AI Document Review for Law Firms: The Practical Guide

Picture this: your client is three months into a commercial dispute. The other side produces 47,000 documents. You have four weeks until the discovery deadline, a paralegal billing at $85 per hour, and a matter budget that was set before anyone anticipated this volume. Running the math: if your paralegal reviews around 50 documents per hour, covering the full set takes roughly 940 hours. That is about $80,000 in paralegal time alone, before attorney supervision. The client did not budget for that. Neither did you.

This is the scenario where AI document review earns its place in a law firm. Not as a magic box that reads files and makes legal judgments, but as a workflow that processes large volumes of electronic documents faster and more consistently than a human team can, surfaces the likely relevant material first, and lets your attorneys spend their time on documents that actually matter.

This guide covers how AI document review works in practice, where it fits in eDiscovery versus due diligence, what courts have said about it, which tools make sense at different firm sizes, and the guardrails you cannot skip.

Two Very Different Use Cases for Document Review AI

When lawyers say "document review," they usually mean one of two things. Understanding the distinction matters because the tools, workflows, and stakes are different.

eDiscovery document review happens in litigation. Opposing counsel produces documents, or you produce documents in response to requests, and someone has to review every document for relevance, privilege, and, where applicable, confidentiality designations. In large commercial disputes, IP cases, employment litigation, or government investigations, the document counts run into the tens or hundreds of thousands.

Due diligence document review happens in transactional work: mergers and acquisitions, real estate acquisitions, corporate finance, or any deal where your client needs to understand what they are buying or lending against. A data room for a mid-market acquisition might contain 5,000 to 30,000 documents, and someone needs to extract key information from contracts, identify change-of-control provisions, flag problematic indemnification clauses, and surface anything that could affect the deal.

AI handles both workflows, but with different tools and different validation requirements. Most of the established guidance from courts applies specifically to eDiscovery; the due diligence side is less litigated but follows the same logic of documented, defensible process.

How Technology-Assisted Review Actually Works

The foundational technology in AI eDiscovery is called technology-assisted review, commonly abbreviated TAR and sometimes called predictive coding. It has been in the legal market since roughly 2010 and has a well-developed track record.

Here is how it works at the process level:

Step 1: Ingest and process. The document set is uploaded to the platform. The platform converts everything to a format it can analyze, running optical character recognition on scanned documents, extracting metadata, and normalizing file formats across email, PDFs, and office documents.

Step 2: Seed set review. Attorneys or paralegals review a subset of documents, typically several hundred to a few thousand, and manually tag them as relevant or not relevant, privileged or not privileged. This seed set teaches the model what the firm considers responsive for the specific case.

Step 3: Model training and ranking. The AI learns from the seed set and ranks every document in the full collection by predicted relevance. The most likely-relevant documents surface at the top.

Step 4: Prioritized human review. The legal team reviews documents in ranked order, starting with the highest-probability-relevant material. The model continues to update as reviewers code more documents. This is called continuous active learning, or CAL, and it is now the more common approach in enterprise platforms.

Step 5: Validation. Before closing out the review, the team runs a statistical validation to estimate recall: roughly, what percentage of the truly relevant documents did the process find? Courts increasingly expect this step to be documented and available for disclosure.

The practical effect is that instead of reviewing all 47,000 documents, your team might need to fully review 10,000 to 15,000 to reach high confidence that you have captured the relevant material. The others can be set aside or spot-checked. On a large matter, that compression is the difference between a feasible review budget and one that breaks the economics of the case.

What Courts Have Said About AI-Assisted Review

Courts have accepted AI-assisted review since at least 2012. The landmark case is Da Silva Moore v. Publicis Groupe, where Magistrate Judge Andrew Peck of the Southern District of New York endorsed predictive coding as an appropriate and cost-effective method for discovery. In the years since, courts have consistently affirmed that computer-assisted review is not inherently inferior to manual review and, in some contexts, produces more consistent and defensible results.

But court acceptance does not mean you can point to the AI and call the review complete. What judges and opposing counsel want to see is a defensible process: documentation of how the model was trained, what validation steps were taken, who made the privilege determinations, and what recall rate the review achieved. The software vendor does not bear that burden. You do.

The courts are not evaluating the software. They are evaluating the process. If challenged, you need to explain your methodology the same way you would explain any quality-control decision in your practice.

Which Tools Fit Which Firm

The eDiscovery software market has consolidated significantly in recent years. Here is how the main platforms currently position for law firms:

Relativity (RelativityOne) is the enterprise standard. It is dominant across large litigation firms, and their aiR for Review and aiR for Privilege tools are now bundled into the standard RelativityOne subscription. If your firm already runs on Relativity, the AI review tools are available to activate. If not, Relativity is priced and structured for large litigation departments and is probably not the right starting point for a small or mid-size firm that handles occasional discovery.

Everlaw is a cloud-native platform positioned for litigation teams that want a cleaner interface and faster onboarding than Relativity. Its AI tools are built into the platform, and reviewers can run AI-assisted ranking, issue tagging, and document summarization without separate licensing. Mid-size litigation firms that handle complex commercial matters often find Everlaw fits better than the enterprise Relativity stack.

Logikcull (now part of the Reveal platform following Reveal's acquisition) targets small and mid-size firms handling eDiscovery without a full litigation support department. Logikcull has published a per-gigabyte pricing model that lets firms upload, process, review, and produce without a large upfront commitment. The self-serve architecture works well for matters where you need to handle discovery on a specific case rather than maintain a standing platform.

Casepoint is another mid-market option with AI-assisted review built in via their CaseAssist feature. It is worth evaluating for firms in the mid-size range that handle a mix of litigation and regulatory investigations.

