E-Discovery
AI E-Discovery for Small Law Firms: A 2026 Guide
Review costs $18,000 per gigabyte with human reviewers. One firm cut that by 85% using AI. Here’s the case law, cost data, and practical workflow for small firm e-discovery in 2026.
Computational scientist with a Ph.D. from UC Irvine and peer-reviewed research in NLP, deep learning, and large-scale data modeling. Over a decade of experience building systems that process complex document sets at scale. Founded Hintyr to bring defensible AI workflows to litigation teams navigating document review, redaction, and production.
AI E-Discovery Is Here. Most Small Firms Haven’t Started.
One hundred thousand documents. One associate at $100 an hour. Five hundred documents reviewed per day. That’s 2,000 billable hours and a $200,000 bill before a single motion gets filed.
For a contingency-fee plaintiff’s lawyer financing the case out of pocket, those numbers kill the case before it starts. Craig Ball, one of the most cited e-discovery practitioners in the country, has described the structural problem bluntly: “Plaintiffs’ lawyers tend toward frugality (as they are spending their own money) and shy away from capital expenditures that cannot be reliably expensed against the matter.”
RAND found that document review consumes 73% of total e-discovery production costs. The ComplexDiscovery 2025 Review Update puts worldwide review spending at $10.8 billion, roughly 64% of all discovery spending.
And the firms shouldering most of that burden are small ones. Solo practices account for about 40% of all law firms, and firms with fewer than six attorneys represent over 75% of the total. AI e-discovery tools exist that could cut those costs dramatically, and platforms like Hintyr were built specifically for firms at this scale. But most small law firms aren’t using any of them.
Here’s the number that should worry you: only 17.7% of solo practitioners use AI tools for any part of their work, per the ABA’s 2024 Artificial Intelligence TechReport. At firms with 500 or more lawyers, that figure is 47.8%. The gap isn’t a curiosity. It’s a competitive vulnerability with ethical teeth.
Forty-two jurisdictions now mandate technological competence under ABA Model Rule 1.1, Comment 8. That’s 40 states plus D.C. and Puerto Rico. When the ABA’s own AI Task Force warns in its December 2025 report of growing “have/have-not stratification” between firms that can afford AI and those that can’t, the competence mandate stops being abstract. It gets concrete fast.
In 2021, Judge Iain Johnston of the Northern District of Illinois issued a 256-page sanctions opinion in DR Distributors v. 21 Century Smoking that split $2.5 million in fees between a defendant and its former defense counsel. The court held the attorneys personally liable for over $1.25 million because they lacked what the judge called “the basic knowledge, training, and skills to handle properly ESI.” That was five years ago. The bar has only risen since.
Over 300 federal and state judges have issued standing orders, general orders, or local rules addressing AI. Courts have endorsed technology-assisted review as defensible for over a decade. And major vendors now include AI features in standard pricing. The question for small firm litigators isn’t whether AI e-discovery is legitimate. That debate ended years ago.
Five Rulings That Made AI Document Review Defensible
The case law on technology-assisted review is settled. If you’ve been waiting for judicial permission, you already have it.
In 2012, Magistrate Judge Andrew Peck issued the first published decision approving TAR in Da Silva Moore v. Publicis Groupe. He told the bar directly: “Counsel no longer have to worry about being the ‘first’ or ‘guinea pig’ for judicial acceptance of computer-assisted review.” That was 14 years ago.
Three years later, Judge Peck went further in Rio Tinto PLC v. Vale S.A., declaring that TAR is “now black letter law.” Two points from Rio Tinto matter for small firms especially. First, TAR cannot be held to a higher standard than manual review or keyword search. Second, the producing party chooses the methodology unilaterally. You don’t need opposing counsel’s agreement.
In 2016, Hyles v. New York City endorsed continuous active learning (CAL) as the preferred approach and set the standard that matters most for resource-constrained firms: “the standard is not perfection, or using the ‘best’ tool, but whether the search results are reasonable and proportional.” CAL also eliminates seed-set disputes, removing a layer of complexity that earlier TAR versions required.
