Batch privilege review with Claude

A Claude Skill that takes a batch of documents from an eDiscovery review platform and runs first-pass privilege review — flagging each document as privileged, not-privileged, or borderline-needs-attorney, plus drafting the privilege log entry for each privileged document. Designed to plug into an existing review workflow rather than replace the review platform.

What you’ll need

• Claude Code or Claude.ai with custom Skills enabled and Tier-A authorization for privileged content per AI policy
• Document export from your review platform (Relativity, Everlaw, DISCO) as a folder of files plus metadata .csv
• Privilege determination criteria documented for the matter (which attorneys, which subject matters, jurisdiction-specific rules)
• Sample of attorney-pre-tagged documents for calibration (50-100 documents recommended)

Setup

1. Drop the Skill. Place privilege-review-batch.skill into your Claude Code skills directory. The Skill exposes two callable functions: classify_batch and draft_log_entries.
2. Configure matter parameters. Edit matter_config.yaml with: list of attorney custodians (in-house and outside), the matter’s relevant subject-matter scope, applicable jurisdictional rules, the privilege standard (attorney-client only, attorney-client + work product, etc.).
3. Calibrate against attorney-tagged sample. Run the Skill on the 50-100 sample documents; compare AI tags to attorney tags; adjust matter config if the Skill systematically over- or under-flags.
4. Run on the full batch. Process the document export folder; produce per-document classification + draft log entries for privileged documents.

How it works

For each document, the Skill:

1. Extracts metadata. From the platform export — author, recipients, date, document type, custodian, prior review tags.
2. Reads the document content. Plain-text extraction from .docx, .pdf, .eml, .txt.
3. Applies privilege criteria. Tests against the matter’s privilege determination rules: was the communication between attorney and client; was the purpose legal advice; does the work-product doctrine apply; are any waiver indicators present.
4. Outputs classification + log entry. Returns privileged / not-privileged / borderline plus a confidence score; for privileged, drafts a privilege log entry following the matter’s log format.

Output

• Classification CSV — one row per document with: document ID, classification, confidence score, basis, suggested log entry text
• Borderline queue — separate CSV of documents flagged for attorney review with the specific concern noted (attorney involvement unclear, subject matter borderline, potential waiver indicator)
• Draft privilege log — formatted log entries for all privileged classifications, ready for attorney review and finalization

Where it fits

This Skill is the first pass in a Stage 6 EDRM review workflow. The expected workflow:

1. AI runs first-pass on the full review universe → classifies 80% with high confidence, flags 20% borderline
2. Attorneys review borderline documents and a sample of high-confidence classifications (10-20%) for validation
3. AI-generated log entries are reviewed and edited by attorneys before production
4. Final attorney sign-off on every withheld document and every log entry before production goes out

AI is the speed multiplier; attorneys remain accountable for every privilege call.

Watch-outs

• Borderline rate matters. A well-tuned config produces 10-20% borderline rate. Higher means the Skill is hedging; lower means over-confidence (validate by sampling high-confidence calls).
• Sample-validate every batch. Even a well-tuned Skill makes errors. Sample 5-10% of high-confidence classifications per batch for attorney validation.
• Court-acceptable workflow. AI-assisted privilege review is increasingly accepted by courts but disclosure norms vary. Verify the production-side disclosure obligations for the matter.
• Privilege log finalization is attorney-only. The Skill drafts log entries; attorneys finalize. Producing AI-drafted log entries without attorney review is professional malpractice.
• Tier-A AI authorization required. Privilege content cannot be processed by non-enterprise AI tiers. Confirm vendor data terms before use.