Sort Through Mountains: E-Discovery Simplified

AI in E-Discovery and Document Review

Part 1: Three Things Most Legal Professionals Have Wrong

Most legal professionals believe AI in e-discovery is either a futuristic luxury for BigLaw firms, a risky black box that courts won't accept, or a tool that replaces junior associates and paralegals. All three beliefs are wrong, and holding onto them is costing firms real money, real time, and real competitive ground. A mid-size litigation team that reviewed 200,000 documents manually in 2019 might spend six figures on that review. Today, the same volume can be processed in a fraction of the time at a fraction of the cost. The technology isn't coming. It's already standard practice in federal litigation, regulatory investigations, and corporate due diligence. What matters now is understanding what AI actually does in document review, what it doesn't do, and how non-technical legal professionals can use it confidently and responsibly.

Myth 1: AI E-Discovery Is Only for Big Firms with Big Budgets

The assumption goes like this: AI-powered e-discovery requires enterprise contracts, dedicated IT staff, and a litigation support team. That was partially true in 2015. It is not true now. Platforms like Relativity, Everlaw, Logikcull, and Casepoint have all released cloud-based versions that small and mid-size firms can access on a per-matter or monthly subscription basis. Logikcull, for example, charges by the gigabyte of data processed, meaning a solo practitioner handling a wrongful termination case with 15,000 emails pays only for what she uses. There is no server to install, no IT department required, and no six-figure commitment. The cost barrier that defined e-discovery a decade ago has largely collapsed.

The real barrier today is not money, it's awareness and workflow habits. Many attorneys at boutique firms still default to manual keyword searching because that's what they learned in law school or early in their careers. Keyword searches are fast to set up but notoriously inaccurate. Studies from the RAND Corporation found that keyword searches in e-discovery miss 60 to 80 percent of relevant documents in complex matters. That's not a small margin of error. That's the difference between finding the smoking-gun email and missing it entirely. AI-assisted review, even at the entry level, dramatically narrows that gap.

Consider a regional employment law firm handling a class-action wage dispute. The opposing party produces 80,000 documents. The firm's two paralegals would need months to review that manually. With Everlaw's AI-assisted review, they can upload the production, train the system on a seed set of relevant and irrelevant examples, and let the tool prioritize the documents most likely to matter. The paralegals now spend their time on the top-ranked documents rather than reading every time card and HR memo in order. That's not a luxury, that's basic efficiency. And it's available for approximately $2,000 to $5,000 for a mid-size matter on most cloud platforms.

Myth 2: AI Document Review Is a 'Black Box' Courts Won't Accept

This myth has more staying power because it once had more truth behind it. Early AI review tools were opaque, attorneys couldn't easily explain to a judge why the system flagged certain documents as relevant. But the legal community didn't just accept that opacity; it developed standards to address it. The most important development was the widespread adoption of Technology-Assisted Review, or TAR, validated through judicial opinions and professional guidelines. Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012) was the first federal case to explicitly approve predictive coding in e-discovery. Since then, dozens of courts have accepted TAR protocols, and the Sedona Conference has published detailed guidelines on validation and transparency.

Modern AI review platforms are built around explainability and auditability. In Relativity, for instance, every document that gets a relevance score can be traced back to the training decisions that produced that score. Attorneys can show opposing counsel or a judge exactly which seed documents were used, how the model was validated, and what quality-control checks were run. This is actually more defensible than keyword search, where the attorney just asserts that 'termination AND bonus' should catch everything relevant, with no statistical validation whatsoever. The black box concern has largely been answered by the industry, even if the perception hasn't caught up.

There's also a practical point that gets missed: AI-assisted review doesn't eliminate attorney judgment, it channels it more effectively. A paralegal or associate still reviews the documents the AI flags as high-priority. A senior attorney still makes privilege calls. The AI accelerates the triage step; humans make the final decisions. This is the model courts have accepted, and it mirrors how judges think about proportionality under Federal Rule of Civil Procedure 26(b)(1). Using a validated AI tool is not cutting corners, it is proportionate, defensible, and increasingly expected in large-volume matters.

