What Is E-Discovery AI? A Guide for Law Firms

Traditional e-discovery is linear. Collect data, process it into a review platform, assign reviewers, export what is relevant. Every step requires human decisions on every document. At scale, that model is expensive and slow.

E-discovery AI breaks that linearity at several points.

Predictive coding (technology-assisted review): A machine learning classifier is trained on a seed set of documents reviewed by a senior lawyer. The model learns relevance patterns and applies them across the full collection. Platforms like Relativity and Everlaw have offered this for years. The result: reviewers focus on borderline documents rather than obvious non-responsive material.

Generative AI summarisation and issue coding: Newer workflows use large language models to read a document and produce a structured output. Issue code: product liability. Key facts: defendant knew of defect by March 2021. Privilege flag: communication with in-house counsel. Relativity's aiR suite, priced around $18-25 per GB per month for hosted data (AI Vortex, 2026), does this at the document level across entire collections.

Privilege detection: AI models trained on privilege patterns surface potentially privileged communications before they leave the review database. This is now table stakes. If a platform does not include automated privilege logging, look elsewhere.

Conversation threading and analytics: Tools like Everlaw analyse email threads, identify key actors, and map communication patterns. This is particularly useful in antitrust and securities matters where the relationship between custodians matters as much as individual documents.

These are not black-box processes. The best implementations link every AI output to the source document, with a full audit trail that survives motion practice.

Legal teams spent years debating whether AI-assisted review was accurate. That debate is over. The real question now is whether your AI workflow is defensible in court.

EDRM's 2026 guidance on AI privilege workflows is direct: design for defensibility from day one (EDRM, 2026). That means documenting how the AI was trained, what seed sets were used, what validation protocol confirmed accuracy, and who reviewed the outputs before production. A workflow you cannot explain to opposing counsel or a judge is a liability.

Small firms sometimes skip these steps because the tooling feels complicated. That is a mistake with real consequences. AI can cut review costs by up to 85% (Hintyr, 2026), but those savings disappear if a privilege challenge results in sanctions or a clawback dispute.

Three things every firm must get right:

1. Transparency: Document every AI decision point. Who trained the model? What was the recall target? What human review validated the outputs?
2. Auditability: Every document flagged, coded, or excluded by AI should carry a log entry. Platforms without audit trails are not fit for contested matters.
3. Lawyer-in-the-loop controls: AI should accelerate the reviewer, not replace the reviewing lawyer's judgment. Any platform that acts autonomously on privilege or production decisions without human sign-off creates professional responsibility exposure.

This is not overcaution. It is the standard the courts are moving toward.

Four platforms dominate the e-discovery AI market in 2026: Relativity, Everlaw, DISCO, and Logikcull. Each takes a different approach.

Relativity's aiR suite adds generative AI on top of its existing predictive coding infrastructure. Document summarisation, privilege detection, and issue coding run in the same hosted environment at $18-25 per GB per month (AI Vortex, 2026). It is the default choice for large matters and BigLaw.

Everlaw targets government and in-house teams with predictive coding and conversation analysis. Pricing runs $20-30 per GB per month, with competitive flexibility for volume deals (AI Vortex, 2026). Its strength is timeline visualisation and custodian network mapping.

DISCO emphasises AI-driven document prioritisation and automated privilege logging, with per-matter pricing. Its workflow approach suits firms that run many small to mid-size matters rather than single large ones.

Logikcull positions itself as the budget option for small firms, with cloud-based deployment and transparent pricing. For firms doing occasional e-discovery, it removes the infrastructure overhead entirely.

All four now claim over 90% accuracy on relevance classification (AI Vortex, 2026). Verify that claim against your document types before accepting it. A model trained on English commercial contracts will not perform identically on Mandarin-language communications or highly technical patent documents.

For firms using Relativity alternatives for law firms, the key question is whether the alternative platform matches Relativity's audit trail depth and privilege workflow documentation.

E-discovery AI is scoped to a specific problem: finding and reviewing ESI for production in litigation. It is not built to make that information usable after the matter closes.

That is a significant gap. A firm runs a product liability matter, processes 800,000 documents, identifies 12,000 relevant, and produces 4,000. The matter closes. That work, and the intelligence embedded in it, is functionally lost. No one builds a knowledge graph from a Relativity review database. The next similar matter starts from scratch.

This is where a platform like Casero addresses a different but adjacent problem. Casero is a UK-based legal intelligence layer that connects emails, documents, and case management systems into living, case-level knowledge graphs. Where e-discovery AI is optimised for volume document review under litigation rules, Casero is optimised for making the intelligence inside those documents continuously accessible and reusable across matters.

Casero's entity extraction automatically identifies people, organisations, dates, events, and obligations from documents and emails, then maps how they relate within a knowledge graph. Every fact traces back to its source document. Its similar cases matching surfaces past matters based on legislation, factual circumstances, and case classification, with multi-dimensional scoring showing exactly why each case matched. That is the institutional memory problem that e-discovery AI does not touch.

The two tools operate at different points in the matter lifecycle. E-discovery AI handles the review. A platform like Casero handles what you do with everything you know once the review is done. See what case intelligence means for law firms for more on that distinction.

The most common failure is treating AI as a one-click process. Upload documents, let the model run, produce. That sequence is missing two critical steps: validation and documentation.

Validation means running a statistically valid sample review after the AI coding pass to confirm recall at an acceptable rate. Most courts expect to see this. Most firms using budget tools skip it.

Documentation means writing down what you did before you get a challenge, not in response to one. A privilege log dispute is a bad time to reconstruct how your AI model made its decisions.

A second common failure is using a single model for heterogeneous document sets. A model trained on English emails performs differently on scanned PDFs, handwritten notes, or technical schematics. Segment your collection and validate each segment separately.

Third: ignoring data privacy obligations during collection. The data going into an e-discovery AI platform often includes personal data of third parties, employees, and clients. In the UK and EU, that raises GDPR questions about lawful basis, data minimisation, and cross-border transfers. Check legal AI data privacy guidance before sending client data to a US-hosted review platform.

Fourth: choosing a platform based on accuracy claims alone. Accuracy percentages are measured against controlled test sets. Ask for the validation protocol, not just the headline number. Ask which document types were in the test set. Ask what happens when the model flags a document as non-privileged and it is later found to be privileged.

Small firms have historically outsourced e-discovery or avoided cases with significant ESI. Both responses are increasingly wrong.

Cloud-based platforms like Logikcull have made per-matter pricing accessible without infrastructure investment. AI review tools can cut document review costs by up to 85% for small firms (Hintyr, 2026). A two-lawyer team can now run a 50,000-document review that would have required a contract review vendor five years ago.

The practical starting point: use a platform with transparent per-GB pricing and no minimum commitment. Run a single matter through it. Document the validation steps. Compare the cost to what you paid for manual review on a comparable matter last year.

That comparison will tell you whether e-discovery AI makes sense for your practice. The firms waiting for the technology to mature have already waited too long. It is operational now.

Casero's pilot tier is free, and pilot partners receive full Professional-tier access with no commitment required. For small UK firms that want to connect e-discovery outputs into reusable case intelligence without building a data infrastructure team, that is a practical starting point. See legal AI knowledge management guidance for lawyers for more on building that layer.