The State of AI E-Discovery
Technology-assisted review (TAR) has been a standard tool in large document review projects for nearly a decade. What changed in 2024-2025 is the underlying engine: predictive coding and keyword search have been supplemented — and in some deployments supplanted — by large language models capable of substantive legal analysis, not just categorization.
The result is a qualitatively different capability. Instead of asking "is this document relevant?" AI can now answer "does this document undermine or support the plaintiff's theory of the case?" — a judgment that until recently required a trained attorney.
Key Takeaways: (1) AI-first review is now competitive on quality with attorney review for initial cut decisions. (2) The new risk is overconfidence — producing without sampling. (3) Privilege review remains the human-required step AI cannot safely take alone.
Where AI Is Actually Accelerating
First-pass responsiveness review. For large document sets, AI-first review with attorney sampling is now producing substantially equivalent results to full attorney review at a fraction of the cost. Several published cases in 2025 accepted AI-first protocols where the requesting party could not show material advantage from attorney-first approaches.
Deposition preparation. AI tools that can surface the five documents most likely to contradict a deponent's expected testimony have become standard in well-resourced litigation practices. The time saving on deposition prep is, anecdotally, the single most-cited productivity win among litigators we spoke with.
Exhibit identification and chronology construction. The mechanical work of building a timeline and populating an exhibit list from a large document set — work that consumed significant associate hours — is now almost entirely AI-handled in practices that have adopted the workflow.
The New Risk Vectors
Overconfidence in AI responsiveness calls. Courts have sanctioned parties for producing AI-reviewed document sets without adequate attorney sampling. The AI does not know what it does not know — a highly confident responsiveness determination can mask systematic misclassification of an entire document category.
Privilege log AI. Several vendors now offer AI-assisted privilege logging, which can dramatically accelerate a painful process. But courts have shown willingness to scrutinize AI-generated privilege logs more carefully than attorney-generated ones, and a mass-produced AI privilege log with errors can expose an entire review to challenge.
Cross-jurisdictional discovery. AI tools trained primarily on US federal court standards do not automatically apply the different discovery regimes in UK Commercial Court, German courts, or Israeli proceedings. Global litigation teams are learning this the hard way.
The Frontier: Agentic Litigation Workflows
The leading e-discovery vendors are beta-testing what they call agentic review workflows — AI that does not just categorize documents but takes actions: drafting deposition outline sections based on document patterns, flagging likely hot documents for partner attention, automatically updating privilege logs as new documents are produced.
These systems are not production-ready for high-stakes matters yet. The failure modes are unpredictable, the audit trails are inadequate, and courts have not yet provided guidance on what attorney supervision is required.
But the next twelve months will bring production deployments in lower-stakes matters. The firms building familiarity now will have a significant advantage when these tools are ready for prime time.