Document Review Automation in Legal Practice

Document review is one of the largest categories of legal work amenable to automation. This article catalogs working patterns.

The document review challenge

Document review (e.g., e-discovery) is typically:

• 80-90% of e-discovery cost (Rand Corp studies)
• High-volume (millions of documents)
• Repetitive (similar tasks)
• Quality-sensitive (privilege, responsiveness)

Working automation patterns

Pattern 1: TAR (Technology-Assisted Review)

TAR (continuous active learning) reduces review cost:

• 30-70% reduction in documents reviewed
• Standard in e-discovery practice
• Court-approved in Da Silva Moore (S.D.N.Y. 2012) and subsequent cases
• Quality control: random sampling + elusion testing

Pattern 2: AI-assisted privilege review

AI flags:

• Attorney-client communications
• Work product
• Common interest doctrine
• Joint defense privilege

Working patterns:

• AI ranks documents by privilege likelihood
• Lawyer reviews top candidates
• Threshold tuning per matter
• Privilege log auto-generated

Pattern 3: AI responsiveness review

AI classifies:

• Responsive vs. non-responsive
• Hot vs. cold
• Issue tags (auto-tag from review protocol)
• Privilege flags

Pattern 4: Deposition transcript analysis

AI assists:

• Deposition summary (long depositions)
• Issue tagging across witnesses
• Contradiction detection
• Citation generation

Pattern 5: Contract review

AI assists:

• Clause extraction
• Deviation from standard
• Risk flagging
• Multi-contract comparison

Quality control

Validation patterns

Working QC:

• Random sampling (5-10% of AI decisions)
• Elusion testing (find privileged docs AI missed)
• Inter-reviewer agreement (lawyer + AI agreement rate)
• Adversarial testing (red team finds AI blind spots)

Bias auditing

Per ABA Formal Opinion 512 + 533:

• AI bias can affect review (e.g., deprioritizing certain document types)
• Audit protected-group patterns
• Document audit process

Vendor options

Major vendors:

• Relativity (RelativityOne)
• Reveal (Brainspace)
• Everlaw
• DISCO
• Logikcull

Open-source:

• EDRM (reference data model)
• BlackBoiler (contracts)
• Clause (open-source contract review)

Acknowledgments

This article summarizes public sources including ABA Formal Opinion 533, Rand Corp e-discovery studies, EDRM reference model.

Dillon Deutsch has built document review systems. https://courtgpt.ai