AI privilege review in 2026 -- what it catches, what it misses

Updated 21 April 2026 | Independent reference | Not legal advice

Privilege review remains the single largest cost driver in modern eDiscovery. A 5 TB corporate matter may contain tens of thousands of documents requiring individual privilege assessment, and the consequences of error -- either inadvertent production or over-withholding -- are significant. AI is genuinely useful here, but the accuracy limitations and failure modes require attorney supervision and a 502(d) backstop.

The privilege review problem at volume

In a typical 5 TB corporate investigation, a document population of 500,000 documents might yield 30,000-80,000 documents flagged as potentially privileged in keyword-based first-pass culling. Manual attorney review at $200-350/hour, even at 50-80 documents per hour, can cost $75,000-160,000 in privilege review alone before any other review cost. Privilege logging requirements add further attorney time.

AI privilege detection replaces or augments the keyword-flagging step with a model that understands privilege signals contextually: attorney-client communications, work product preparation, legal advice requests, and the forwarded-chain propagation of privilege. The potential cost reduction is significant -- but only if the AI is accurate enough and the validation methodology is documented.

How AI privilege detection works

Modern AI privilege detection, as implemented in Relativity aiR for Privilege, EverlawAI's privilege module, and similar tools, uses a combination of signals. Structural signals: presence of attorney email addresses or names in the To/From/CC fields, header analysis for legal hold or counsel communications. Content signals: LLM analysis of the document text for legal advice requests, attorney-client discussion patterns, legal strategy language, and work product indicators. Propagation signals: forwarded chain analysis that flags a document as privileged if it contains a privileged communication embedded in a reply chain, even if the outer email is routine.

Relativity's aiR for Privilege is the most mature implementation as of April 2026, with published benchmarks citing 88-95% first-pass accuracy on attorney-client and work-product documents in typical corporate litigation populations. EverlawAI's privilege module reports similar figures. Accuracy varies significantly by document type.

Accuracy by document type

Rule 502(d) as the essential backstop

Federal Rule of Evidence 502(d) allows a court to order that inadvertent production of privileged or work-product material does not constitute a waiver in the pending proceeding or in any other federal or state proceeding. This is the essential backstop for any AI-assisted privilege review.

Judge Peck's model 502(d) order language (developed and refined through the Da Silva Moore, Biomet, and Rio Tinto matters) is the standard template: it covers inadvertent production, requires prompt notification and return of produced privileged documents, and applies non-waiver protections to all subsequent proceedings. Any AI-assisted privilege review should be covered by a 502(d) order before production begins. The Sedona Conference has published recommended 502(d) order language, and most courts in jurisdictions with frequent complex commercial litigation have a standing template available.

Hard cases: where AI privilege detection fails

• Business-advice versus legal-advice line. In-house counsel routinely provide both legal advice and business advice in the same communication. Attorney-client privilege applies to the legal advice but not the business advice, and the line is a question of fact for each document. Current AI models struggle with the mixed-purpose document because the privilege determination turns on purpose, not content.
• Joint defence and common-interest privilege. Communications between co-defendants or parties with a common legal interest may be privileged under the joint-defence doctrine. AI models are not consistently trained on joint-defence privilege signals and frequently miss this category.
• Crime-fraud exception. Where a party claims the crime-fraud exception to privilege, the court conducts an in camera review. AI cannot assess the crime-fraud exception because it requires a legal determination, not a document-classification determination.
• Partial redaction within a document. A document may contain both privileged and non-privileged content. AI tools that classify at the document level, not the text-segment level, will either over-withhold or under-redact on partial-privilege documents.