Privilege Review

Privilege review is the pass over a responsive document set that identifies material protected by attorney-client privilege or the work-product doctrine, so it can be withheld from production and recorded on a privilege log instead. It sits inside the broader eDiscovery workflow — after responsiveness review, before production. Producing a privileged document by mistake is one of the most expensive errors an eDiscovery program can make: depending on jurisdiction and the orders in place, it can waive privilege over an entire subject matter rather than just the one document that went out.

It is not responsiveness review, and it is not redaction. Responsiveness review asks whether a document falls inside the scope of the request; privilege review asks whether a responsive document can be withheld anyway. Redaction is what happens to the partly-privileged subset — the document is produced with the protected passages masked instead of withheld whole. It is also not a confidentiality screen: trade secrets, personal data, and anything stamped “confidential” are handled by protective order and redaction, not by privilege.

What’s actually privileged

Two doctrines, routinely conflated:

• Attorney-client privilege. Protects confidential communications between attorney and client made for the purpose of obtaining or giving legal advice. Runs both directions. Held by the client, and waivable by the client.
• Work-product doctrine. Protects material prepared by or for an attorney in anticipation of litigation — mental impressions, strategy notes, witness interviews. Broader in what it covers, weaker in what it grants: work product is a qualified protection that can be overcome by a showing of substantial need, while privilege is absolute until waived.

Three misconceptions that drive most of the disputes:

• In-house counsel communications are privileged only when the purpose is legal advice. The same lawyer writing about litigation exposure may be privileged; writing about pricing strategy usually is not.
• Adding an attorney to the cc line does not create privilege. The communication has to seek or convey legal advice, not merely inform a lawyer that something happened.
• Waiver is often inadvertent — sharing with a third party, producing the same document in another matter, or putting the advice at issue as both sword and shield.

Where it sits in the matter

1. Responsiveness review identifies documents within the scope of the request
2. Privilege review flags the responsive documents that will be withheld
3. Privilege log generation records each withheld item — date, author, recipients, document type, subject matter, basis — without disclosing the protected substance
4. Production delivers the responsive, non-privileged set to opposing counsel
5. Log delivery accompanies the production and lists what was held back

What changed on 1 December 2025

Amendments to the Federal Rules of Civil Procedure took effect on 1 December 2025 and moved privilege logging to the front of the case. Rule 26(f) now requires the parties’ discovery plan to state their views and proposals on how and when privilege claims will be made, and Rule 16(b)(3)(B)(iv) lets the court write the timing and method for complying with Rule 26(b)(5)(A) into the scheduling order. Because the Rule 26(f) conference happens no later than 21 days before the scheduling conference, the log format is now negotiated in the first weeks of the case rather than fought over after review is finished.

What is on the table at that conference: whether documents may be logged categorically instead of one row per document, whether metadata fields can stand in for narrative descriptions and which fields count, and whether whole categories — post-complaint communications with outside counsel, for instance — can be excluded from logging entirely.

The operational consequence is that privilege review design became a day-one decision. If the plan is to run an AI privilege screen and auto-draft log descriptions, that is a position to take at the Rule 26(f) conference, because it determines which metadata has to survive processing and what description granularity the team has committed to deliver. Agreeing to narrative descriptions and then discovering the tooling produces metadata-shaped output is an expensive order to unwind. See privilege log format for the field-level shape of each option.

What it costs

The Winter 2026 eDiscovery Pricing Survey run by ComplexDiscovery with EDRM (53 respondents, late December 2025 through 21 February 2026) puts per-document managed review at $0.50-$1.00 for 30.2% of respondents onsite, with 22.6% reporting above $1.00; remote review runs slightly cheaper, with 13.2% below $0.50. Roughly a third of respondents answered “do not know” on the per-document questions, so treat the bands as directional rather than as a quote.

Privilege candidates are typically an estimated 5-15% of the responsive set. On a 1M-document responsive population that is 50,000-150,000 documents through privilege review, or roughly $25,000-$150,000 at survey rates — before the log. The log is usually the larger number, because each entry needs a document-specific subject-matter description, and that is attorney work rather than reviewer work.

How AI changes privilege review

Privilege is a better fit for LLM assistance than responsiveness, because the question has firmer edges: who sent it, who received it, was legal advice sought or given. The tooling has split into three shapes.

• A dedicated privilege product. Relativity aiR for Privilege returns a prediction per document plus rationales, citations, and considerations, and drafts the privilege log description. Setup is the load-bearing part: Known Attorneys and Known Law Firms lists, and a role configuration that classifies each organization and individual as privilege-conferring, privilege-breaking, or neutral. Relativity publishes a Fortune 100 telecom case study reporting 99% recall, 91% precision, and review 80% faster than manual. On cost, note that aiR became standard in RelativityOne in October 2025 but is metered in aiR Units at one unit per document — included is not unlimited, and the allowance and overage rate are quote-only.
• General coding AI pointed at privilege criteria. Everlaw has no separate privilege model; its AI Assistant Coding Suggestions run in batch against the privilege coding criteria you define. DISCO’s Cecilia Auto Review is a first-pass tagging engine that DISCO benchmarks at 3,800 documents per hour with precision and recall 10-20% higher than typical human reviewers — privilege is a tag you define there, not a purpose-built classifier.
• Assistants that assess privilege but do not run productions. CoCounsel and Claude will analyze a document set for privilege, but neither does Bates numbering, production sets, or log delivery. They fit a focused set of a few thousand documents, not a matter-scale review.

One re-platforming note: Casetext’s CARA is gone. Thomson Reuters retired Casetext on 1 April 2025 and the technology now ships as CoCounsel inside Westlaw, so any privilege workflow still written around CARA needs a new home.

The final call stays with a lawyer in every one of these. ABA Formal Opinion 512, issued 29 July 2024, holds the lawyer responsible for competence, supervision, and candor when generative AI is in the loop, and Relativity’s own documentation requires human confirmation of every prediction before production release.

Common pitfalls

• Treating every attorney communication as privileged. Over-claiming invites a motion to compel and adverse inferences. Guard: classify custodians by role before the screen runs, and sample-audit the flagged set for business-advice content.
• Boilerplate log descriptions. “Re: legal advice” repeated across thousands of entries draws an automatic challenge. Guard: require a named subject matter in every entry, and reject entries that are identical across more than a handful of documents.
• Forgetting work product. Material prepared by non-attorneys at attorney direction can be work product even when it is not privileged. Guard: log the two bases as separate fields so an entry can carry one, the other, or both.
• Inadvertent production. Guard: get an FRE 502(d) order entered before the first production, not after the problem surfaces — it allows recovery without subject-matter waiver.
• Negotiating log format after review is done. Under the December 2025 rules this is now a missed deadline, not just a bad sequence. Guard: bring a proposed format, and the tooling that produces it, to the Rule 26(f) conference.
• AI flags shipped without attorney sign-off. Guard: attorney QC on 100% of documents the model called privileged plus a random sample of the not-privileged set, with the sample rate recorded in the matter file.
• Assuming bundled AI is free. Guard: get the per-document unit allowance and the overage rate in the contract before scoping a matter around AI privilege review.

Related

• eDiscovery — the broader process privilege review fits inside
• Privilege log format — the deliverable that records withheld documents
• Redaction workflows — the adjacent process for partly-privileged documents
• Relativity — the platform with a purpose-built privilege model