eDiscovery document review where the agents explain every call, your reviewers decide, and the privilege log holds up on first serve
A privilege log that holds up on first serve, every entry tied to its passage.
See one case, screen by screen ↓eDiscovery document review software that reviews documents for relevance and privilege and drafts the privilege log, each call cited. Eight agents code the review set, check each other's privilege calls, write the log row by row as reviewers decide and watch productions for slips. Review attorneys make every privilege call, and the supervising attorney signs the log and every production.
How to choose AI solutions for back-office and enterprise work →
Updated
Every document that answers the other side's requests for production has to be reviewed, then produced, withheld or redacted. Each withheld or redacted document needs a privilege log entry that lets the other side assess the claim under Rule 26(b)(5)(A): date, author, every recipient and their affiliation, the privilege asserted and a description that does not give the advice away. Review is most of what eDiscovery costs, and the log is usually written at the end, in bulk, from tags and metadata.
That is where it breaks. A reviewer calls an email privileged because the general counsel is on it and misses the outside PR firm on the cc line. Descriptions come out as "email re legal advice". Since 1 December 2025 the timing and method of the log belong in the Rule 26(f) discovery plan and can be written into the scheduling order, so the log is measured against what the parties agreed. A log the other side cannot assess brings a motion to compel, a re-review and, where a third party was left off, an argument that the privilege was waived.
Estimated split for one rolling production volume on a mid-size matter, by hand and with the solution.
Eight agents code, check each other and write the log as reviewers decide; the supervising attorney signs.
Arden Components, Inc. v. Kestrel & Vane Industrial Ltd., in the District of Delaware. Kestrel & Vane is answering 24 requests for production from a review set of 186,240 documents across 38 custodians, and Volume 3 is due Friday 23 October. The agents read, cite and propose; the review attorney decides; the supervising attorney approves and signs. Here is how one email goes, screen by screen, in the working solution.
Today shows 45 of 46 batches closed and the eight release gates for Volume 3. Under "Needs you", Hannah has one item: batch 14 has a call waiting, because the checker flagged a possible waiver. Below, the agents' latest work is listed: the checker's flag, the batch planner's forecast, the QC sampler's elusion sample.
“Batch 14: 1 call waiting for you · KV-0031942 · the checker flags a possible waiver: outside PR firm on counsel's advice”
Batch 14 is a second-level privilege batch of 64 documents from the Okafor and Marchetti threads. KV-0031942, "RE: Arden termination — statement and next steps", is open in the middle with its passages numbered. The first-pass reviewer coded it responsive to RFP 7, 8 and 23 at 0.96 and privileged, attorney-client, at 0.91: "General Counsel gives advice in ¶2 of her 17 Mar message." Under the document, the batch's privilege log is being written as each call is made.
The header names every participant's role: Sofia Marchetti, attorney · Leah Brennan, third party.
The checker reads the thread: 3 messages, 3 people, 2 domains. Leah Brennan is not counsel. The engagement register shows Halden Brook Communications was engaged by Corporate Communications on 6 Nov 2023 for media relations, with no Kovel letter. Attorney-client privilege is likely waived; work product is tested separately, because Arden's notice of 11 March made litigation foreseeable and ¶2 is counsel's assessment of litigation risk. It recommends producing with one redaction, at 0.87, with each source cited.
“Hand to the review attorney — A privilege call is never final until a person makes it.”
The screen says it plainly: "Agents disagree: first pass withhold, checker produce with one redaction." Hannah chooses Overturn, then "Produce with one redaction", then reason code TP-1, "Third party on the communication — no Kovel engagement", and adds a note for the record. Her decision rewrites the log row and is logged with her name.
Paragraph 2 becomes a redaction box: "REDACTED — WORK PRODUCT · PRIV-02980 · Protocol §4.4". The row that read "Withheld · Attorney-client · Email re legal advice" now names Leah Brennan, Halden Brook Communications (outside PR consultant; third party), and its description passes ESI Order ¶11. The attachment, the holding-statement draft, follows its parent and is produced in full. Batch 14 is complete, and so is review for Volume 3: 46 of 46 batches.
“Email chain forwarding counsel's 17 Mar 2024 assessment of litigation risk following Arden's 11 Mar 2024 notice of breach, prepared in anticipation of litigation; shared with an outside communications consultant (Halden Brook Communications) for a holding statement.”
It is Hannah's second Halden Brook overturn in the matter, so the checker drafts protocol note PN-12: halden-brook.com is a third party, and work product is tested for counsel's litigation assessments after 11 Mar 2024. The preview finds 318 documents in the Volume 4 queue across 9 batches: 37 calls would change, 29 privileged to produce and 8 to redact as work product. The 64 decided Volume 3 documents from the same firm show 0 conflicts.
