We respect attorney-client confidentiality. No tracking pixels in our emails.
A practical guide to using AI for eDiscovery, covering TAR vs CAL, platform selection, cost benchmarks, and validating your methodology for court.
Move from this guide to practical legal AI tool shortlists and comparison pages.
Opposing counsel just produced 2.3 million documents. Your discovery deadline is 14 days. A manual review team would take 60 days and cost $400,000. This is the math that makes AI eDiscovery not optional.
The question is not whether to use AI for document review. The question is which platform, which methodology, and how to document it well enough that opposing counsel cannot challenge your process.
This is our practical guide to using AI for eDiscovery in 2026, written for litigation partners, legal ops managers, and eDiscovery professionals managing large document productions.
LawyerAI built this guide. We earn no affiliate revenue from these tools.
Here are the 4 rules we set for ourselves before writing this:
We re-review this list every quarter.
Short answer: Relativity aiR fits Am Law firms managing large, complex cases. Everlaw fits mid-size litigation with strong collaboration needs. Logikcull fits self-serve small cases where simplicity and predictable pricing matter more than advanced AI features.
The term "AI eDiscovery" covers several distinct technologies. Conflating them leads to poor tool selection and unrealistic expectations.
Technology-Assisted Review (TAR) refers broadly to machine learning approaches that help prioritize documents for human review. Within TAR, there are two main variants:
Generative AI layers are newer. Tools like Relativity aiR for Review and DISCO's generative AI features use large language models to synthesize document themes, draft privilege logs, identify hot documents, and answer natural-language questions about a document set. These are useful for analysis and synthesis, but they introduce hallucination risk. A generative AI summary of a document set should be treated as a starting point for attorney analysis, not a final product.
Understanding which technology you are using matters for court. If you are relying on TAR to establish the thoroughness of your review, you need to be able to explain the methodology, the training process, and the validation metrics. Courts have become increasingly sophisticated about eDiscovery methodologies.
We score AI tools across five dimensions, each rated 1-5. See /blog/how-we-score-legal-ai-tools for full methodology.
| Tool | Category | Starting Price | Best For | 5D Score |
|---|---|---|---|---|
| Everlaw | Cloud eDiscovery | $25-45/GB + $250/seat/month (vendor-reported) | Mid-size cases, strong collaboration | 4.1/5 |
| Relativity AI | Enterprise eDiscovery | $50-120/GB/month (vendor-reported) | Large enterprise cases | 4.3/5 |
| DISCO eDiscovery | AI-native eDiscovery | $35-50/GB (vendor-reported) | Mid-size cases, cleaner interface |
Before selecting a platform, know which TAR methodology you intend to use. The choice affects both the platform you select and how you document your methodology.
TAR 1.0 works best when the document population is relatively homogeneous and your legal team can invest time upfront in a rigorous seed set review. The seed set should include a statistically meaningful sample of both responsive and non-responsive documents. A poorly constructed seed set produces a poorly calibrated model.
TAR 2.0 / CAL is generally preferable for complex, multi-topic litigation where the document population is diverse and the legal theory is still developing. The continuous learning loop means the model improves as you learn more about the case. Most modern platforms — including Everlaw, Relativity, and DISCO — use CAL or a hybrid approach.
The practical distinction: if you are running a focused single-issue employment dispute with 500,000 documents, TAR 1.0 with a careful seed set may be sufficient. If you are running a complex antitrust matter with 5 million documents across 15 custodians and 4 legal theories, CAL is the appropriate methodology.
Before TAR can run, documents must be processed: converted to text, de-duplicated, de-NISTed (removing known system files), and metadata extracted. This processing step is where cost per GB calculations become most relevant.
Processing costs vary by platform and by data complexity. Email with attachments is more expensive to process than clean text files. Encrypted files, unusual formats, and large media files add cost and time.
Get a processing estimate before committing to a platform for a specific matter. Most platforms offer project-specific pricing in addition to enterprise licensing. The per-GB rates quoted above are ranges; your actual cost depends on data type and volume.
Data residency is a critical consideration for matters involving international parties. Know where your data will be processed and stored. EU-based documents may require EU-based processing under GDPR. Government matters may require FedRAMP-authorized infrastructure.
TAR is more effective — and less expensive — when applied to a culled document population rather than the raw data dump. Keyword culling reduces the document population before TAR begins.