Myth 3: AI Will Replace Paralegals and Junior Associates in Document Review

This fear is understandable. If AI can review 100,000 documents in hours, what does a first-year associate do? The answer is: more interesting work. The document review that AI handles best is the most mechanical kind, first-pass relevance sorting of large, uniform populations. What it cannot do is understand litigation strategy, recognize when a document is relevant in a way that wasn't anticipated in the original review protocol, catch privilege issues that require nuanced legal judgment, or communicate findings to a client. Every AI-assisted review still requires humans to set up the matter, train the model, QC the outputs, make final privilege calls, and synthesize findings into a coherent narrative. Those tasks don't disappear, they become the core of what reviewers do.

The more accurate framing is that AI is changing the skill mix required, not eliminating the need for legal professionals. Paralegals who understand how to set up a Logikcull matter, build a coding guide, train a TAR model, and interpret quality-control metrics are more valuable than those who don't, not less. The same is true for associates. Firms that have adopted AI-assisted review report that junior attorneys spend less time on rote first-pass review and more time on substantive analyzis, deposition prep, and client communication. That's a better use of expensive legal training. The professionals who should be worried are those who refuse to engage with these tools at all.

Myth vs. Reality: A Direct Comparison

What Actually Works: How Legal Professionals Use AI in Document Review Today

The professionals getting the most out of AI in e-discovery are treating it as a workflow partner, not a magic button. Here's what that looks like in practice. A litigation paralegal at a mid-size firm receives a production of 60,000 documents in a commercial lease dispute. She uploads the files to Everlaw, creates a matter, and builds a coding guide with her supervising attorney, defining what 'relevant,' 'privileged,' and 'hot' mean for this specific case. She then reviews a seed set of 500 documents herself, coding each one. The AI model learns from those decisions and begins ranking the remaining 59,500 documents. Within 48 hours, she has a prioritized queue. She works through the top-ranked documents, periodically spot-checking the low-ranked population to confirm the model isn't missing anything important.

On the generative AI side, tools like Harvey AI, CoCounsel (built on GPT-4), and Relativity's aiR, legal professionals are using AI to do something beyond ranking: they're asking it to summarize, extract, and synthesize. A contract attorney reviewing due diligence materials for an M&A deal can ask CoCounsel to identify every document that references indemnification obligations and produce a summary table. That used to take a team of reviewers a week. With a well-configured AI tool, it takes hours. The attorney still reads the flagged documents and makes the final judgment, but the needle-finding step is dramatically faster.

The teams that struggle are those who treat AI as a set-and-forget solution. They upload documents, let the model run, and assume the output is correct. It isn't, not without quality control. Every serious AI review protocol includes elusion testing (sampling the documents the AI ranked as not relevant to confirm it hasn't missed anything), precision testing (checking whether the AI's high-ranked documents are actually relevant), and periodic recalibration as reviewers discover new issue patterns. This is not technically complex, it's a structured review discipline. Any paralegal or associate who learns it becomes an asset to every litigation team they join.