“Wait for the supervising attorney — Nothing changes until Elena Brandt approves.”
PN-12 waits on Quality with its numbers: 318 documents, 37 calls, 0 conflicts. Beside it is how the review is measured: estimated recall 95.7% from an elusion sample of 2,401, reviewer agreement 96.5%, and the agents' privilege calls overturned 11.2% of the time in batch 1 and 3.1% in batch 46 as approved notes reached the first pass. Jonah Weiss is at 93.4%, and his batch 9 can be re-checked with one click.
“Overturns fall as approved protocol notes reach the first-pass reviewer. Nothing is retrained in the dark: every change traces to a note.”
Elena runs the opposing-counsel read on log v3: 3,091 entries checked against Rule 26(b)(5)(A) and ESI Order ¶11, every participant checked for a role and affiliation, and descriptions compared with what courts have sent back. The result: 0 entries at high risk, 41 medium and 3,050 low. The medium ones use thread or attachment wording that the order allows, and a reviewer accepted each one.
“No entry at high risk. … The log is ready for sign-off.”
The sign-off shows what she is approving: 3,091 entries, 2,604 withheld and 487 redacted, all 3,091 re-read by the checker, 1,108 rewritten, and the method, ESI Order ¶11, document by document. She picks reason code AP-1, "Meets ESI Order ¶11 and Rule 26(b)(5)(A)", and types her name. Volume 3 then shows 7 of 8 release gates passed; her approval of the production opens only when every gate has passed.
“I have reviewed the log and the checker's read. I approve it for service with Volume 3.”
The clawback agent keeps comparing what was produced with every later privilege decision. On 20 Oct it found that KVI-0104377 to KVI-0104379, produced in Volume 2, are a 98.6% near-duplicate of a withheld memo. Under ¶3 of the Rule 502(d) order, notice is due Tuesday 27 Oct, 6 days away. The notice to opposing counsel is drafted for Elena to sign; the agent never contacts the other side itself.
“Notice due Tue 27 Oct · 6 days left · found Tue 20 Oct 16:40 by the clawback agent”
The same matter, seen by four of the people who carry it — what a production week looked like, and what it looks like now.
Eight agents code, check each other and write the log as reviewers decide; the supervising attorney signs.
Builds batches by custodian, thread and family, assigns them by load and skill, and back-plans review to the production date, proposing a re-balance when a batch falls behind.
Codes every document for responsiveness, request by request, and for privilege, highlights the passage behind each call and gives a confidence.
Re-reviews every privilege call and log line: tests waiver against the engagement register, tests work product separately, proposes the narrowest treatment, rewrites descriptions, runs the opposing-counsel read and drafts protocol notes from overturns.
Writes the log row at the moment of decision: date, author, every recipient with their affiliation, basis, treatment and a description, logging threads and attachments as the ESI order says.
Proposes each redaction box with its basis: privilege, work product, personal data under FRCP 5.2, or third-party pricing under the protective order.
Draws elusion and reviewer samples at the confidence in Settings, estimates recall, spots reviewer drift and re-queues a batch for a re-check.
Numbers, endorses and assembles each volume with load file, privilege log and cover letter, checks Bates continuity and families, and serves by secure link.
Watches produced volumes against later decisions and near-duplicates of withheld documents, starts the Rule 502(d) clock, finds every copy and drafts notices and sequestration certificates.
Signs the log and every production. The agents propose; a named person decides.
Ask in plain words about a control number, a Bates range, the log or the next production — or tell it what to draft. Every answer points to its source and opens the screen it is about.
Why was KV-0031942 flagged?
The first-pass agent coded it attorney-client privileged (0.91) because Sofia Marchetti gives advice in ¶2 of her 17 Mar message. The privilege checker then found Leah Brennan, Halden Brook Communications, on the cc line. Halden Brook was engaged by Corporate Communications on 6 Nov 2023, not through counsel, and has no Kovel letter (ENG-0112). So attorney-client privilege is likely waived, but ¶2 can still be protected as work product. Its recommendation: produce with one redaction, confidence 0.87. The call is yours.
Which log entries would the other side challenge?
One entry is at high risk: KV-0031942. Its description, "Email re legal advice", does not let Arden assess the claim under Rule 26(b)(5)(A). It also leaves out Leah Brennan of Halden Brook, a third party on the thread, which ESI Order ¶11(b) requires. 41 more entries are medium risk.