By 2020, Livingston v. City of Chicago confirmed that keyword pre-culling before running TAR is defensible. The City filtered 1.5 million emails down to 192,000 before applying Relativity’s Active Learning. That approach, narrowing first, then applying AI, is exactly the cost-cutting workflow small firms need.
And for validation, the In re Broiler Chicken Antitrust Litigation protocol from 2018 provides a court-approved template. Special Master Maura Grossman’s framework requires a 3,000-document validation sample across three strata: responsive, non-responsive, and AI-excluded. It’s a replicable process you can follow step by step.
The thread across all five decisions: courts expect a reasonable, proportional, documented review process. Not a perfect one. A reasonable one.
How One Firm Cut Document Review Costs by 85%
RAND priced document review at $18,000 per gigabyte in 2012. The 2026 EDRM Winter Pricing Survey puts GenAI-assisted review at $0.11 to $0.50 per document. That’s not an incremental improvement. It’s a structural shift in how review gets priced.
The numbers at scale tell the story. A 100,000-document manual review at standard rates ($100/hour, 500 documents per day) runs roughly $200,000 and 2,000 hours of attorney time. Purpose Legal completed a 300,000-document review using Relativity’s aiR in one week: 85% cost reduction, over 4,000 hours of manual work eliminated, and more than $70,000 saved. DISCO’s Auto Review processes 32,000 documents per hour with precision and recall above 90%.
Those are enterprise-scale numbers, but the market shift underneath them affects every firm. Major vendors now bundle AI review into their base platform pricing. Price floors for AI-assisted review have dropped so fast that the per-document cost gap between a five-attorney firm and a 500-attorney firm is narrower than it’s ever been. (For a tool-by-tool pricing breakdown, see our comparison page.)
What matters isn’t any single vendor’s price list. It’s the structural reality: the cost barrier that kept small firms out of AI-assisted review two years ago has largely collapsed. The question now is which tools actually fit a small firm’s workflow, not whether you can afford to use them.
That’s where Hintyr sits. The AI agent handles research, tagging, redaction, Bates numbering, and production exports through a single conversational interface, with a Discovery Portal for sharing productions at no additional per-seat cost. Instead of paying for five separate tools and stitching them together, you get one agent that handles the full review cycle. For a small firm running multiple matters, that consolidation is where the cost savings compound.
This matters for proportionality arguments too. Under FRCP Rule 26(b)(1), discovery must be “proportional to the needs of the case.” Judge Peck put the point sharply.
But refusing to adopt technology doesn’t just cost your client more. It weakens your legal position. A court that sees you choosing a $200,000 manual review when AI-assisted alternatives exist at a fraction of the cost isn’t going to sympathize when you argue the request is disproportionate. And that calculus runs both ways: plaintiffs’ firms that adopt AI review can take on cases that the math previously ruled out.
ABA Opinion 512 and Your AI Obligations
ABA Formal Opinion 512, issued in July 2024, is the national ethics framework for lawyers using generative AI. It covers six duties: competence, confidentiality, supervision, candor, fees, and a specific warning about hallucinations.
The competence standard is lower than many attorneys assume. You don’t need to become an AI expert. Opinion 512 requires “a reasonable understanding of the capabilities and risks of any GenAI tool” you use, or the ability to “draw on the expertise of others who can provide this guidance.” The Texas Professional Ethics Committee put it more colorfully in Opinion 705: “This principle apparently wasn’t obvious to the ever-increasing number of lawyers who have been caught submitting made-up citations.”
The supervision duty is where it gets practical. Opinion 512 warns that lawyers who rely on AI “risk many of the same perils as those who have relied on inexperienced or overconfident nonlawyer assistants.” That analogy is useful. You wouldn’t let an unsupervised paralegal sign off on privilege calls. The same logic applies to AI.
This is where tool design matters. Hintyr’s AI agent includes source citations in every document-backed response, each linking to the exact page in the source file. That makes supervision practical: click the citation, read the passage, confirm the agent’s interpretation. For redaction, the agent never applies changes without explicit human approval. And every agent response can be exported to case notes with one click, creating the documented record that both Opinion 512 and the Heppner discoverability ruling demand.
