Use AI only for bounded passes

Run small, inspectable passes instead of one prompt asking for a legal conclusion.

For every pass, specify the allowed source files, required output fields, and refusal rule. A useful refusal rule is: “If the contract does not state the answer, mark it not found and name the source locations checked.” AI output belongs in a working layer; it must not silently overwrite the signed source or the professional review.

  • Create a clause inventory with page and section anchors.
  • Extract defined terms and build a definition-to-use map.
  • Compare named clauses with the approved playbook without inventing missing standards.
  • Identify conflicting dates, amounts, notice periods, renewal mechanics, and cross-references.
  • Draft questions and alternative language for an authorized reviewer to assess.
  • Produce a change log that links each proposed edit to a verified issue.

Copyable issue register

Use one row per material issue:

| ID | Source anchor | Verified text | Playbook position | Business consequence | AI contribution | Human finding | Owner | Disposition | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | CR-01 | Agreement § / page | exact quotation | standard / fallback / prohibited / absent | concrete operational effect | extraction / comparison / draft | verified judgment and uncertainty | named person | accept / negotiate / escalate / hold |

Add columns for jurisdiction, confidentiality class, affected defined terms, linked schedules, proposed language, approver, approval date, and residual risk when they change the decision. Store the prompt and output identifier only when policy permits; otherwise retain a reproducible instruction summary.

Worked example: renewal and data deletion

The following example is synthetic. A software agreement says the service renews automatically for another year unless either party gives notice 30 days before renewal. A schedule says customer data is deleted “after termination” but gives no deadline. The playbook requires a 60-day renewal notice and deletion within 30 days.

The AI pass extracts both clauses and flags the variance. It also proposes “delete promptly” as replacement language. The reviewer rejects that suggestion because it is less measurable than the playbook. CR-01 records the renewal anchor, a calendar consequence, the 30-day variance, and a proposal for 60 days. CR-02 records the deletion clause, the missing deadline, affected data categories, and a proposal for a 30-day deletion obligation plus confirmation.

The authorized reviewer then checks whether another schedule changes either term, assesses governing law and commercial leverage, and chooses a negotiation position. The evidence artifact preserves the original text, AI extraction, rejected wording, final language, approver, and remaining uncertainty. The example demonstrates a method; it does not determine what any real contract requires.

Verification tests before approval

Run these tests against the actual files, not the AI summary:

A pass means the review is traceable and ready for its named professional decision. It does not mean the agreement is safe, enforceable, complete, or suitable.

  • Completeness: every required clause and schedule is present or explicitly marked missing.
  • Anchor: a second reviewer can reach every quoted term from the recorded page and section.
  • Definition: each defined term is used consistently and no capitalized term is silently reinterpreted.
  • Cross-reference: every cited section exists and points to the intended obligation.
  • Number: amounts, dates, notice periods, caps, formulas, and currencies match the source.
  • Contradiction: the main agreement, schedules, amendments, and proposal do not create an unresolved conflict.
  • Authority: each departure has an authorized owner and recorded decision.
  • Confidentiality: the AI use and retained evidence comply with the applicable data-handling rule.

Hold and escalation conditions

Hold the review when the governing document set is incomplete, the source version is disputed, privileged or confidential material cannot be handled safely, the playbook is absent, an issue depends on jurisdiction-specific interpretation, or no one has authority to accept the risk. Escalate unverifiable citations, conflicting terms, material calculations, regulatory questions, unusual indemnities, security obligations, and any AI suggestion that changes economic or legal effect.

Two Prune may use a contract-review task inside a realistic AI-enabled work assessment. It can evaluate how a participant frames the task, navigates evidence, uses AI, verifies claims, exercises judgment, and communicates a deliverable. It does not provide legal advice, approve contracts, or make employment decisions.

Sources

American Bar Association: Formal Opinion 512 — Generative Artificial Intelligence Tools

  • National Institute of Standards and Technology: AI Risk Management Framework Core

Sources: [1] [2]

Sources

  1. 1.ABA Formal Opinion 512 · www.americanbar.org
  2. 2.NIST AI RMF Core · National Institute of Standards and Technology

Explore this topic