Frequently Asked Questions

• Q: Do I need to tell opposing counsel I'm using AI for document review? A: In most jurisdictions, you do not have a blanket obligation to disclose the use of TAR, but you are required to describe your review methodology if asked, and courts increasingly expect transparency in ESI protocols. The Sedona Conference recommends proactive disclosure as a best practice. Check your jurisdiction's local rules and any standing orders from your assigned judge.
• Q: What if the AI misses a relevant document and it's not produced? A: This is why quality-control testing (elusion testing) is non-negotiable. No review methodology, manual or AI-assisted, guarantees 100% recall. Courts evaluate whether you used a reasonable, proportionate process, not whether you achieved perfection. Document your QC steps and you have a defensible position. Skip them and you don't.
• Q: Can I use ChatGPT or Claude to review client documents? A: Not safely without explicit data processing agreements and enterprise-tier accounts with privacy protections. ChatGPT's standard consumer version and Claude's free tier are not appropriate for confidential client materials. Purpose-built legal AI tools like Relativity aiR, Harvey, and CoCounsel have attorney-client privilege protections and data security commitments built into their contracts.
• Q: How long does it take to set up an AI-assisted review on a new matter? A: On a platform like Everlaw or Logikcull, a paralegal with basic training can have documents uploaded, processed, and a TAR model running within one to two business days for a standard production. The seed set review (coding 300 to 500 documents to train the model) typically takes four to eight hours of focused attorney or paralegal time.
• Q: What's the difference between keyword search and predictive coding? A: Keyword search retrieves documents containing specific words or phrases you define upfront, fast but rigid and prone to missing context. Predictive coding (TAR) learns from human examples what relevant documents look like, then scores the entire population by predicted relevance, slower to set up but far more accurate, especially in complex matters with varied language patterns.
• Q: Is AI e-discovery appropriate for small matters with only a few thousand documents? A: Generally, no, the setup time for TAR doesn't pay off on very small collections. Most practitioners use manual review or basic keyword search for matters under 5,000 to 10,000 documents, and reserve AI-assisted review for larger productions where the efficiency gains are substantial. Platforms like Logikcull are designed to make that threshold judgment easy by showing you upfront what processing will cost.

Key Takeaways from Part 1

1. AI e-discovery tools are accessible and affordable for small and mid-size firms, cloud platforms like Logikcull and Everlaw charge by usage, not by enterprise contract.
2. Courts have accepted AI-assisted review since 2012, and modern platforms produce auditable logs that are more defensible than traditional keyword search.
3. AI handles mechanical first-pass triage, humans still make every legal judgment, privilege call, and strategic decision. The skill mix shifts; the need for legal professionals does not disappear.
4. Quality control is not optional. Elusion testing and precision testing are what make AI-assisted review defensible, and they require human judgment, not technical expertise.
5. The biggest barrier to adoption is not cost or court acceptance, it's workflow inertia. Professionals who learn these tools now are building a durable competitive advantage.

Three Myths That Are Costing Legal Teams Time and Money

Most legal professionals approach AI-assisted e-discovery with one of three assumptions: that it's too unreliable to trust, that it's only for massive litigation handled by BigLaw firms, or that once you deploy it, it runs itself. All three beliefs lead to bad decisions, either avoiding tools that would genuinely help, or misusing them in ways that create professional risk. The reality of how AI performs in document review is more nuanced, more accessible, and more demanding of human judgment than any of these myths suggest. Understanding where the myths come from, and what the evidence actually shows, is how you go from cautious skeptic to confident practitioner.

Myth 1: AI Document Review Is Too Unreliable for Real Legal Work

This myth has a legitimate origin. Early keyword-search approaches in e-discovery were genuinely unreliable. A 2006 TREC Legal Track study found that keyword search alone achieved recall rates as low as 20%, meaning attorneys were missing up to 80% of relevant documents. That finding stuck in legal culture. But the tools being used today are categorically different. Modern Technology-Assisted Review systems, including those embedded in platforms like Relativity and Everlaw, use machine learning models trained on attorney decisions. Studies published in peer-reviewed legal technology journals consistently show TAR achieving recall rates of 75% or higher, often exceeding manual review teams working under time pressure.

The more important comparison isn't AI versus perfection, it's AI versus human reviewers working at scale. Human document review is itself error-prone. Research published in the Richmond Journal of Law and Technology found that junior associates reviewing documents under deadline pressure had inconsistent recall rates, with individual reviewers disagreeing on relevance calls up to 40% of the time. Fatigue, inconsistent training, and sheer volume all degrade human performance. AI doesn't get tired at hour seven of a 10-hour review session. It applies the same decision logic to document number 50,000 that it applied to document number one.