On confidentiality, the Heppner ruling from February 2026 changed the calculus. Judge Jed Rakoff held that conversations between a defendant and Anthropic’s Claude are not protected by attorney-client privilege. The FBI seized 31 AI-generated documents. For firms using consumer AI tools for case work, every input is potentially discoverable.
Oregon’s Formal Opinion 2025-205 draws a practical line: open AI models require informed client consent before inputting confidential data. Closed, enterprise-grade models carry lower risk but still need safeguards. For a deeper analysis, see our guide to privilege risks when using AI tools.
Purpose-built review platforms sidestep much of this risk because the AI operates entirely within your case data. It doesn’t train on your inputs or expose them outside the platform. Hintyr’s AI agent, for example, searches only the documents in your case and retains findings in a case memory shared with your team, not with the model provider or other users.
The billing rules are straightforward. You can’t charge clients for time spent learning a generic AI tool. You can’t bill for time the AI saved you. If per-matter AI costs are hard to isolate, treat them as overhead. No exceptions.
Looking forward: the Colorado AI Act (effective June 2026, delayed from February) lists legal services as a “consequential decision” area. The proposed Federal Rule of Evidence 707, governing machine-generated evidence, goes to committee vote May 7, 2026. California’s COPRAC approved AI amendments to five Rules of Professional Conduct in March 2026, with public comment open until May 4.
And the regulatory direction is clear. Getting ahead of it now costs less than scrambling to comply later. If you’re working through AI disclosure requirements, we’ve covered that separately.
A Practical Workflow for Your Next Document Review
Here’s what AI-assisted review actually looks like for a small firm, step by step. This covers the review and production stages. It assumes you’ve already issued litigation holds, identified custodians, and collected ESI defensibly. If you haven’t, start there. The sanctions in DR Distributors were triggered by preservation and collection failures, not review methodology.
Pre-cull and organize. Before running any AI review, narrow the collection. Livingston endorsed keyword pre-culling on a 1.5 million document collection, and it cuts your reviewable corpus and costs in the same stroke. In Hintyr, the Tag Wizard lets you define filter conditions on file type, date range, custodian, or email metadata, preview the matches, and create smart tags that classify documents automatically. The AI agent can then be scoped to specific tags, so subsequent analysis runs only on the filtered set.
Run AI-assisted categorization. The most capable review tools now offer what practitioners call agentic review: AI that responds to natural-language queries with answers anchored to specific passages in the source documents.
In Hintyr, the AI agent handles this through conversation. Ask it to find every document mentioning a contract clause, and it returns cited answers with clickable links to the exact pages. Ask it to tag the cited files as “Responsive,” and it does so in the same conversation. Research, categorization, privilege flagging, and custodian assignment all happen in one interface, without switching screens.
Review privilege flags with human judgment. The AI agent can mark documents as attorney-client privilege or work product based on criteria you describe, but the decision about what to mark remains yours. No platform’s AI substitutes for that judgment. Document-linked notes let you record privilege observations directly alongside each file: when you open a flagged document, your per-document analysis appears automatically. Build your privilege log as you review, not after.
Validate your results. The Broiler Chicken protocol gives you a court-approved template: draw a stratified sample from the responsive and non-responsive populations and have qualified reviewers grade each document. Grossman and Cormack’s research established 75% recall as the defensible threshold, meaning the AI should correctly identify at least three out of every four relevant documents.
Hintyr’s TAR validation handles the statistics for you. L1 (Control Set) testing measures precision and recall from your responsive population. L2 (Elusion Test) samples the discard pile to check whether your review is meeting its recall target. Set your confidence level and margin of error (the 95%/5% defaults are built in), and Hintyr calculates the required sample size automatically, draws the random samples, and opens a grading panel with live statistics that update as you grade. When finished, export the results as a spreadsheet for your case file. That’s the Broiler Chicken protocol as a built-in workflow rather than a manual exercise.
