Unnamed Law Firm

In civil litigation, document review is one of the largest and most unpredictable cost drivers for defendants. When a large global law firm collected and processed over 1 million documents on behalf of a defendant, the dispute over keyword search terms threatened to make costs untenable. Opposing counsel pushed for a broad keyword list that would have required reviewing nearly 200,000 documents — a tenfold increase over what defense counsel considered necessary. The firm lacked in-house expertise in predictive coding and needed a credible, court-defensible methodology to negotiate an alternative before the inflated review costs became unavoidable.

The firm engaged Consilio to design and execute a predictive coding (Technology-Assisted Review) workflow using machine learning and predictive analytics, negotiating an agreement where opposing counsel accepted the broader keyword list in exchange for allowing TAR to cull the set before review. Consilio guided the attorney team — who had limited TAR experience — through the full workflow: protocol design, quality assurance routines, and the specific technical vocabulary needed to negotiate with opposing counsel. The implementation ran four iterative training rounds: an initial random sample of 2,057 documents immediately set aside 162,000 as non-responsive; subsequent rounds used biased draws, additional random samples, and a disagreement-reversal technique to refine model accuracy. A parallel second-level review workstream was established so junior attorneys could begin reviewing high-probability responsive documents concurrently, compressing the overall timeline.

Predictive coding eliminated 89% of the 200,000-document corpus from first-level review, reducing the final review pool to approximately 21,500 documents while hitting the agreed 90% ±5% recall target — a threshold defensible under court scrutiny. Key outcomes:

• $209,000+ saved in first-level review costs (at ~$1.25/document)
• 162,000 documents (83%) set aside after just the first training round
• Documents produced one day earlier than a conventional keyword-only workflow would have allowed
• Opposing counsel, initially skeptical of TAR, accepted the methodology after Consilio's experts explained the protocol — removing a significant negotiation risk before review began

• Secure opposing counsel buy-in early: TAR only delivers cost savings if the protocol is agreed before review begins; expert-facilitated explanation of recall targets and workflow is often what converts a skeptical adversary.
• Catch human coding errors before they corrupt the model: Document-family coding shortcuts (reviewing a family without per-document decisions) introduced noise that required a dedicated QA pass to detect and correct.
• Use disagreement-reversal rounds to surface reviewer inconsistency: When the model's probability scores conflict sharply with human coding, those documents are most likely to reveal reviewer errors that marginal training rounds will not fix.
• Parallel second-level review compresses timelines: Starting high-confidence responsive document review concurrently with ongoing training rounds can recover days otherwise lost to sequential staging.