AI for Legal Document Review: What Works, What Does Not, and What Is Coming

What AI Does Well in Legal Document Review

— Standard clause extraction: identify and extract defined clause types (termination, liability limitation, notice periods, governing law) with high accuracy when trained on representative examples

— Deviation flagging: compare extracted clauses against a baseline standard and flag deviations for lawyer review — faster and more consistent than manual red-lining

— Key date and obligation extraction: pull payment dates, renewal dates, notice deadlines, and key obligations from complex documents

— Consistency checking: identify inconsistencies within a document (definitions used before being defined, cross-references that do not match)

— Summary generation: produce structured summaries of contract terms for business stakeholders who need accessible overviews

— Bulk classification and triage: classify large document populations by type, risk level, or priority for targeted human review

What AI Cannot Reliably Do

— Legal advice: AI can identify and summarise terms; it cannot advise on enforceability, applicable law interpretation, or negotiating strategy

— Novel or complex clause interpretation: unusual clause structures, jurisdiction-specific interpretation nuances, and ambiguous language require lawyer judgment

— Risk assessment requiring context: whether a clause represents acceptable risk depends on context (counterparty, deal size, relationship) that AI cannot fully assess

— Negotiation support: AI cannot evaluate what is achievable in negotiation or prioritise where to spend negotiating capital

— Final sign-off: AI-assisted review should accelerate lawyer review, not replace it — final legal judgment requires a qualified lawyer

System Design Requirements for Legal AI

— Human review mandatory: every AI extraction that will inform a legal decision must be reviewed by a qualified lawyer — design the review interface to make this fast, not optional

— Audit trail: log every AI extraction, the source text it was extracted from, and the reviewer who confirmed or corrected it

— Confidence scoring with review triggers: extractions below a defined confidence threshold should automatically require review regardless of category

— Version tracking: legal documents are often revised during negotiation — the system must track which version each extraction came from

— Data isolation: legal documents are often covered by privilege — strict access controls and tenant isolation are required