The approach: develop a keyword list in consultation with your legal team that captures the most likely relevant custodians, time periods, and topics. Apply the keywords and review only the documents that hit on at least one keyword. This typically reduces the document population by 60-80% before TAR begins.
The risk: keyword culling can miss relevant documents that do not use the expected terminology. This is a known limitation. In practice, keyword culling followed by TAR on the culled population — with a statistical sample drawn from the non-responsive set to validate recall — is a defensible and widely accepted methodology.
The RAG retrieval-augmented generation approach used by some newer platforms can reduce reliance on keyword culling by using semantic search to find conceptually relevant documents that do not contain the exact search terms. DISCO and Relativity both offer semantic search capabilities alongside traditional keyword search.
Once your document population is culled and uploaded, TAR begins. The workflow varies by platform, but the core loop is the same:
Quality control benchmarks: establish upfront what "good enough" means. The standard metric is elusion rate — the percentage of documents in the "non-responsive" bucket that are actually responsive. An elusion rate below 1-2% is generally considered defensible, but the appropriate threshold depends on the matter. Consult with your litigation team and, if necessary, agree on a protocol with opposing counsel.
Track precision as well as recall. High recall (finding most responsive documents) at low precision (tagging many non-responsive documents as responsive) inflates your review burden. The goal is to optimize both.
Modern platforms layer generative AI features on top of TAR. These features are useful for:
These features use large language models and carry hallucination risk. Relativity aiR for Review can summarize documents and identify themes, but attorneys should treat AI-generated summaries as starting points, not finished work product. A privilege assertion made on the basis of an AI-generated privilege log entry that mischaracterizes the document content creates exposure.
The appropriate use: AI-generated summaries and privilege log drafts as first-pass inputs that attorneys review and verify, not as final outputs.
Before producing documents, validate that your TAR methodology found what it was supposed to find.
The standard validation approach: draw a statistically significant random sample from the documents the TAR model classified as non-responsive (the documents you are not producing). Have attorneys review this sample and identify any responsive documents. The percentage of responsive documents in this sample is your elusion rate.
A common sample size for validation: 95% confidence interval with a 2% margin of error requires approximately 2,400 documents for a large population. Statistical sampling calculators are widely available.
Document the validation. You will need this documentation if opposing counsel challenges your TAR methodology.
Courts increasingly require parties to disclose their eDiscovery methodology, particularly when TAR is used. Best practice is to address this in a Rule 26(f) conference or ESI protocol stipulation at the beginning of the case, rather than defending your methodology after production is challenged.
The documentation should include:
Platforms that generate this documentation automatically — Relativity, Everlaw, and DISCO all have reporting features — save significant attorney time in the event of a methodology challenge.
What works: Cloud-native platform with a clean interface that litigation teams find easier to learn than Relativity. Strong collaboration features — multiple reviewers can work simultaneously with real-time updates. CAL implementation is solid. Storybuilder feature is genuinely useful for timeline development and deposition prep. Good for mid-size firms and boutique litigation practices that do not have dedicated eDiscovery staff.
Real limitations: At $25-45/GB for processing plus $250/seat/month (vendor-reported), costs add up quickly for large matters. Not the right choice for cases above 5 million documents where Relativity's enterprise infrastructure becomes relevant. Customer support is good but not the 24/7 enterprise support that large firms expect.
What works: The market-leading platform for enterprise eDiscovery. Handles document populations that would overwhelm other platforms. The aiR for Review generative AI layer adds genuine synthesis capability for complex matters. FedRAMP authorization available for government matters. Deep integrations with other litigation tools. The Relativity ecosystem (RelativityOne, Relativity Server) gives administrators significant control over configuration.
Real limitations: Requires a dedicated Relativity administrator — either in-house or through a managed services provider. Cost at $50-120/GB/month (vendor-reported) is the highest in this comparison. The learning curve for non-expert users is steep. Smaller firms without eDiscovery infrastructure should look at Everlaw or DISCO before committing to Relativity. See Everlaw vs Relativity AI for a detailed comparison.
What works: AI-native from the ground up, which shows in the interface and in features like semantic search and AI-generated document summaries. Cleaner interface than Relativity, with less configuration overhead. $35-50/GB (vendor-reported) is positioned between Everlaw and Relativity. DISCO Cecilia AI provides natural-language document querying.
