Discovery review workflow worth building out:
AI builds the document index—type, date, author, subject, preliminary relevance. Semantic search by issue, not by keyword. The attorney makes every privilege and relevance call on the flagged subset.
Semantic search finds what keyword search misses. A party that never wrote "fraud" may have committed plenty of it. Search for what they meant, not what they wrote.
The framing that works: AI handles the organization, attorney handles the judgment. The output should be a cleaner pile, not a smaller obligation.