Every vendor in this space says “defensible.” Very few say what it means.
Here’s the working definition we use, and the questions to ask any AI review platform, ours included.
Defensible doesn’t mean perfect
Courts have never required perfect review. Rule 26(g) of the Federal Rules of Civil Procedure asks for a reasonable inquiry, and Rule 26(b)(1) ties discovery to what’s proportional to the needs of the case.
Human review isn’t perfect either. Research on manual review going back more than a decade has found that trained reviewers often disagree with each other about the same documents.
When technology-assisted review (TAR) arrived, courts accepted it on that same basis. The Southern District of New York approved predictive coding in Da Silva Moore v. Publicis Groupe (2012). Three years later, in Rio Tinto v. Vale, the same court called it black-letter law that a producing party may use TAR if it chooses.
AI review gets measured the same way. The question isn’t whether the tool is perfect. It’s whether you can show your process was reasonable.
What you need to be able to show
A reason for every call
TAR gives you a relevance score. It can’t tell you why a specific document scored the way it did.
Agentic review writes a rationale for each decision and cites the exact text it relied on. If a partner, opposing counsel, or a judge asks why a document was withheld, the answer is already on the record.
A log of who did what
Every agent decision and every reviewer action should be logged with a timestamp and the person or agent responsible. When a call changes, you can see who changed it and when.
That log is what you’d export for a meet-and-confer, a discovery dispute, or a regulator.
Human sign-off where it matters
Agents do the first pass. Lawyers make the high-stakes decisions.
Privilege calls, close responsiveness questions, and anything going out the door should stop at a checkpoint for a person to confirm. The platform should make those stops required, not optional.
Quality checks you can describe
You should be able to explain how you measured the work. That usually means sampling the documents marked not responsive, estimating recall, and recording how disagreements between agents and reviewers were resolved.
A simple test: if you can’t describe your validation in one paragraph, it won’t hold up well in a dispute about it.
Privilege handling that fails safe
Privilege candidates should be surfaced for a lawyer to decide, never auto-withheld or auto-produced. Where you can, pair your process with a Federal Rule of Evidence 502(d) order so an inadvertent production doesn’t turn into a waiver fight.
Questions to ask any AI review platform
- Can I see the reason behind each decision, with a citation to the document?
- Is every agent and reviewer action logged, and can I export that log?
- Where are human checkpoints required, and can a user skip them?
- How is recall estimated, and can I run my own validation sample?
- How are privilege candidates handled?
- Is our data used to train models?
- Where does inference run, and who controls that environment?
If a vendor can’t answer these plainly, that tells you something.
How we built Discovery AI around this
Every agent call in Discovery AI comes with a written rationale cited to the source text. Every agent decision and reviewer action is logged. Agents stop at human checkpoints where the stakes are high, and your team signs off before anything is produced. The audit trail exports for opposing counsel or a regulator.
That’s what we mean by defensible by construction. The record exists because the workflow creates it as the work happens, not because someone pieced it together after a challenge.
This guide is general information, not legal advice. Your obligations depend on your jurisdiction, your court, and any ESI protocol in your matter.