For due diligence work specifically, the dominant tools are different from the eDiscovery stack:

Kira Systems is one of the most mature contract analysis platforms for M&A review. It excels at extracting specific clause types across large sets of agreements, which is exactly what a due diligence review requires: find every change-of-control clause in 400 contracts, or every assignment restriction, fast.

Harvey is a generative AI platform that has gained traction in large-firm M&A and commercial work. It handles more open-ended analysis than Kira's structured extraction and is useful for summarizing documents and identifying issues that do not fit neatly into predefined clause categories.

Spellbook targets smaller transactional firms and integrates directly into Microsoft Word, making it accessible for attorneys who do not want to export documents into a separate platform for analysis.

The Privilege Question

Privilege review is the most consequential part of any document review, and it is where AI tools require the most careful human oversight.

AI platforms can flag documents as potentially privileged based on patterns: the presence of attorney names, phrases associated with legal advice, communication structures that suggest attorney-client exchanges. But AI does not understand the substance of a communication the way an attorney does, and it produces both false positives (flagging non-privileged documents as privileged) and false negatives (missing documents that are actually privileged).

The standard approach is to use AI to surface likely-privileged documents for priority human review, not to treat AI privilege designations as final determinations. Many platforms maintain a separate privilege-review queue. Every document on that queue needs eyes on it from a licensed attorney before it goes on the privilege log.

A privilege log based on AI classification alone, without attorney review of the underlying documents, is not a privilege log you want to defend in court. The AI narrows the pile. The attorney makes the call.

Confidentiality and What ABA Formal Opinion 512 Requires

Before uploading client documents to any AI-assisted review platform, you need to understand where the data goes and what contractual protections cover it.

ABA Formal Opinion 512, issued in July 2024, makes clear that confidentiality obligations under Model Rule 1.6 apply fully to AI tool usage. The analysis is the same whether you are using a general-purpose AI tool or a dedicated legal platform: you need to understand how the vendor handles data, whether uploaded documents can be used to train models, and what enterprise data controls exist.

For eDiscovery and document review platforms that handle client litigation or deal documents, look for:

• Data stored in a secure, isolated environment rather than shared infrastructure across customers
• Contractual prohibitions on using your uploaded documents for model training
• Compliance with relevant security standards (SOC 2 Type II certification is standard in this space)
• A data processing agreement that covers your firm's obligations under Rule 1.6

Relativity, Everlaw, Logikcull, Casepoint, Kira, and Harvey all publish their security posture and operate with architecture designed for confidential legal data. But you need to review the specific terms for your firm's agreement, not assume vendor marketing covers everything. The confidentiality obligation runs from you to your client, not from the vendor.

State bar ethics guidance is evolving quickly. As of mid-2026, more than 35 state bar associations have issued guidance on AI use in legal practice, and confidentiality in third-party tool usage is a recurring theme. Verify that your state bar has not issued guidance that imposes obligations beyond what the ABA opinion covers.

A Practical Checklist for a Defensible Review

Whether you are doing eDiscovery or due diligence, these process steps keep your workflow documented and defensible:

Before the review starts:

• Document the platform, version, and configuration used for the matter
• Record the seed set size, who coded the seed set, and the criteria applied
• Confirm your vendor agreement covers data protection for this matter type

During the review:

• Keep an audit log of who reviewed which documents and when (most platforms maintain this automatically; confirm the logs are available for export)
• Designate a supervising attorney responsible for quality control and methodology decisions
• Track any second-level review or override of AI classifications

Before closing the review:

• Run the platform's validation or quality-control report and document the recall estimate
• Have a licensed attorney review every document on the potential-privilege queue
• Confirm the production set was reviewed for responsiveness and privilege before export

Data handling at close:

• Confirm your data retention and deletion obligations under the vendor agreement and your firm's policies

This list is process documentation, not legal advice about your specific obligations in a specific matter. Your jurisdiction, your court's standing orders, and your ESI protocol with opposing counsel may impose additional requirements.

Where to Start If Your Firm Is New to AI Document Review

If your firm handles occasional litigation or eDiscovery but does not have a standing platform relationship, here is a realistic starting sequence:

For a specific matter: Logikcull's per-gigabyte model lets you spin up a project, process documents, run AI-assisted review, and produce without a long-term contract or a platform administrator. The per-GB pricing makes costs predictable and passable to the client. It is a practical first step for a small firm handling a moderately complex discovery matter.

For transactional work: If you handle deals and currently do due diligence manually, Kira or Spellbook are worth evaluating. Start with a specific clause type or document category where you have enough volume to measure time savings concretely, rather than trying to automate the full due diligence workflow on the first matter.

Do not start with your highest-stakes matter. Build confidence in the platform on a matter where the volume and risk allow you to validate output carefully before you rely on the AI to cover a set you cannot fully review manually. Once you have seen how the tool performs on your matter type, you can scale up with confidence.

The Bigger Picture

AI document review is not new, but it is becoming accessible at a price and technical level that works for small and mid-size firms, not just large litigation departments. The tools have matured, the court acceptance is well-established going back over a decade, and the main variable is whether your firm has the process discipline to use them in a way that holds up to scrutiny.

The AI can rank documents and surface patterns faster than any human team. It cannot decide what is relevant to your case, make privilege determinations, or certify your discovery responses. Those responsibilities stay with the attorney. The firms that get genuine value from AI document review are the ones that treat it as a workflow accelerator under attorney supervision, not a substitute for that supervision.

If your firm handles any significant volume of litigation discovery or transactional due diligence, the question in 2026 is not whether AI document review is worth investigating. It is which tool fits your matter types and whether your firm has the process in place to use it defensibly.