Click to enlarge

Myth 2: AI E-Discovery Is Only for BigLaw and Enterprise Litigation

The assumption that AI-powered document review requires a massive IT infrastructure, a six-figure software contract, and a dedicated litigation support team is outdated by at least five years. The market has shifted dramatically. Platforms like Casetext (now part of Thomson Reuters), Logikcull, and CloudNine offer subscription-based e-discovery tools with AI review features priced for small firms, solo practitioners, and in-house legal departments at mid-market companies. Logikcull, for example, has published pricing starting at $250 per matter, a number accessible to a small employment litigation practice or a regional corporate legal team handling a vendor dispute.

The use cases extend well beyond high-stakes commercial litigation. HR professionals handling internal investigations routinely deal with email archives, Slack exports, and Teams message logs that need to be reviewed for relevance and privilege. A compliance officer at a healthcare company responding to a regulatory inquiry might need to process 30,000 emails from three departing employees. A school district's general counsel facing a public records request might have 15,000 documents to sort through. None of these scenarios require a BigLaw budget. All of them benefit from AI-assisted sorting, deduplication, and concept clustering available in mid-market platforms.

The practical entry point for non-BigLaw teams is often a single matter. You upload documents to a platform like Logikcull or Everlaw, let the system deduplicate and process them, then use the AI search and concept clustering features to prioritize review. You don't need to negotiate an enterprise contract or hire a litigation support specializt. Many platforms offer self-service onboarding with video tutorials, and some, including Casetext's CARA and Thomson Reuters' CoCounsel, let you interact with documents through a natural-language interface that feels more like asking a question than running a database query.

Click to enlarge

Myth 3: AI Review Runs Itself Once You Set It Up

This is the myth most likely to create professional liability. TAR systems, particularly the continuous active learning models used in platforms like Relativity, do improve as they process more attorney decisions. But 'improving' is not the same as 'running without oversight.' Every TAR workflow requires what practitioners call a 'seed set': an initial batch of documents reviewed and coded by a qualified attorney that trains the model's baseline understanding of relevance. If your seed set is poorly coded, inconsistently applied, or drawn from an unrepresentative sample, the model learns the wrong lesson and propagates that error across tens of thousands of documents. Garbage in, garbage out, at scale.

Ongoing quality control is non-negotiable. Standard protocol in well-run TAR projects includes elusion testing, pulling random samples from the documents the AI has marked non-relevant to verify the system isn't excluding documents it should be flagging. Courts and opposing counsel have increasingly demanded transparency about TAR protocols, including sampling methodology and validation metrics. If you can't explain how you validated your AI review, you have a defensibility problem regardless of how good the technology is. The Sedona Conference's TAR Case Law Primer and Commentary provide detailed guidance on what courts expect, and 'we trusted the software' is not a sufficient answer.

What Actually Works: Building a Defensible AI Review Process

The legal teams getting the most value from AI document review share a few common practices. They treat AI as a prioritization engine, not a decision-maker. The AI's job is to surface the most likely relevant documents first, so attorneys spend their limited review time on the highest-value material. This is sometimes called a 'priority review' workflow: the platform ranks documents by predicted relevance, and reviewers work from the top down. When the marginal relevance rate drops below a defined threshold, say, fewer than 2% of documents in a batch are being marked relevant, the team has statistical confidence they've found most of what matters.

They also invest in clear coding guidelines before review begins. This sounds obvious, but it's where many projects fail. If three different reviewers have three different interpretations of what 'relevant to the contract dispute' means, the AI model will be trained on inconsistent signals and produce inconsistent results. A one-page coding guide with three to five concrete examples of relevant and non-relevant documents, agreed upon before anyone codes a single file, dramatically improves model performance. In TAR-based review, this document is sometimes called a 'relevance protocol,' and producing it forces the legal team to have a conversation about scope that benefits the entire matter, not just the e-discovery process.