Draft the clawback notice for KVI-0104377
KVI-0104377 to KVI-0104379 were produced in Volume 2 on 25 Sep. They are a 98.6% near-duplicate of withheld KV-0019220, Wexford Lane's draft termination memo. Under Rule 502(d) Order ¶3, notice is due within 7 days of discovery (20 Oct), so by Tue 27 Oct. The notice asks Corbin & Ashe to sequester and certify within 5 days. Elena Brandt signs it.
How much is left before Volume 3?
One call. 45 of 46 batches are closed. Batch 14 has KV-0031942 waiting for Hannah Cole, plus its attachment, which follows it. After that, the open gates are 618 personal-data redactions to accept and Elena Brandt's sign-off of log v3. Volume 3 is due Fri 23 Oct.
13 screens from the working solution, on its sample data. Pick one to see it large.
The rules and thresholds the agents follow are settings for each matter. Changes to the review rules are logged and need the supervising attorney.
| Setting | Default | Choose from |
|---|---|---|
| Privilege candidates always go to a person | Always on | Cannot be turned off: an agent never makes a privilege call final |
| First-pass confidence for a person to review | 0.90 | Set per matter; below it, responsive calls go to a reviewer |
| Code not responsive automatically | On, at 0.95 or above | On · Off; these are seen only if they fall in a QC sample |
| Families follow the parent (FM-1) | On | On · Off; the checker can show a different basis for an attachment |
| Thread logging | Per ESI Order ¶11(c) | Per ESI Order ¶11(c) · Every message separately |
| High challenge risk | 0.70 | The score at which a log entry blocks sign-off |
| Near-duplicate threshold for the clawback watch | 95% | Produced pages compared with withheld documents |
| Reviewer QC sample · elusion sample | 2.5% · 95% confidence, ±2% | Share of each reviewer's decisions re-checked · confidence and margin |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample matter · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (ComplexDiscovery / EDRM 2025 pricing surveys — $0.50–$1.00 a document for traditional review, often more). Rules: FRCP 26(b)(5)(A); FRCP 16(b)(3)(B)(iv) and 26(f)(3)(D) as amended 1 Dec 2025; FRE 502(d). People, companies, firms, matters and documents named on this page are fictional — characters and sample data in the working solution.
eDiscovery document review software that reviews documents for relevance and privilege and drafts the privilege log, each call cited. Eight agents code the review set, check each other's privilege calls, write the log row by row as reviewers decide and watch productions for slips. Review attorneys make every privilege call, and the supervising attorney signs the log and every production.
The first-pass reviewer codes each document for responsiveness by request and for privilege, with the passage behind each call and a confidence. The privilege checker then re-reviews every privilege call. Neither makes a call final: a review attorney upholds, overturns or escalates each one with a reason code.
The privilege checker reads every participant against the attorney list and your engagement register: who counsel engaged, when, and under which letter — Kovel, common interest or ordinary business. Where a third party is present, it tests attorney-client privilege and work product separately and recommends the narrowest treatment, such as producing with one paragraph redacted.
Every entry is written at the moment of decision with the date, author, every recipient and their affiliation, the basis and a description, then re-read by the privilege checker against Rule 26(b)(5)(A) and the method in your ESI order. Before sign-off it reads the whole log the way opposing counsel will and scores each entry's challenge risk. The target is zero entries sent back after service; on the sample matter, the checker rewrote 1,108 of 3,091 descriptions before the log was served.
Recall is measured, not assumed. The QC sampler draws an elusion sample from what was coded not responsive — on the sample matter, 31 responsive in 2,401, for an estimated recall of 95.7% — and re-checks a share of every reviewer's decisions, so drift shows up while there is time to fix it.
After every volume, the clawback agent compares what was produced with later privilege decisions and with near-duplicates of withheld documents. When it finds a slip, it starts the clock in your Rule 502(d) order, lists every copy by Bates number and drafts the notice for counsel to sign. When the other side claws back, it finds and locks every copy, including deposition exhibit folders, and drafts the certification.
Yes. The agents recommend; review attorneys decide every privilege call with a reason code. A protocol note drawn from an overturn changes nothing until the supervising attorney approves it, and only the supervising attorney can sign the log, approve a production or send a clawback notice. The agents never contact opposing counsel.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your review protocol, orders, attorney list, engagement register and a sample of past logs — and it goes live once every quality gate has passed. We will walk you through it on a sample of your own review set first.
Every deadline met, and every page traced to the request it answers.
1-min demoPublic Records Request ManagementEvery withheld line cites its law, and every release goes out before the clock runs.
1-min demoLegal Intake & TriageRoutine questions answered in minutes; anything risky reaches a lawyer first.
1-min demoWe'll run eDiscovery Document Review & Privilege Log on a sample of your own review set, protocol and past privilege logs.