Practical guide

AI contract review verification checklist

An AI-assisted contract review is not verified because the model found many clauses or produced a polished summary. Verification means an accountable reviewer can trace every material finding to the controlling document, reproduce the comparison, test the risky details, and see who accepted any remaining uncertainty. Use this checklist after issue spotting and before a contract recommendation is.

By Two Prune

An AI-assisted contract review is not verified because the model found many clauses or produced a polished summary. Verification means an accountable reviewer can trace every material finding to the controlling document, reproduce the comparison, test the risky details, and see who accepted any remaining uncertainty. Use this checklist after issue spotting and before a contract recommendation is treated as ready for negotiation, approval, or escalation. It is an operational aid, not legal advice.

Set the review boundary

Complete these fields before checking the output:

If the document set or authority chain is incomplete, stop. More model output cannot repair an undefined decision boundary.

The American Bar Association’s Formal Opinion 512 discusses lawyers’ duties when using generative AI, including competence, confidentiality, communication, supervision, accuracy, and reasonable fees. It also emphasizes understanding a tool’s limitations and reviewing its output. Applicability depends on jurisdiction and engagement, so the opinion supports the verification discipline here without determining a real contract’s legal effect. ABA Formal Opinion 512

  • [ ] Decision required: sign, negotiate, reject, escalate, or request information
  • [ ] Exact agreement version, date, parties, schedules, amendments, and incorporated documents
  • [ ] Business purpose, deadline, governing-law assumption, and materiality threshold
  • [ ] Approved playbook or fallback positions used for comparison
  • [ ] Named legal reviewer, business owner, risk approver, and final decision maker
  • [ ] AI tool, tenant, data-handling setting, retention rule, and approved information class

Sources: [1]

Run the TRACE checklist

TRACE turns a draft review into an inspectable evidence packet.

### T — Text and source control

### R — Rules and comparison basis

### A — Arithmetic, dates, and cross-references

### C — Consequence and control

### E — Evidence and exception closure

NIST’s AI Risk Management Framework Core calls for documented roles, clear accountability structures, human oversight, monitoring, and review. TRACE applies those general governance outcomes to a contract-review artifact; passing it does not establish that a legal control or AI system is safe or effective. NIST AI RMF Core

  • [ ] Every quoted clause has a page, section, or paragraph anchor.
  • [ ] The quoted wording matches the frozen source, including defined terms and exceptions.
  • [ ] The main agreement, schedules, amendments, order form, and incorporated policies were searched.
  • [ ] Optical-character-recognition errors, tables, footnotes, and handwritten changes were checked against the original rendering.
  • [ ] Missing or unreadable material is labelled missing rather than inferred.
  • [ ] Each variance names the playbook rule, approved fallback, policy, or legal question used for comparison.
  • [ ] The review distinguishes a source fact, an AI suggestion, a reviewer judgment, and an unresolved question.
  • [ ] External law, regulation, or policy claims link to a current authoritative source and state jurisdiction and date.
  • [ ] No clause is marked acceptable merely because it resembles language seen elsewhere.
  • [ ] Fees, caps, credits, percentages, currencies, and formulas were recalculated independently.
  • [ ] Effective, renewal, notice, cure, payment, deletion, survival, and termination dates were placed on a timeline.
  • [ ] Every internal cross-reference exists and points to the intended obligation.
  • [ ] Defined terms are used consistently across the agreement and attachments.
  • [ ] Conflicting provisions have a recorded precedence analysis or escalation.
  • [ ] Each material finding explains the operational, financial, security, privacy, or legal consequence.
  • [ ] The proposed disposition is accept, negotiate, reject, clarify, or escalate—not simply “review.”
  • [ ] Proposed language is checked for unintended changes to scope, economics, remedies, and dependencies.
  • [ ] Confidentiality and privilege handling match the organization’s approved process.
  • [ ] An authorized person, not the AI system, owns interpretation and risk acceptance.
  • [ ] Accepted and rejected AI suggestions remain distinguishable in the record.
  • [ ] Each issue has an owner, due date, disposition, approver, and remaining uncertainty.
  • [ ] Deviations from the playbook include the approver and rationale.
  • [ ] The final recommendation links back to the issue register and verified source set.
  • [ ] Unresolved exceptions are visible to the decision maker and cannot silently disappear on export.

Sources: [2]

Copyable verification record

Use one row for every material finding and one summary row for the final decision.

| ID | Source anchor | Verified text or value | Comparison basis | Consequence | AI contribution | Human verification | Disposition | Owner / approver | Residual uncertainty | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | CR-01 | Agreement § / page | exact quotation or calculation | rule / fallback / question | concrete effect | extract / compare / draft | check performed and result | accept / negotiate / reject / escalate | named people | what remains unknown |

Attach the frozen source manifest, calculation sheet, date timeline, definition map, prompt or reproducible instruction summary, output identifier when policy permits, and approval record. A screenshot alone is weak evidence because it may omit the source version and surrounding language.

Worked test: renewal and deletion terms

This synthetic example shows how to use the checklist. A software agreement renews for one year unless notice is given 30 days before renewal. A data schedule says customer data is deleted “after termination” but states no deadline. The approved playbook seeks 60 days’ renewal notice and deletion within 30 days.

The AI output correctly extracts the 30-day renewal term but reports that data will be deleted within 30 days. TRACE fails that second finding at Text and source control because the contract contains no such deadline. The reviewer records “deadline absent,” not the playbook value, and checks all schedules for a conflicting term.

The model also proposes replacing the deletion clause with “delete promptly.” TRACE fails the proposal at Consequence and control because “promptly” removes the measurable playbook requirement. The issue register preserves the rejected suggestion, proposes a 30-day obligation plus confirmation, and assigns the privacy owner and legal reviewer.

The timeline test then reveals that the commercial team’s planned decision date falls after the existing notice deadline. The renewal issue is escalated immediately rather than buried in the final summary. A completed packet contains the verified clauses, the absence finding, the calendar calculation, proposed language, owners, approvals, and remaining jurisdiction-specific questions.

Independent spot check

Before release, a second reviewer should sample all critical findings and a risk-based selection of the rest. For each sample, ask them to locate the source without help, repeat the calculation or comparison, and state whether the recorded consequence follows from the evidence. Expand the sample when any critical error, invented source fact, wrong version, or unexplained contradiction appears.

A checklist pass means the review artifact is traceable and ready for its named professional decision. It does not mean the agreement is enforceable, commercially acceptable, complete, or suitable for another jurisdiction.

Hold conditions

Hold the review when the controlling version is disputed, required schedules are absent, confidential information cannot be handled under an approved process, OCR quality prevents reliable anchoring, the comparison playbook is missing, a material issue depends on unresolved jurisdiction-specific advice, calculations cannot be reproduced, or no authorized person can accept the risk.

Two Prune can use a contract-review task in a realistic AI-enabled work assessment and 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