Finally, effective teams document everything. Which platform did you use? What was your seed set size? How many rounds of review did the model go through? What elusion testing did you conduct, and what were the results? This documentation serves two purposes: it protects you if opposing counsel challenges your review methodology, and it creates institutional knowledge your team can apply to the next matter. Firms that treat each e-discovery project as a one-off lose the efficiency gains that come from refining a repeatable process. The teams that build internal playbooks, even simple ones, compound their advantage over time.

Practice Task: Build a TAR Readiness Checklist for a Real Matter

Frequently Asked Questions

• Q: Do I need opposing counsel's agreement to use TAR? A: Not automatically, but transparency is increasingly expected. Many courts and the Sedona Conference recommend disclosing your TAR protocol to opposing counsel proactively. Some jurisdictions and judges require it. Check your local rules and consider a meet-and-confer conversation about review methodology early in the case, it reduces challenges later.
• Q: What's the minimum document volume where AI review makes economic sense? A: Most practitioners cite 5,000–10,000 documents as the practical floor. Below that, manual review is often faster and cheaper than the setup time for a TAR workflow. Above 10,000 documents, the efficiency math shifts decisively in AI's favor. Some platforms, like Logikcull, are efficient enough at smaller volumes to be worth using for 2,000–3,000 documents when time is the constraint.
• Q: Can AI tools identify privileged documents automatically? A: AI can flag documents that contain attorney names, law firm domains, legal terminology, or patterns associated with privilege, but it cannot make final privilege determinations. Privilege review requires attorney judgment. Use AI to create a 'privilege candidate' list that a qualified reviewer then evaluates. Never rely on automated privilege filtering as your sole screen.
• Q: What happens if the AI misses a relevant document and it surfaces later? A: This is the defensibility question. If you followed a documented, reasonable TAR protocol, including elusion testing and validation, courts have generally treated missed documents as an acceptable outcome of a good-faith process, not sanctionable conduct. The risk increases dramatically if you can't show what methodology you followed. Documentation is your protection.
• Q: Can I use general AI tools like ChatGPT for document review instead of dedicated e-discovery platforms? A: General AI tools are not designed for e-discovery and should not be used as substitutes for dedicated platforms. ChatGPT and Claude are useful for drafting review protocols, summarizing individual documents you manually paste in, or generating search term lists, but they have no chain-of-custody controls, no audit logging, no deduplication, and no TAR workflow. Use purpose-built e-discovery platforms for actual review.
• Q: How do I handle cross-border data in AI document review? A: Cross-border e-discovery involving data from EU countries triggers GDPR considerations, including restrictions on transferring personal data outside the EU. Some jurisdictions have additional data sovereignty requirements. Before uploading documents containing non-US employee or customer data to any cloud-based platform, verify where the platform's servers are located and whether the vendor has appropriate data processing agreements in place. This is a legal question, not just a technical one.

Key Takeaways from Part 2

1. TAR recall rates consistently match or exceed manual review at scale, the honest comparison is AI versus real-world human review, not AI versus a theoretical perfect review.
2. AI e-discovery tools are accessible to small firms, solo practitioners, and mid-market in-house teams, matter-based pricing on platforms like Logikcull starts at a few hundred dollars.
3. AI review requires active human oversight: a quality seed set, a written relevance protocol, ongoing elusion testing, and documented validation methodology.
4. Courts have approved TAR since 2012, but they expect transparency about methodology, 'we trusted the software' is not a defensible answer.
5. General AI tools like ChatGPT are useful for drafting protocols and summarizing individual documents, but dedicated e-discovery platforms are required for actual document review workflows.
6. A one-page relevance protocol written before review begins improves both human and AI consistency, and gives you documentation to produce if your methodology is challenged.

