# What are the best practices for using AI in eDiscovery?

> Write your criteria down before you run anything, validate the output against a human-reviewed sample, keep a reviewable record of why each document was coded the way it was, and keep a lawyer accountable for the final call.

AI in eDiscovery fails in predictable ways, and the practices that prevent it are mostly procedural rather than technical.

**Start with a written criterion.** Most disappointing results come from a vague instruction, not a weak model. State the question the way you would state it to a contract reviewer: what makes a document responsive, what does not, and what the edge cases look like.

**Validate before you rely.** Run the criterion on a sample, have a human review the same sample independently, and compare. Disagreements tell you whether the problem is the instruction or the population.

**Measure both directions.** Track what the AI missed and what it over-called. A tool tuned to catch everything will bury reviewers in false positives; one tuned for precision will quietly drop documents.

**Insist on reasons, not just labels.** A coding decision you cannot explain is a coding decision you cannot defend. The reasoning behind each call should be recoverable months later.

**Decide disclosure early.** Whether and when to tell the other side how the review was run is a strategy question best settled at the discovery-plan stage, not after a challenge.

**Keep a lawyer accountable.** Competence guidance is consistent on this point: the tool assists, the lawyer remains responsible for the work product.

**Protect the sensitive material.** Confirm where data is processed, who can access it, and whether anything is retained for training.

Related reading: [How AI fits into modern eDiscovery: a practical guide](/stories/how-ai-fits-into-modern-ediscovery-a-practical-guide).

If you want to see what this looks like in a working review, [book a Claira demo](/demos).