AI privilege review is the use of artificial intelligence tools to identify, within a document population collected for litigation or investigation, those documents that are potentially protected from disclosure by attorney-client privilege, the work product doctrine, or other applicable privileges. In the eDiscovery context, privilege review is a mandatory step before production — a party cannot produce privileged documents to opposing counsel, and inadvertent production can in some circumstances constitute a waiver of the privilege.
The privilege review challenge is a volume problem. Modern eDiscovery document collections routinely involve hundreds of thousands to millions of documents. Reviewing every document individually for potential privilege — the traditional approach — is time-consuming and expensive. AI privilege review tools address this by applying automated screening to the full document population, identifying documents with indicators of potential privilege for focused attorney review, and allowing faster batch decisions on clearly non-privileged documents.
The essential and non-negotiable limitation of AI privilege review is that the AI identifies documents for human review — it does not make the privilege determination. Whether a specific document is legally protected by attorney-client privilege or the work product doctrine is a legal judgment that requires attorney analysis. AI is the screening filter; the attorney is the decision-maker.
Privilege review in complex litigation is one of the most error-prone steps in the eDiscovery process. Two types of errors are possible, each with different consequences:
Under-designation (privilege waiver risk). If AI or human reviewers fail to identify a privileged document, it may be produced to opposing counsel. Under the common law majority rule and FRE 502(b), inadvertent production of a privileged document may or may not constitute waiver depending on whether reasonable steps were taken to prevent disclosure and to correct the error promptly. Even with FRE 502(b) protection in federal court, privilege waiver litigation is expensive and uncertain. Claw-back procedures impose burdens on both parties. The risk is real: complex privilege analysis across millions of documents creates significant opportunity for error.
Over-designation (privilege log burden and challenge risk). If reviewers over-designate documents as privileged — withholding documents that are not actually privileged — the opposing party may challenge the privilege designations. Courts may order in camera review of withheld documents, and improper withholding can result in sanctions. An inflated privilege log also creates credibility problems with the court.
AI privilege review can reduce both error types when properly implemented — reducing under-designation by systematically applying consistent privilege indicators across the full document population, and reducing over-designation by calibrating the AI model to realistic privilege thresholds rather than flagging all attorney-adjacent communications as privileged.
How It Works
AI privilege review tools use a combination of approaches to identify potentially privileged documents:
Attorney and law firm name matching. The reviewing party provides a list of attorneys who were involved in the matter — in-house counsel, outside counsel, and their assistants. The AI identifies all documents that include any of these names in the communication participants, author fields, or document body. This is the most basic privilege indicator, but it systematically over-designates — not every communication involving an attorney is privileged.
Law firm domain recognition. Email domains associated with law firms (partner at smithjones.com, attorney at biglaw.com) are identified as privilege indicators. Combined with attorney name matching, domain recognition catches communications with outside counsel even when attorney names are not in the party's provided list.
Semantic privilege analysis. More sophisticated AI tools apply NLP-based semantic analysis to document content, identifying language patterns associated with legal advice communications: "legal advice," "attorney-client communication," "confidential communication," "pursuant to attorney-client privilege," and related semantic clusters. This approach catches potentially privileged communications that do not involve names or domains in the provided attorney list.
In-house counsel communications. Attorney-client privilege extends to communications with in-house counsel seeking legal advice — but not to in-house counsel communications that are primarily business advice rather than legal advice. AI tools trained on this distinction can flag in-house counsel communications for careful attorney review, recognizing that the legal vs. business advice distinction is particularly contested in in-house privilege disputes.
Work product identification. Work product protection covers documents prepared in anticipation of litigation by or for an attorney. AI tools identify documents with creation dates after litigation became reasonably anticipated, authored by or for legal counsel, with subject matter related to the disputed claims — flagging these for work product review.
Key Considerations for Law Firms
Customize the attorney list before running privilege review. The accuracy of AI privilege review depends critically on the completeness of the attorney list provided to the system. Include all in-house counsel, all outside counsel who worked on the matter, and all relevant legal assistants and paralegals who worked under attorney supervision. An incomplete attorney list will systematically miss privilege indicators for communications with attorneys whose names are not in the system.
Validate through statistical sampling. Before making production decisions based on AI privilege review, validate the AI's performance through statistical sampling: pull a random sample of AI-designated non-privileged documents and review them for missed privilege indicators. This "elusion testing" for privilege identifies systematic misses in the AI's screening approach and allows recalibration before production.
Attorney review of all designated documents is mandatory. Every document the AI flags as potentially privileged requires attorney review before it is withheld from production and logged in the privilege log. The attorney review must assess: Is this communication actually privileged? Was it made in confidence? Was it for the purpose of seeking or providing legal advice? Has the privilege been waived? AI flags; attorneys decide.
Privilege log accuracy is enforceable. The privilege log — the list of withheld documents with descriptions sufficient to assess the privilege claim — must accurately reflect the actual privilege basis for each withheld document. Courts enforce privilege log accuracy; inadequate logs can result in court-ordered in camera review and potential privilege waiver orders.
FRE 502 orders and protective agreements. In federal court, parties should enter into FRE 502(d) orders at the beginning of the case to protect against inadvertent production. These orders, approved by the court, provide that production of privileged documents in the case does not constitute waiver in the case or in any other proceeding. A 502(d) order provides the maximum available protection and should be standard practice in any matter with large-scale AI-assisted review.
Limitations and Risks
AI cannot assess the legal-advice vs. business-advice distinction. The most contested privilege question in complex litigation — particularly involving in-house counsel — is whether a specific communication was primarily for legal advice (privileged) or primarily for business advice (not privileged). This distinction requires legal analysis of the communication's content and context. AI tools cannot reliably make this distinction; every in-house counsel communication flagged by AI requires careful attorney analysis.
Foreign privilege rules vary significantly. Attorney-client privilege and its equivalents vary dramatically across jurisdictions. UK legal professional privilege differs from US attorney-client privilege. German Berufsgeheimnis differs from French professional secrecy. Cross-border discovery involving non-US documents requires jurisdiction-specific privilege analysis that AI tools trained primarily on US privilege doctrine cannot reliably provide.
Waiver analysis is entirely attorney work. AI privilege review identifies potentially privileged documents, but privilege waiver analysis — whether privilege has been waived by prior disclosure, subject matter waiver, crime-fraud exception, or other doctrine — requires attorney legal analysis that is entirely outside the scope of current AI privilege review tools.
Encryption and unreadable documents. AI privilege review depends on reading document content. Encrypted documents, corrupted files, and documents in proprietary formats that cannot be processed by the platform will not receive AI privilege screening and require separate handling.