What Most Legal Professionals Get Wrong About AI and E-Discovery

Most legal professionals hold three beliefs about AI in e-discovery that sound reasonable but fall apart under scrutiny. First: that AI document review is too risky for high-stakes litigation because it misses too much. Second: that only large firms with big IT budgets can actually use these tools. Third: that once you deploy AI review, you can hand off the work and stop worrying. All three beliefs lead to bad decisions, either avoiding tools that would genuinely help, or trusting them in ways that create real liability. Each myth has a corrected version that changes how you should think about AI in your document review workflow starting this week.

Myth 1: AI Review Is Too Risky Because It Misses Documents

The fear is understandable. A missed document in litigation can be catastrophic. So when attorneys hear 'AI review,' they imagine a black box making autonomous decisions about what gets produced and what gets buried. That fear is based on a real risk, but it's the wrong risk. The actual danger isn't that AI misses more than humans. Multiple studies have shown the opposite. A landmark TREC Legal Track study found that human reviewers achieve recall rates around 60-70%, while well-implemented technology-assisted review (TAR) consistently reaches 75% or higher. Human review at scale is exhaustively expensive and surprisingly inconsistent.

Courts have repeatedly validated TAR when it is properly supervised and documented. The Da Silva Moore case in 2012 was the first federal court to approve predictive coding in e-discovery. Since then, dozens of courts have followed. Judges aren't demanding human-only review, they're demanding defensible process. That means logging your training decisions, validating recall rates, and being able to explain your methodology. The risk isn't using AI. The risk is using AI without a documented, supervised protocol.

The better mental model: think of AI review as a highly consistent junior associate who flags documents based on patterns you trained it to recognize. It doesn't get tired at hour nine of a document review marathon. It applies the same standard to document 1 and document 80,000. Your job is to train it well, audit its output, and document the process. That is defensible. Random fatigue-driven human review at scale often is not.

Myth 2: AI E-Discovery Tools Are Only for BigLaw

The assumption is that enterprise e-discovery platforms like Relativity or Everlaw cost so much that solo practitioners and small firms are priced out entirely. That was largely true a decade ago. It is not true now. Everlaw offers subscription pricing that scales down to smaller matters. Logikcull, built specifically for small and mid-size firms, charges per gigabyte with no setup fees and no long-term contracts. Casetext's CoCounsel, now part of Thomson Reuters, is accessible to individual attorneys at a monthly subscription rate comparable to a single hour of associate billing.

Even general-purpose AI tools like Claude Pro or ChatGPT Plus can help with document-related tasks that don't require a full e-discovery platform: summarizing deposition transcripts, drafting privilege log descriptions, identifying key themes across a set of contracts you paste in manually. These aren't TAR platforms, but they're real productivity tools available for $20-$25 per month. For a solo practitioner handling a commercial dispute with a few hundred documents, this is a legitimate starting point.

The realiztic picture: your tool choice should match your matter size and budget. A 500-document breach of contract case does not need Relativity. A 2-million-document antitrust investigation probably does. The market now offers genuine options across that entire spectrum. The barrier isn't cost, it's knowing which tier of tool fits which type of matter.

Myth 3: AI Review Runs Itself Once You Set It Up

This is the most operationally dangerous myth. Attorneys who believe AI review is a 'set it and forget it' process either end up with defensibility problems or miss the point entirely. TAR requires active attorney involvement throughout. You train the model by reviewing seed documents and coding them. You validate results by sampling across the document population. You monitor for drift, situations where the model's predictions diverge from your evolving understanding of the case. This isn't a one-time setup. It's an iterative loop that runs until you've hit your recall target and can document why you stopped.

The good news: the attorney hours required are far fewer than linear human review. You're guiding a process rather than reading every document yourself. But guidance requires judgment, and judgment requires engagement. Courts that have sanctioned parties over AI-assisted review failures have generally found not that the AI was bad, but that no qualified person was actually steering it. AI in e-discovery is a force multiplier for attorney judgment, not a replacement for it.

