AI privilege review uses AI to flag documents that may be privileged and to draft privilege log descriptions, while attorneys make every final privilege call. Done well, it speeds triage and makes log entries more consistent. Done badly, it risks waiver or an inadequate log.
The safe pattern: define scope and a role map, let AI triage, route borderline documents to attorneys, quality-check both "privileged" and "not privileged" results, and keep records.
QC note: Requires attorney review before publishing. This article is general information, not legal advice, and does not guarantee compliance with any rule, order or ethics obligation.
What the Rules Require
Privilege review sits inside rules that AI does not change:
• Describing withheld material: Federal Rule of Civil Procedure 26(b)(5)(A) requires a party withholding documents to expressly claim privilege and describe them in a way that, without revealing privileged information, lets other parties assess the claim
• Inadvertent disclosure: Federal Rule of Evidence 502(b) addresses when disclosure does not waive privilege, including whether reasonable steps were taken to prevent and fix it
• Court orders: Rule 502(d) lets a federal court order that disclosure in the litigation does not waive privilege
• Lawyer duties: ABA Formal Opinion 512 (2024) discusses competence, confidentiality, supervision and fees when lawyers use generative AI
Local rules, ESI protocols and state rules may add requirements. Check them for each matter.
The Seven-Stage Workflow
Stage 1: Define scope. Record the privileges at issue (attorney-client, work product), governing law, date range and the log format agreed or ordered.
Stage 2: Build a role map. List in-house and outside counsel, their email addresses and domains, plus agents such as consultants whose involvement may matter. AI tools rely on this to spot legal communications.
Stage 3: Run AI triage. Let the tool score documents as likely privileged, likely not privileged or uncertain, with reasons and cited passages.
Stage 4: Route to attorneys. Send every likely privileged and uncertain document to attorney review. The tool's output is a suggestion, never the call.
Stage 5: QC both sides. Sample documents the AI marked not privileged, because a miss there leads to disclosure. Sample attorney calls too.
Stage 6: Draft and review the log. Let AI draft descriptions from approved fields, then have attorneys edit each entry so it describes without revealing.
Stage 7: Record and protect. Keep tool versions, settings, samples and reviewer decisions, and seek a Rule 502(d) order where appropriate.
Where AI Helps Most
• Spotting counsel: finding lawyer names and domains across long email chains
• Consistency: applying the same reasoning to near-duplicates and families
• Log drafting: producing first-draft descriptions in a consistent format
• Prioritization: putting the hardest documents in front of senior reviewers first
Relativity's documentation, for example, describes its Privilege product as predicting whether documents may be privileged and drafting log descriptions, using generative AI, machine learning and social network graphs.
Where AI Struggles
• Business versus legal advice: an email copying a lawyer is not automatically privileged
• Third parties: consultants or vendors in a thread can affect privilege
• Partial privilege: documents needing redaction instead of withholding
• Family and thread boundaries: attachments and earlier messages may differ in status
• Jurisdiction: privilege law varies, and the tool may not know which law applies
Worked Example: One Log Entry
Illustrative example. An email thread between a product manager and in-house counsel about a draft supplier contract.
• AI triage: likely privileged; counsel identified from role map; legal advice language cited
• Attorney call: privileged, attorney-client; one attachment is a supplier's form contract and is produced
• AI draft description: "Email from in-house counsel to product manager providing legal advice regarding draft supplier contract terms."
• Attorney edit: adds date, author, recipients and privilege basis fields per the agreed log format; confirms the description reveals no advice content
The final entry fields, such as control number, date, author, recipients, document type and privilege basis, should follow your matter's protocol.
QC Sampling Plan
Illustrative example. On a set of 40,000 documents:
• AI not privileged: random sample of 500 reviewed by attorneys to estimate misses
• AI privileged, attorney confirmed: sample 200 for second-level review
• Uncertain: 100% attorney review
• Log entries: senior attorney reviews a sample before service
Sample sizes and acceptance thresholds are illustrative; the supervising attorney sets them based on the matter's risk.
Vendor Questions to Ask
• How does the tool use our role map, and can we update it during review?
• Does any client data train models, now or later?
• Where is data stored and for how long?
• Can we export AI reasons, scores and reviewer decisions for each document?
• How are thread and family relationships handled?
For related reading, see our guide on how law firms supervise AI legal review.
Roles and Responsibilities
• Supervising attorney: sets scope, sampling thresholds and signs off on the log
• Review attorneys: make privilege calls on routed documents and edit log entries
• Litigation support: configures the tool, records versions and settings, runs exports
• Case team: maintains the role map as counsel and consultants change
ABA Formal Opinion 512 notes that managerial lawyers should set clear policies on permitted AI use and supervisors should make sure lawyers and staff are trained on the tools they use.
Pre-Production Checklist
• Scope, governing law and log format recorded
• Role map built and approved by the case team
• AI settings and version recorded
• All likely privileged and uncertain documents reviewed by attorneys
• Not-privileged set sampled and results recorded
• Log entries edited and approved by an attorney
• Rule 502(d) order considered
• Clawback procedure ready
Common Mistakes to Avoid
• Letting AI make the call: privilege decisions belong to attorneys
• Skipping the not-privileged sample: that is where inadvertent disclosure comes from
• Revealing advice in descriptions: a log entry must describe, not disclose
• Stale role maps: new counsel or consultants change results
• No records: without logs of the process, defending reasonableness is harder
Limitations
AI tools differ in method and accuracy, and vendor documentation describes features rather than proving results. Privilege law varies by jurisdiction. Requires attorney review before publishing.
Next Step
If your team wants privilege-aware review with citations preserved for each recommendation, see Lexintis legal AI by LuMay.
• Federal Rule of Evidence 502 (Cornell LII)