Prepare for production. This is where small firms typically lose the most time switching between tools or paying vendors for tasks the platform should handle. In Hintyr, the entire production pipeline stays in one place.
Start with redaction. Tell the AI agent to find and flag sensitive content: “Redact all Social Security numbers,” “find and redact all email addresses,” or target a specific individual by name. The agent searches your files, flags matches with confidence scores, and opens a review panel where you approve each one individually. The agent never redacts on its own.
Next, apply Bates numbering. Tell the AI agent to number all files in a tag with the prefix and starting number you specify. It handles the sequence automatically, stamping onto a separate layer so originals stay untouched.
Then build the production. The Export Wizard walks you through filter criteria, document format, Bates settings, and output options in a guided dialog. Or skip the wizard and tell the AI agent what you need: “Create a production of all responsive non-privileged PDFs from 2024 with Bates numbering starting at PROD000001.” The agent configures the export, shows a summary for confirmation, and submits the package. The output includes documents with applied redactions and Bates stamps, a load file, and a manifest.
Share the result through the Discovery Portal, where opposing counsel accesses only the tags you authorize through a controlled interface. No external file-sharing service needed. No per-seat cost for outside users.
Keeping the full production chain in one platform isn’t convenience for its own sake. It’s defensibility. Every time you export documents to a separate tool for redaction or Bates numbering, you introduce metadata-handling risk and the kind of errors that generate spoliation motions.
Document everything. GenAI prompts may themselves be discoverable, as In re OpenAI Copyright Litigation (S.D.N.Y. Sept. 2025) demonstrated. Every prompt. Every output. Part of the case record.
Hintyr provides three layers of documentation for this. Case notes capture observations, privilege analysis, and review strategy in a built-in editor right next to the AI agent and document viewer. Group notes are shared and collaboratively edited; private notes stay visible only to you for attorney work product. Document-linked notes attach annotations to individual files, so your per-document analysis opens automatically whenever any team member views that file.
And when the AI agent produces a finding worth keeping, click “Add to Case Notes” to save the full response, with citations and formatting, as a quoted block in your active note. That single-click capture builds a running audit trail of every AI-assisted analysis you relied on. When you need to share work product, export notes to PDF for reports, court filings, or client communication.
One honest caveat. The old threshold that said “under 5,000 documents, just review manually” assumed tools that took days to learn and required dedicated support staff. That assumption doesn’t hold when AI review is conversational. Hintyr’s agent lets you type a question in plain English and returns cited answers from your documents, no training period, no workflow to memorize.
The barrier has shifted. It’s no longer about volume. What AI still can’t do, no matter how easy the interface, is exercise professional judgment: reading between the lines, weighing credibility, recognizing what’s missing from the record. That part is still yours.
The point isn’t that AI is always the answer. It’s that pretending it doesn’t exist is no longer defensible. For a deeper look at the malpractice exposure from improper AI use (or non-use), see our separate guide. For a side-by-side look at how different platforms handle small firm e-discovery, see our tool comparison.
Full disclosure: we build agentic document review tools at Hintyr, so we have a stake in this conversation. We’ve tried to ground every claim in this post in case law, ethics opinions, and industry data you can verify independently.
If you want to see what agentic review looks like in practice, asking questions, getting cited answers, and moving from research to redaction to production in a single conversation, the documentation walks through every capability. The legal framework is settled and the economics have flipped. What remains is execution. Yours.
Disclaimer: This blog post is published by Hintyr for informational purposes only and does not constitute legal advice. The discussion of ethics rules, case law, and e-discovery practices is general in nature and may not reflect the rules applicable in your jurisdiction. Attorneys should consult their state bar’s ethics opinions and qualified legal counsel before making AI compliance or e-discovery decisions. No attorney-client relationship is created by reading this post.
AI-powered review you can actually verify.
Every document-backed answer links to its source passage. Statistical validation measures your review with court-recognized metrics. And production, from redaction to Bates numbering to opposing counsel delivery, runs in the same platform without a single file export.