Privilege review is the most expensive line item in most discovery budgets, which makes it the first place legal teams want AI leverage — and the place where a mistake costs the most. A single missed privileged document can waive protection, and under the wrong facts, waive it for the entire subject matter.
The good news: the legal standard for protecting yourself is process-based, and AI-assisted review can satisfy it — if the process is built for defensibility from the start.
The standard: reasonable steps, not perfection
FRE 502(b) protects against waiver from inadvertent disclosure when the holder took reasonable steps to prevent disclosure and promptly rectified the error. Courts applying it look at the process: what screening was done, how it was validated, and whether the response to discovery of an error was prompt.
Nothing in that standard requires human eyes on every page. What it requires is a process you can describe, defend, and document. That's the design target.
What a defensible AI-assisted process looks like
- Tiered output, not binary calls. The AI classifies documents as privileged, potentially privileged, or clear — and everything except "clear" gets attorney review. The AI narrows the haystack; it doesn't make the call.
- A rationale log per document. For every classification, the system records why — the indicia it found, the names it matched, the thread it traced. Six months later, this log is your privilege-log draft and your 502(b) narrative.
- Validation sampling on the "clear" tier. Attorneys review a statistically meaningful sample of documents the system cleared. Agreement rates get documented. This is the single strongest piece of evidence that your steps were reasonable.
- A clawback agreement anyway. A 502(d) order or clawback agreement costs nothing and caps the downside. AI-assisted or not, entering discovery without one is the actual unreasonable step.
The failure mode to design against
The risk isn't that AI screening misses documents — sampling catches systematic misses. The risk is an undocumented process: screening that happened somehow, with settings nobody recorded, validated by nobody in particular. That process fails the 502(b) analysis even if it made no errors, because you can't prove what it did.
This is why audit-trail quality should outweigh raw accuracy when evaluating tools. A skill that's right 97% of the time with a deterministic, per-document log beats one that's right 99% of the time as a black box.
Supervision protocol, one page
- AI performs first-pass classification with rationale logs on the full production set.
- Attorneys review all privileged and potentially-privileged classifications.
- Attorneys review a sampled subset of clear classifications; agreement rate is recorded.
- Disagreements route back as configuration corrections, and affected tiers are re-screened.
- The privilege log is drafted from the rationale logs and finalized by attorneys.
Write it down, date it, follow it. That single page is most of your defensibility.
Related reading
This post supports our complete guide: AI Agent Skills for Legal Teams: The Complete Evaluation Guide. The same evidence-first discipline runs through everything LibSkills ships — if you're deciding where a niche practice like this fits in your own market, start with the free Sub-Niche Opportunity Report.