Real limitations: Smaller market share than Relativity and Everlaw means less third-party support ecosystem (fewer managed review providers with DISCO expertise). For very large matters where Relativity's processing infrastructure is needed, DISCO may not be the right choice.
What works: Self-serve platform that does not require a dedicated eDiscovery professional to operate. Flat $250/GB pricing is predictable and easy to budget for smaller matters. Good for in-house legal teams and small litigation firms that occasionally need eDiscovery without an enterprise contract. No minimum commitment.
Real limitations: Acquired by Reveal, which introduces some uncertainty about platform roadmap. At $250/GB, it is significantly more expensive per GB than enterprise platforms for large matters — this pricing model is designed for low-volume users. Not appropriate for matters above 1 million documents where the cost becomes prohibitive and the platform's capabilities hit limits.
What works: Strong government and FedRAMP-authorized eDiscovery. Appropriate for law firms and agencies handling government matters with strict data sovereignty requirements. CJIS-compliant features for criminal justice matters.
Real limitations: Limited adoption in the private sector means fewer third-party practitioners with Casepoint expertise. Pricing not published; requires a sales conversation for any estimate. Private sector firms without government clients are better served by other platforms.
If you are an Am Law firm managing cases with more than 5 million documents → Relativity AI. Enterprise infrastructure, deepest ecosystem, FedRAMP option.
If you are managing a mid-size case (500,000 to 5 million documents) → Everlaw. Strong collaboration features, accessible interface, competitive pricing for this range.
If you are running a small self-serve case and need predictable pricing → Logikcull. Flat per-GB pricing, no minimum, no dedicated admin required.
If your matter requires FedRAMP or government data handling → Casepoint or Relativity (FedRAMP edition).
If you are an AI-native shop prioritizing semantic search and modern interface → DISCO eDiscovery.
What is TAR and when should I use it? TAR (technology-assisted review) refers to machine learning approaches that prioritize documents for review based on relevance predictions. Use TAR when your document population exceeds 100,000 documents and manual review would be cost-prohibitive or time-prohibitive. For populations below 100,000 documents, the setup overhead for TAR may not justify the savings over a well-managed keyword review. Courts have accepted TAR as a valid methodology; document your approach carefully.
How much does AI eDiscovery cost per GB? Vendor-reported ranges as of 2026: Everlaw $25-45/GB (plus seat licensing), Relativity $50-120/GB/month, DISCO $35-50/GB, Logikcull $250/GB (flat). These are processing costs; review costs (attorney time) are separate. For large matters, negotiate custom pricing — per-GB rates are typically lower for high-volume commitments. Always get a project-specific estimate before starting.
Can AI replace a document reviewer? Not entirely. TAR can reduce the number of documents that require human review by ranking documents by likely relevance and concentrating review on the most likely responsive documents. But attorney eyes are still required for privilege determinations, relevance calls on close questions, and the validation sample that confirms your methodology found what it should. The attorney signs the production certification; the tool does not.
What is the difference between Everlaw and Relativity? Everlaw is cloud-native, easier to learn, and well-suited to mid-size matters with collaborative review teams. Relativity is the enterprise standard for large Am Law matters, with deeper configurability, a larger third-party ecosystem, and more processing capacity. Relativity requires dedicated administrative expertise that Everlaw does not. For a detailed comparison, see Everlaw vs Relativity AI.
How do I validate my TAR methodology for court? The standard approach: draw a statistically significant random sample from the documents TAR classified as non-responsive, have attorneys review that sample, and calculate the elusion rate (percentage of responsive documents in the non-responsive bucket). Document sample size, confidence interval, and elusion rate. Disclose your methodology in the Rule 26(f) conference or ESI protocol. Most modern platforms generate validation reports automatically. Opposing counsel is entitled to understand your methodology; courts have increasingly required disclosure.
LawyerAI evaluations are independent. We do not accept payment that influences our editorial scores. Featured placements are clearly labeled and do not affect our 5-dimension methodology (Accuracy / Speed / Usability / Value / Security). We re-review tools every 6 months.
If you believe any information is inaccurate, contact editor@lawyerai.directory.
| 4.0/5 |
| Logikcull | Self-serve eDiscovery | $250/GB | Small cases, self-serve | 3.6/5 |
| Casepoint | Government eDiscovery | Not published | Government/FedRAMP | 3.8/5 |
