10,000-doc review reaches the partner this week.
Attorneys still confirm every privilege call. The system does the surfacing at scale so they only see the documents that need a judgment.
A matter lands as 47,000 documents.
Custodian mailboxes, file shares, and collected exports stream into the matter workspace. No paralegal opens a Bates-numbered folder.
- batch 01 [Custodian A] mailbox 12,840 docs · 4.1 GB Loaded
- batch 02 [Custodian B] mailbox 9,217 docs · 3.2 GB Loaded
- batch 03 Shared drive export Contracts, board decks, memos Loading
Sonnet handles the parse pass across all 47,000 docs. Opus is reserved for the judgment calls. Running Opus on every document would blow the unit economics on a matter this size. The model split is what makes this viable at volume, and it is the only reason a tens-of-thousands-document review can clear in days instead of weeks.
Every doc becomes structured metadata.
PDFs, scans, email chains, attachments. Sender, recipients, dates, threading, and full text extracted to a per-doc record the rest of the pipeline can reason over.
RE: Q3 forecast (draft)
Body excerpt
"Looping in [Outside counsel] for guidance on the disclosure language before we circulate to the board. The revenue assumptions in tab 3 are the piece I want their read on..."
Thread
Parent of DOC-018475, DOC-018482
{ "doc_id": "DOC-018472", "doc_type": "email_with_attachments", "sent_at": "2023-08-14T16:42:00Z", "from": "custodian_a", "to": ["custodian_c", "outside_counsel"], "thread_id": "TH-04918", "attachments": 14, "ocr_confidence": 0.97, "counsel_addressee": true, "requires_judgment": true }
Parsing is high-volume, low-judgment. Sonnet runs on all 47,000 docs because the work is extraction, not interpretation. Anything that looks like it might involve counsel or attorney work-product gets a flag (requires_judgment: true) and waits for Opus in the next stage. That gate is what keeps the spend rational.
Privilege candidates surfaced for attorney confirmation.
The judgment layer. Opus reviews every flagged document against the firm's prior privilege calls on similar matters. It does not mark anything "privileged." It surfaces candidates for an attorney to confirm. The log is attorney-confirmed or it does not exist.
DOC-018472 · email to outside counsel
Counsel on To-line, request for legal advice on disclosure language. Pattern matches 14 prior privilege calls on similar disclosure-review threads.
DOC-021908 · in-house counsel memo
Authored by [In-house counsel] for [Custodian D]. Work-product pattern.
DOC-034115 · ambiguous · third party on thread
Counsel on thread, but a third-party vendor is also addressed. Possible waiver. Flagged for attorney review, not auto-classified either way.
DOC-041203 · vendor invoice
No counsel involvement, no legal-advice content. Not a privilege candidate.
Every doc gets tagged against your matter taxonomy.
Sonnet handles the straightforward tags. When a document sits between two issues or hits a sensitive boundary, Opus is called in to resolve. Tags map to the issues the matter team defined at kickoff.
DOC-018475 · thread reply
Matter #MT-2841 · taxonomy v3
Attorneys confirm. The log writes itself.
Each candidate lands in a queue with the system's reasoning, the prior-call pattern it matched, and the surrounding thread. The reviewing attorney confirms, downgrades, or asks a question. The log entry only exists once an attorney has signed off. Every decision is logged to the reviewing attorney.
Matter #MT-2841 · privilege log
1,184 candidates · attorney-confirmed entries only
Confirmation queue (sample)
- DOC-018472 Pending
- DOC-021908 Pending
- DOC-034115 Pending
- DOC-019003 Confirmed
Every confirmation, downgrade, and edit is timestamped to the reviewing attorney. The exported privilege log is a record of attorney decisions, never a model output.
A one-pager, plus the 34 docs that matter.
Of 47,000 documents, the partner gets a single page on what the record actually shows, plus links to the documents that drive it. Everything else stays available, but nobody has to read it to be briefed.
Matter #MT-2841 · partner brief
Prepared from 47,000 docs · 34 hot-doc citations
What the record shows
- [Custodian A] looped in counsel on disclosure language two weeks before the board meeting (DOC-018472, DOC-018475).
- Internal control owner raised the revenue-recognition issue in writing on 2023-07-22 (DOC-016401), 23 days before disclosure was finalized.
- Forecast assumptions in tab 3 were edited by three custodians in a 48-hour window (DOC-019200 through DOC-019218).
- 1,184 privilege candidates surfaced; 1,096 confirmed by reviewing attorney, 62 downgraded, 26 still in queue.
Hot documents (top 6 of 34)
- DOC-016401 Control owner raises revenue recognition concern in writing · pre-disclosure Hot
- DOC-018472 [Custodian A] requests counsel guidance on disclosure language Hot
- DOC-019215 Tab 3 forecast revision · assumption swap, no documented rationale Hot
- DOC-022104 [Custodian C] acknowledges the open question to a peer the morning of the board meeting Hot
- DOC-027880 Post-meeting summary references the unresolved item, then drops it Hot
- DOC-031044 Investor-relations Q&A prep avoids the same item by name Hot
Questions for the partner
- Do we treat the control owner's 2023-07-22 memo as the start of the relevant period for [Plaintiff]'s theory?
- How aggressive do we want to be on the tab 3 revision thread in deposition prep?
- Two of the 26 still-open privilege candidates may be the strongest documents for the other side. Do we want a second attorney pass before producing?
Same attorneys. The doc pile disappears.
Attorneys still confirm every privilege call. They just stop being the bottleneck on getting to the documents that matter.
Want one for your firm?
Record a 5-minute voice memo about a typical matter. We'll show up to our call with a system already designed for it.
Tell us about your practice Backed by our 6-week guarantee