AI does first-pass document review and summarization well enough to save a paralegal several hours a week, and remains unfit for anything citable: fabricated authorities are still routine in 2026, so every case, statute, and clause reference must be verified in Westlaw or Lexis before it leaves your desk.
Spellbook, Claude, Harvey, and ChatGPT were tested on document review, summarization, drafting, and legal research tasks for paralegals. This AI for paralegals verdict was issued in August 2026.
- ✓First-pass contract review flagging missing clauses and unusual terms
- ✓Summarizing long deposition transcripts and medical records into chronologies
- ✓Drafting routine documents like engagement letters and discovery requests
- ✓Classifying, renaming, and indexing unstructured document productions
- ×Legal research citations with fabricated case authorities
- ×Privilege and confidentiality judgment requiring facts the model cannot know
- ×Jurisdiction-specific procedure, local rules, and filing deadlines
Scorecard
Fabricated citations look identical to real ones, making output unreliable for legal research.
Fits well into review and drafting workflows but not into citation-dependent tasks.
Handles long documents like 400-page records effectively for summarization.
Quick to adopt for drafting and review, but requires discipline to avoid citation errors.
Wrong citations are presented with the same confidence as correct ones.
A persistent matter file that stores verified citations, the procedural calendar with sources, and prior privilege decisions. Every AI session would read this context first and auto-check any cited authority against the verified list, replacing the manual per-session workaround paralegals use today.