AI in Legal E-Discovery 2026: Relativity aiR, DISCO, Everlaw, and TAR After CAL

E-discovery has been quietly running production AI longer than almost any legal workflow — predictive coding has been court-accepted since the Da Silva Moore opinion in 2012, and Technology-Assisted Review (TAR) has been a discipline in major matters for more than a decade. The generative AI wave has accelerated change in a category that was already mature, and the 2026 picture is one of meaningful capability layered on top of a workflow with well-developed defensibility expectations.

This is the state of e-discovery AI in 2026, the platform incumbents, what TAR Continuous Active Learning evolved into, and where generative review actually fits.

The platform incumbents#

Relativity is the dominant e-discovery platform globally, and Relativity aiR is the umbrella for their generative AI features — review, privilege, and case-strategy assistance. RelativityOne (the cloud platform) is where most of the AI capability ships and is the deployment shape most large matters use.

DISCO built a more modern, cloud-native review platform with strong AI capabilities from earlier in their history; they have continued to invest aggressively in generative review and have a notable customer base in mid-market litigation and in corporate legal departments.

Everlaw is the third major platform with deep AI integration and has been particularly aggressive on the generative side with their “Everlaw AI Assistant” features for document analysis, deposition prep, and case narrative.

Reveal (which acquired Logikcull in 2023 and combined with the Reveal e-discovery and investigations platform) operates across the e-discovery and compliance investigations space with AI built across the platform.

Nuix continues to be strong in early case assessment, investigations, and forensic review with AI layered into a different shape of the workflow.

AI document review in e-discovery

TAR and Continuous Active Learning#

Technology-Assisted Review evolved through several generations. TAR 1.0 was simple predictive coding — review a seed set, train a model, classify the remainder. TAR 2.0 introduced Continuous Active Learning (CAL), which removes the rigid seed-train-classify cycle in favor of a model that updates continuously as the review team marks documents. CAL has been the dominant pattern for the last decade and remains so in 2026 for the core relevance review task.

The reason CAL won is operational. It does not require a careful sampling protocol up front. It improves continuously as reviewers work. It handles concept drift across the review. It is straightforward to defend in front of a court because the methodology is well-understood and there is substantial case law endorsing it.

What CAL does not do well is the parts of review that are not binary relevance — privilege, confidentiality, issue tagging, key document identification. Those have always required additional models or human judgment.

Generative AI on top of TAR#

The 2023 to 2025 wave layered generative AI on top of the existing TAR workflow. The use cases that have settled into production:

Summarization of documents and document families — instead of (or in addition to) a reviewer marking relevance, the LLM produces a one-paragraph summary of what the document is about. This speeds review and improves consistency.

Privilege screening — LLM-based pre-screening of documents that may be attorney-client privileged or work product, which then go to a focused human privilege review. This has been one of the more controversial use cases because privilege exposure is severe if a model produces false negatives.

Deposition and witness preparation — surfacing the most relevant documents for a witness, generating preliminary outlines, summarizing what the witness’s documents suggest about their likely testimony.

Case-narrative drafting — LLM-assisted drafting of section summaries, exhibit lists, and case-theory documents grounded in the document set.

Cross-document factual extraction — given a question, find documents that bear on it, extract the relevant facts, and produce a citation-grounded answer.

The Mata v. Avianca lesson and defensibility#

E-discovery practitioners have arguably internalized the Mata v. Avianca lesson more deeply than other legal segments because defensibility has always been the discipline of the practice. The lesson — that LLMs hallucinate citations and that lawyers remain professionally responsible for the accuracy of their filings — translates directly into the e-discovery deployment shape.

Every credible generative e-discovery product in 2026 ships with retrieval-grounded generation (the model must cite source documents), citation verification, and prominent disclaimers about human review. The audit log preserves which documents the model considered and which it cited. The model output is treated as an assist to the reviewer rather than a substitute for the reviewer’s professional judgement.

The defensibility expectation in court has caught up. Several 2024 and 2025 opinions have addressed the use of generative AI in discovery and the broad position has been that the methodology is acceptable when properly documented, validated, and supervised — much the same posture courts took toward TAR more than a decade ago.

Generative review summarization

Privilege and the high-stakes screening problem#

Privilege review is the single most expensive part of many large e-discovery projects because the cost of a false negative — producing a privileged document — can be matter-defining. Several vendors and major firms have invested heavily in privilege-specific models.

The 2026 state of the art is a combination of name and email-domain rule-based identification (which captures most attorney communications cleanly), entity-relationship extraction across the document set, and LLM-based contextual review of borderline cases. The human privilege review still happens; the AI narrows the population to review and surfaces the rationale.

The honest defensibility position is that privilege screening by AI is acceptable when paired with human review of the AI-flagged set and a defensible sampling protocol over the AI-cleared set. The big-firm patterns documented in 2024 and 2025 broadly track this shape.

Cross-border and data sovereignty constraints#

Large multinational matters increasingly run into data sovereignty constraints — German employee data subject to works council and BetrVG protections; EU data subject to GDPR and the EU AI Act; Chinese data subject to PIPL and the Data Security Law; cross-border transfer rules that limit what data can leave a jurisdiction even for review.

The 2026 vendor response has been in-country review environments (Relativity, DISCO, and Reveal all offer EU and APAC-resident deployments), federation patterns that keep data in place while running models against it, and protocols for ensuring the model itself does not transmit data inappropriately. The operational complexity of large cross-border matters has grown, not shrunk, with AI.

Investigations and regulatory work#

E-discovery technology has expanded into adjacent categories — internal investigations, regulatory inquiries, M&A due diligence — that share document-heavy review at their core. Reveal, Nuix, and the platform incumbents have leaned into this expansion. The AI capabilities transfer cleanly; the workflow and stakeholder shapes differ.

Where the gaps remain#

Audio and video discovery (depositions, recorded calls, recorded meetings) remains harder than text. Transcription is fine; making transcription as searchable and reviewable as text-native documents takes work. Several vendors are pushing in this direction and the gap is closing.

Hyper-large international matters with mixed-language document sets still strain the platform incumbents. The non-English-language model quality varies by language pair and is meaningfully worse for some Asian and Middle Eastern languages than for European languages.

The economics of small matters — under USD 200,000 in legal fees — still do not fully support AI-driven review for most firms. The fixed setup cost of a sophisticated AI deployment is too high for the matter size.

Where pdpspectra fits#

Our AI integration practice helps legal departments and litigation support providers build e-discovery and investigations workflows — retrieval over document populations, summarization pipelines, evaluation, and the audit and governance scaffolding that defensibility requires.

E-discovery AI is the most-mature production legal AI and is still evolving. Talk to our team about your discovery and investigations AI roadmap.