Using AI, a Tech Company Reduced Review by 80% and Produced to Regulators with Confidence

Key Results

• Document set reduced from 750,000 to 158,000 documents
• Review speed increased from 10 to up to 60 docs/hour from bulk issue coding
• AI privilege model trained for reuse, improving speed and consistency
• 310 key documents identified across six core topics
• Production delivered with confidence, accuracy, and defensibility

The Challenge: A Sweeping Request and a Tight Timeline

At the outset, the FTC issued an expansive request that pulled in more than 750,000 records including emails and Slack messages. Agreement on keywords proved difficult, as regulators pushed for maximum disclosure.

Lighthouse supported outside counsel through nine rounds of negotiation, including line-by-line responses to regulator objections, landing on a refined population of 158,000 documents—an 80% reduction in review scope.

Review at Scale: Fast, Focused, and Credible

One immediate challenge with prioritized document review for this matter was the large number of issue codes. Previously, similarly complex issue codes had slowed review to just 10 documents per hour. Moreover, outside counsel expected the FTC to heavily scrutinize the issue code distribution, making accuracy and nuance critical.

To guide reviewers away from default or overly broad issue codes, a classifier was used to apply issue codes to responsive documents. Three refinement cycles ensured the issue codes were applied accurately, and as counsel directed.  

This method not only passed regulator scrutiny, review was accelerated by 4.5-6x.

Key Document Identification and Proactive QC

In parallel to review, Lighthouse began 4 rolling deliveries of key documents supporting six topics. Over four weeks, the team surfaced a highly-curated set of 310 documents to support case strategy.  

The work to identify key documents was also used to identify documents that were likely under-coded by contract reviewers—enabling the team to course-correct in real time.

Building an AI Privilege Model for This Matter—and the Next

The team next trained an AI privilege model to support privilege review, not just for this matter, but for future matters as well.  

The process began with a linguistically curated sample set, which counsel coded with future applicability in mind. The trained AI privilege classifier analyzed documents, delivering outputs in Relativity, allowing reviewers to reference privilege scores as part of their decision-making.

Technical Execution: Overcoming Real-World Complexity

Behind the scenes, Lighthouse had to address several technical hurdles. The review was hosted in RelativityOne, and initial data connectors required custom adjustments to work with the client’s Slack and metadata structure.

In particular, Lighthouse developed a workaround to solve for Slack transcripts, which didn’t have file extensions, thus breaking standard ingestion processes.