AI document review is the oldest, most judicially tested use of AI in legal practice. Technology-assisted review (TAR) received its landmark judicial endorsement in Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012), where Magistrate Judge Andrew Peck held that computer-assisted review is an acceptable way to search for relevant ESI — and by Rio Tinto v. Vale (S.D.N.Y. 2015) he could write that it is "black letter law" that parties may use TAR. First-generation TAR trained a classifier on lawyer-coded seed sets; TAR 2.0's continuous active learning (CAL) improved on it by constantly reprioritizing the review queue as reviewers code. Canadian courts and the Sedona Canada Principles endorse proportionate, technology-assisted approaches to the same effect.
The judicial acceptance of TAR was hard-won and instructive. Early objectors argued machine classification was a black box; courts answered that human review was the real black box — inconsistent, unauditable, and empirically worse — and that what mattered was measurable output quality, not the reviewer's species. That framing, built across a decade of decisions and Sedona Conference commentary, is precisely why LLM review is entering practice with far less friction: the defensibility framework (documented process, statistical validation, proportionality) was already built, and the new tools simply plug into it with better economics.
The LLM generation changes the mechanics again. Instead of training a classifier per matter, large-language-model review (Relativity aiR, Everlaw's AI review features, and peers) applies natural-language relevance criteria directly: you describe what makes a document responsive, privileged, or hot, and the model classifies and explains its call document by document. Early published benchmarks and vendor validations show LLM review matching or beating human contract-reviewer accuracy on responsiveness calls, with recall/precision measurable the same way TAR always was. The through-line across all three generations: courts accept machine classification when the process is transparent, validated, and proportionate.
This guide sits in our AI in Legal Practice library beside its transactional sibling, AI contract drafting — due diligence review shares most of discovery's mechanics with fewer procedural constraints.
Defensibility has never meant perfection — human review misses documents too, and studies since the TREC Legal Track have shown human reviewer agreement rates far lower than most lawyers assume. What courts require is a reasonable, documented, validated process:
The emerging LLM-specific question is prompt disclosure: whether the relevance instructions given to the model are discoverable process or protected work product. Protocols negotiated in 2024–2026 have gone both ways; until appellate law settles it, assume your prompts may be seen and write them accordingly.
The cost comparison is stark. Traditional first-pass review with contract attorneys runs roughly $25–$60 per reviewer-hour through staffing agencies, and at 40–60 documents per hour that translates to roughly $0.50–$1.50 per document before QC layers, project management, and re-review. LLM first-pass review is priced per document or per gigabyte and typically lands at a fraction of that — often cited in the $0.05–$0.35 per-document range depending on platform and volume — while running in hours instead of weeks. On a 500,000-document matter, first-pass review that cost $400,000+ and six weeks with a 40-person contract team becomes a five-figure compute bill and a validation exercise measured in days.
The second-order economic effect is on case strategy itself. When first-pass review cost half a million dollars, review costs were settlement leverage and motions to limit scope were worth fighting; when it costs a twentieth of that, marginal custodians and date ranges stop being worth a discovery dispute, and the proportionality calculus under FRCP 26(b)(1) shifts for both sides. Litigators who understand the new cost curve negotiate ESI protocols differently — and clients who understand it are already asking why review line-items still look like 2019.
The honest caveats: senior-lawyer time shifts to criteria design, sampling, and QC rather than disappearing; hot-document and privilege calls still get human review; small matters may not justify platform minimums; and the contract-attorney market is not vanishing so much as moving up-stack into validation and QC roles. Due diligence in M&A shows the same curve — AI-assisted diligence platforms now extract change-of-control, assignment, and termination provisions across data rooms at a speed that has repriced diligence fixed fees across the market.
Pricing models also matter to comparability: per-document pricing punishes over-collection while per-gigabyte pricing punishes rich media, and both reward the culling and deduplication discipline that good information governance was always supposed to deliver. Firms quoting AI review to clients should quote the full stack — processing, hosting, model passes, validation lawyer time — because a headline per-document rate that omits validation reads as bait-and-switch when the invoice arrives.
A sensible adoption path: start with a retrospective pilot — rerun a completed matter's review with the AI tool and compare calls against the human coding you already trust. Negotiate ESI protocols that permit (or at least do not preclude) AI review before you need it. Build a standard validation playbook — target recall, sampling design, elusion testing, documentation template — so every matter follows the same defensible script. Keep client communication explicit: review methodology, cost implications, and confidentiality terms of the platform all belong in the engagement conversation, and the billing consequences of a 90% cost reduction are governed by the reasonable-fee analysis in our guide to AI billing ethics.
Watch the privilege frontier in particular. First-pass privilege screening by LLMs is improving fast, but privilege calls carry asymmetric consequences — a missed responsive document is a discovery dispute, a produced privileged document is a crisis — so the standard of care that is emerging pairs model screening with human review of everything the model flags as close, plus a 502(d) order in every US federal matter as the structural backstop. Canadian counsel, lacking a 502(d) analogue, lean harder on negotiated clawback terms and the implied undertaking, which makes protocol drafting the real privilege-protection work.
Vendor diligence mirrors every other AI tool: no training on your data, controlled retention, data residency (a live issue for Canadian matters and cross-border productions), SOC 2 attestation, and audit logs. Litigation-support teams that master this stack are becoming a genuine competitive edge for mid-sized firms competing against bigger discovery budgets — the same asymmetry AI creates in marketing, where our comparison guides and AI SEO service show smaller firms outranking larger ones. To plan either side, talk to LexScale.ai.
LexScale.ai helps law firms across Canada and the United States adopt AI for growth — from client-facing intake and content systems to the visibility that puts your firm inside AI answers.
Book a Free Strategy Call →This article is general information, not legal or ethics advice. Professional-conduct rules on AI are evolving and vary by jurisdiction — always verify current requirements with your state bar, law society, or regulator before adopting any AI workflow.
Related Articles
AI Hallucinations in Legal Research: Risks & Fixes · Bar Rules on AI for Lawyers: US & Canada Guidance · Can Lawyers Use ChatGPT? Ethics, Risks & Workflow · Client Confidentiality and AI Tools for Lawyers · Court Rules on AI in Filings: US & Canada Guide · Billing Ethics for AI-Assisted Legal Work
Ready to grow your firm with AI?