eDiscovery Document Relevance Tagging
A litigation matter can pull tens or hundreds of thousands of documents into the review pool, and every one of them needs a relevance call — responsive or not, privileged or not — before production. Reviewing that volume manually at the billing rates associates or contract reviewers command is one of the largest line items in litigation cost, and under a production deadline, review teams face real pressure to move faster than careful reading allows, which is exactly the condition under which a privileged document slips into a production set or a genuinely responsive document gets tagged as irrelevant and never surfaces.
STARTING PRICE
From €799
Complex tier · Multi-system orchestration, custom logic, and higher-volume or higher-risk processing.
Get a quote →Saves roughly Reduces review team hours by an estimated 40-60% on the first-pass relevance sweep for large document populations.
How the automation works
We build a first-pass relevance and privilege scoring layer that reads the full document population against the matter's defined issue codes and privilege criteria, ranking documents by likely responsiveness so reviewers start with the highest-value material instead of working the pile in arbitrary order. Documents scoring as likely privileged are flagged and routed to a distinct privilege-review queue rather than the general responsiveness queue, since a privilege call carries different consequences than a relevance call. Every tag the model assigns is a suggestion with a confidence score attached — a human reviewer confirms or overrides every document before it moves to production, and nothing is produced or logged to a privilege log without that sign-off.
Process flow
- 01
Document population ingested trigger
The full document set for the matter is ingested and normalized, with metadata (custodian, date, file type) captured for review workflow routing.
- 02
Score relevance against issue codes ai
Each document is scored for likely responsiveness against the matter's defined issue codes, with the scoring model surfacing its confidence rather than a binary in/out call.
- 03
Flag likely privileged documents ai
Documents matching privilege indicators — attorney correspondence, legal advice markers, litigation-hold communications — are flagged and routed separately from the general relevance queue for dedicated privilege review.
- 04
Rank the review queue ai
Documents are ranked by likely responsiveness so reviewers work highest-value material first, with low-scoring documents still queued for review rather than excluded from the pool.
- 05
Human review and sign-off output
A reviewer confirms or overrides every relevance and privilege tag before it's finalized — the model's score is a starting point for the reviewer's judgment, never the production decision itself.
- 06
Generate production and privilege logs output
Confirmed tags generate the production set and privilege log entries, with full audit trail of the original AI suggestion, the confidence score, and the reviewer's final call for defensibility.
Inputs
- Full document population for the matter
- Defined issue codes and relevance criteria
- Privilege indicators (attorney names, legal-advice markers)
- Custodian and matter metadata
Outputs
- Ranked review queue by likely responsiveness
- Privilege-flagged document queue
- Reviewer-confirmed production set
- Privilege log with audit trail
Works with
Prefer a fully custom build instead of an off-the-shelf integration? We scope both options during your free consultation — most jobs like this one work fine on standard connectors, but higher-volume or non-standard systems sometimes need bespoke API work, reflected in the complex tier.
Where this goes wrong if you get it wrong
- This tool must never produce a document or exclude it from the review pool based on the AI's relevance or privilege score alone — technology-assisted review is defensible in litigation precisely because a qualified human reviewer confirms every tag, and skipping that step turns a genuine efficiency gain into a discovery sanctions risk if a privileged document is produced or a responsive document is withheld.
- Privilege scoring trained to catch obvious markers — attorney names in the correspondence header, 'privileged and confidential' in a subject line — can miss privilege that arises from context rather than explicit markers, such as a business email forwarding legal advice without repeating the privileged content itself; these context-dependent privilege calls need a reviewer's substantive read, not a keyword match.
- Relevance scoring calibrated against issue codes defined early in a matter can drift out of alignment as the legal theory develops or new issue codes get added mid-review, and documents scored against an outdated issue-code set need to be re-scored, not left with a relevance tag from criteria the case no longer turns on.
- The defensibility of a technology-assisted review process depends on being able to show the methodology and validation sampling to opposing counsel or the court if challenged — every score, override and reviewer decision needs to be logged with enough detail to reconstruct and defend the process, not just to produce the final tagged document set.
Frequently asked questions
Does the AI make the final relevance or privilege call on any document?
No. Every tag it assigns is a suggestion with a confidence score, and a qualified reviewer confirms or overrides it before the document moves to production or a privilege log entry — this is what makes technology-assisted review defensible in litigation.
How are privileged documents handled differently from ordinary relevance scoring?
Documents matching privilege indicators are routed to a separate, dedicated privilege-review queue rather than the general responsiveness queue, since a privilege call carries different legal consequences and typically needs a more senior reviewer.
Can this reduce document review costs in large litigation matters?
Yes — ranking the review queue by likely responsiveness and separating privilege review means reviewer time concentrates on the documents most likely to matter, which is where most technology-assisted review cost savings come from, though every document still gets reviewed.
Is this process defensible if challenged by opposing counsel?
The methodology, confidence scores, and every reviewer override are logged, which supports the kind of validation and audit trail courts have generally required to find technology-assisted review defensible, but you should confirm your specific approach with litigation counsel for the matter.