Myth vs. Reality: Side by Side

What Actually Works in Practice

Effective AI-assisted document review follows a consistent pattern across firm sizes and matter types. It starts with a clear scope decision: what custodians, date ranges, and document types are in play. Before any AI tool touches the data, you establish a written review protocol that specifies your training methodology, your validation approach, and your acceptable recall threshold. Most practitioners target 75% recall or higher, though the right number depends on matter stakes. This upfront work takes a few hours and is the single most important thing you can do to make the entire process defensible.

The training phase is where attorney judgment shapes everything. You review a seed set, typically 500 to 2,000 documents, coding each as relevant or not relevant. The platform learns from your decisions. You then run a validation sample to test how well the model is performing before applying it to the full population. If the recall rate is below your threshold, you do another training round. This iterative loop is not a flaw, it's the mechanism that makes TAR more accurate than linear review. Skipping validation is where firms get into trouble.

Once production is complete, your documentation package becomes your shield. Keep records of every training decision, every validation round, your final recall calculation, and the name and version of the platform you used. If opposing counsel or a court questions your process, you hand them the log. Firms that do this consistently report not only better defensibility but faster discovery cycles, matters that used to take months of contract review often close in weeks. The technology works. The discipline around it is what separates good outcomes from bad ones.

Click to enlarge

Frequently Asked Questions

• Can I use ChatGPT or Claude for actual client document review? For low-sensitivity, non-confidential documents, general AI tools can help with summarization and issue-spotting. For actual litigation production involving confidential client files, use a platform with a signed BAA or data processing agreement. Everlaw, Logikcull, and Relativity all offer these. Never paste privileged client data into a consumer AI tool without confirming your firm's data policy.
• What recall rate should I target for TAR? Most practitioners and courts accept 75% recall as a reasonable threshold for proportionate discovery. Higher-stakes matters, criminal defense, large antitrust, securities fraud, often warrant 80-85%. Document your rationale for whatever threshold you choose.
• How do I handle privilege review with AI? Most enterprise e-discovery platforms include privilege tagging workflows. AI can flag likely privileged documents based on attorney names, firm domains, and key phrases, but a qualified attorney must make the final privilege call on each document. AI speeds the identification; humans make the determination.
• What if opposing counsel challenges my TAR process? A documented protocol is your defense. Courts consistently uphold TAR when parties can demonstrate a reasonable, supervised methodology. Prepare a short written description of your process before production, this is standard practice in matters where TAR is disclosed.
• Do I have to disclose to opposing counsel that I used AI? Rules vary by jurisdiction and are evolving rapidly. Some courts require disclosure; others don't. Check your local rules and any standing orders from your assigned judge. When in doubt, proactive disclosure of your methodology, without necessarily specifying every tool, is generally the safer posture.
• What's the difference between TAR 1.0 and TAR 2.0? TAR 1.0 (also called simple active learning) uses a fixed seed set reviewed upfront, then applies the model to the full population. TAR 2.0 (continuous active learning) updates the model continuously as reviewers code documents during review. TAR 2.0 is generally considered more accurate and is now the default in most modern platforms.

Key Takeaways

• AI-assisted review (TAR) has demonstrated recall rates equal to or better than human linear review, the risk of AI isn't missing documents, it's failing to document your process.
• Accessible, affordable tools exist at every budget level: Logikcull and Everlaw for matters requiring a full platform, Claude Pro or ChatGPT Plus for early case assessment and document summarization.
• TAR is not autonomous, it requires iterative attorney training, validation sampling, and a written protocol to be defensible in court.
• Courts have approved TAR in hundreds of cases since 2012; judicial acceptance is not the obstacle it once was.
• Your documentation package, training logs, validation results, recall calculations, is your liability shield if your process is ever challenged.
• Match your tool to your matter: a 500-document commercial dispute and a 2-million-document antitrust investigation need different solutions, and the market now provides both.