Legal document review is the central cost driver in two major legal practice contexts: contract analysis in transactional work and document review in litigation. In both contexts, the volume of documents requiring review frequently exceeds what human reviewers can process cost-effectively — M&A due diligence might require reviewing 2,000 contracts in a three-week timeline; litigation discovery might involve 500,000 emails requiring relevance and privilege classification. AI document review makes these volumes manageable.
For transactional attorneys, AI contract review changes the economics of due diligence. A private equity firm evaluating an acquisition target with 800 vendor contracts no longer needs to choose between a thorough contract review and a commercially viable timeline. AI can perform an initial pass across all 800 contracts in hours, extracting key provisions and flagging issues for attorney review. The attorney's time is then spent on the flagged issues rather than initial screening — a fundamentally different and more efficient use of legal expertise.
For litigators, AI eDiscovery review reduces review costs by automating the initial relevance determination across large document sets. Traditional manual review of litigation documents costs the client in attorney hours at a rate that can be prohibitive for mid-size commercial litigation. Technology-assisted review (TAR), which uses AI trained on attorney-coded documents to classify the remaining set, has been validated by courts and is now an accepted methodology for large-scale discovery.
How It Works
AI document review operates through different mechanisms depending on the document context.
For contract review, AI tools use a combination of natural language processing and trained classification models to identify specific clause types within contract documents. When Luminance or Kira Systems reviews a contract, the AI reads the document, identifies sections by function (definitions, representations and warranties, limitation of liability, termination provisions), extracts the relevant text from each section, and classifies the extracted content against a taxonomy of issue categories. The output is a structured data set: this contract has a limitation of liability that caps damages at 12 months of fees, with carve-outs for IP indemnification and data breaches; this contract has an auto-renewal provision with a 45-day notice period; this contract is governed by UK law and requires arbitration in London.
For due diligence review, AI contract analysis performs this extraction across an entire document set simultaneously, producing a structured summary of all identified issues across all reviewed agreements. Kira Systems is designed for this high-volume due diligence workflow, allowing transactional teams to upload a document set and receive a populated issue summary organized by clause category and risk level.
For eDiscovery, the mechanism is different. Technology-Assisted Review (TAR) uses a training process: attorneys review and code a representative sample of documents as relevant or non-relevant, and the AI learns from these training decisions to classify the remaining document population. The two primary TAR protocols are TAR 1.0 (simple active learning, with a fixed training phase followed by batch classification) and TAR 2.0 (continuous active learning, where the AI continues to learn from attorney decisions throughout the review). Everlaw and Relativity AI implement TAR within their respective eDiscovery platforms.
Privilege review is a distinct AI function within eDiscovery — identifying documents that may be protected by attorney-client privilege or work product doctrine before production to avoid inadvertent disclosure. AI privilege review uses trained models to identify documents that contain attorney names, legal advice patterns, and privilege markers, flagging them for human privilege review rather than production.
Key Considerations for Law Firms
- Training data quality determines review quality. AI contract review tools perform best on document types they were trained on. A tool trained primarily on US commercial agreements will perform less well on UK law agreements, franchise agreements, or highly bespoke structures that diverge from standard commercial forms. Evaluate tools with your specific document types before committing.
- Sampling is required to validate AI review. Both contract review and eDiscovery AI require sampling protocols to validate that the AI is performing at an acceptable accuracy level. For contract review, sample a random set of AI-reviewed documents to verify that the AI is capturing all clause types accurately. For eDiscovery TAR, implement a statistical recall validation methodology to confirm that the review has identified a sufficiently high proportion of relevant documents.
- Human oversight requirements differ by matter sensitivity. For low-stakes contract review (vendor screening, standard template compliance), AI review with light attorney supervision may be appropriate. For high-stakes transactions (M&A due diligence, major litigation), AI review should be treated as a screening pass with thorough attorney review of all flagged and a statistically significant sample of unflagged documents.
- eDiscovery AI tools require TAR validation methodology. Courts scrutinize eDiscovery methodology increasingly closely. Attorneys implementing TAR must be prepared to explain and defend the validation protocol — seed set construction, recall measurement, stopping criteria — and may need to produce validation documentation to opposing counsel or the court.
- Confidentiality during AI review. Uploading contract or litigation documents to a cloud-based AI review tool inputs potentially privileged, confidential, or trade secret information to a third-party platform. Confirm the platform's data handling, retention, and security practices before uploading sensitive documents.
Limitations and Risks
AI document review requires human supervision for high-stakes decisions. The Mata v. Avianca scenario — where attorneys relied on AI output without adequate review — illustrates the consequences of treating AI review as a substitute rather than a complement for attorney judgment. For document review, the equivalent risk is producing a document set in litigation that the AI classified as non-privileged but that includes privileged communications, or signing off on an M&A due diligence report based on AI extractions that missed a material contract term. The attorney is accountable for these decisions regardless of whether AI performed the initial review.
Accuracy rates vary significantly by document type and training data. AI contract review tools are tested on common commercial agreement types; accuracy on bespoke transactions, specialized regulated agreements (energy contracts, securities agreements, healthcare contracts), and non-English-language documents is typically lower. Vendors that report accuracy rates on controlled test sets may not have tested on the specific document types in a firm's practice.
eDiscovery AI tools require TAR validation methodology that not all legal teams understand or can implement correctly. A TAR protocol that uses a poorly constructed seed set, applies incorrect stopping criteria, or lacks recall validation may produce a defensible-looking workflow with an unacceptably low recall rate. The responsibility for implementing TAR correctly falls on the supervising attorney, not the vendor.