There is no responsible universal accuracy percentage for AI contract review. Performance can change with the document type, scan quality, length, jurisdiction, task definition, model, prompts, and evaluation set.
Define accuracy before measuring it
Different tasks require different tests:
| Task | Evidence of success |
|---|---|
| Clause extraction | The quoted text and page match the source. |
| Classification | The clause is assigned to the correct category. |
| Missing-term detection | The complete document was searched and the expected term is truly absent. |
| Risk explanation | The consequence follows from the text and labels assumptions. |
| Legal research | The authority is current, applicable, and supports the proposition. |
| Cross-document review | Definitions, amendments, exhibits, and conflicts are reconciled correctly. |
A fluent summary may be wrong on one of these tasks even when it reads well. For examples of what a first pass can and cannot surface, see what AI contract review can catch.
Build a useful test
Use a representative, labeled set that includes ordinary clauses, unusual drafting, scans, tables, cross-references, missing exhibits, and known non-issues. Have a qualified reviewer define expected findings before looking at the tool's output.
Record false negatives, false positives, unsupported legal conclusions, citation errors, and failures to label uncertainty. Report results for the tested version and dataset rather than converting them into a permanent product claim.
Check every live report
Even a strong evaluation does not make an individual output self-proving. For each matter:
- confirm the correct file and version were analyzed;
- verify material findings against quoted text;
- inspect connected definitions and exceptions;
- identify facts and documents outside the upload; and
- escalate legal judgment and high-consequence uncertainty.
The NIST AI Risk Management Framework provides a voluntary structure for managing AI risk. It does not certify contract-review accuracy or validate a provider's scoring system.
What an honest provider should disclose
Look for a defined test corpus, task labels, reviewer method, model or product version, error categories, and limitations. A claim such as “lawyer-grade,” “complete,” or “more accurate” is not reproducible without that context.
Inkvex provides a structured first pass for acquisition buyers, franchise buyers, and commercial tenants. Verify its quoted findings and use qualified counsel for the legal conclusions and transaction decision.
Inkvex provides legal information, not legal advice.
Design a reproducible accuracy test
Choose a labeled set that resembles the intended documents. Record expected clauses, cross-document issues, missing items, and outside facts before testing. Include clean digital files, scans, tables, amendments, and deliberately incomplete packages. Have a qualified reviewer establish the reference answers and document disagreements.
Measure issue-level recall, unsupported-finding rate, quote accuracy, source-location accuracy, and calibration. Report results by document type and file quality rather than collapsing everything into one percentage. Repeat runs when model variability matters and preserve the version, prompt, settings, and date.
Review each live result anyway
Even a tested system can fail on a new document. Verify the file count and extraction, compare every material quote with the source, follow definitions and exceptions, and label facts the contract cannot prove. Accuracy in locating a clause is different from accuracy in explaining legal effect. The latter may require governing law, transaction facts, and counsel. Until a provider publishes a relevant labeled study, treat broad accuracy claims as marketing rather than evidence. For when to rely on a report and when to call counsel, see when to trust AI contract review.
Go deeper
Read the guide, then move into the real workflow, pricing, audience page, and glossary that support the next decision.
Got a contract to review?
Upload it for a first-pass contract review. The normal processing target is under 3 minutes. Large, scanned, or unusually dense files can take longer. Free Trial available.
Related Articles
This article is for informational purposes only and does not constitute legal advice. For high-stakes agreements, consult a qualified attorney.