11/20/2026 | Press release | Distributed by Public on 11/20/2026 17:15
Relativity, a legal data intelligence company, recently released How Reliable is Your AI? A Legal Team's Guide to AI Evaluation, a new e-book offering a practical framework for testing and continuously evaluating legal AI.
For organizations seeking value from their AI investments, adoption alone is not enough; evaluation is essential. The e-book describes the "checkbox trap," in which organizations onboard AI without systematically checking whether it performs reliably at the scale required and whether it stays useful as tools, data and needs evolve.
The e-book argues that legal is uniquely positioned to close that gap. Its defensibility culture, shaped by e-discovery and technology-assisted review, established validation, sampling and statistical quality-control norms a decade before AI evaluation entered the mainstream vocabulary.
To give teams a shared vocabulary, the e-book distinguishes evaluation from validation and accuracy from capability and reliability. It also explains why broad benchmarks, while helpful, may not reflect a team's own data and tasks, which is why targeted evaluation of a specific use case falls to the team itself. Accountability for AI output remains with people, and human judgment must stay engaged in both the use and the evaluation of AI.
To help teams put this into practice, the e-book introduces the 5C Evaluation Framework for Legal AI, a five-stage approach designed to help legal teams scrutinize new tools, validate existing ones and build evaluation into the rhythm of their work. The five stages are Claim, Context, Calibrate, Conduct and Continue.
Drawing on insights from litigators, legal operations leaders, data scientists and service providers who are doing this work now, the e-book also offers questions to ask any AI vendor before adoption and addresses questions the industry has yet to settle, such as whether AI can be used to evaluate AI and the need for shared evaluation standards. It concludes that the evaluation work teams do today may itself become evidence later.
To learn more, download the full e-book. For a closer look at the difference between accurate, capable and reliable AI, read this recent post on The Relativity Blog.