Modern E-Discovery: From TAR to Generative AI

E-discovery is the highest-leverage place to deploy AI in litigation. A single mid-size commercial case can involve hundreds of thousands of documents. Manual linear review is now a malpractice question, not a strategy choice. This lesson covers the modern AI-enabled review workflow, the difference between TAR and generative review, and how to defend your methodology when opposing counsel pushes back.

This is an advanced lesson. We assume you know what an ESI protocol is and that you have at least supervised a review project before.

What You'll Learn

• How TAR 1.0, TAR 2.0, and generative AI review differ in practice
• The end-to-end modern review workflow used by the leading platforms in 2026
• How to defend your AI methodology under FRCP Rule 26
• How to set realistic recall, precision, and budget targets

TAR, TAR 2.0, and Generative AI Review

Three categories of technology now coexist in serious e-discovery work.

TAR 1.0 — Predictive coding. Older approach. A subject matter expert reviews a control set, the system trains on those decisions, and the model is then applied to the rest of the corpus. Strengths: well understood by courts, defensible, predictable cost. Weaknesses: brittle to changes in case theory, requires careful seed selection.

TAR 2.0 — Continuous active learning (CAL). Reviewers continuously code documents; the system continuously reprioritizes the queue toward responsive material. Strengths: faster to start, adapts to evolving case theory, better recall in most matters. Weaknesses: harder to define when review is "complete."

Generative AI review. Large language models read documents and provide reasoning. They can code documents for issue tags, summarize them, extract entities and dates, and identify privilege candidates. Strengths: explains its decisions, handles non-keyword concepts, drafts privilege logs and chronologies. Weaknesses: hallucination risk, much higher cost per document, slower throughput than TAR.

In 2026, most large reviews use a combination. TAR 2.0 prioritizes the corpus and culls obvious junk. Generative AI handles privilege, issue coding, and synthesis on the prioritized set. The two are complements, not substitutes.

The Modern Review Workflow

Here is the standard 2026 workflow at firms using Everlaw, Relativity with aiR for Review, or Reveal.

Step 1 — Collection and processing. Custodial and non-custodial data is collected. Files are processed, deduplicated, threaded, and OCR'd. AI is mostly absent here — this is plumbing.

Step 2 — Early case assessment (ECA). AI clustering and entity extraction surface the most important custodians, dates, and topics in days rather than weeks. Use this output to refine the search scope and negotiate ESI protocols.

Step 3 — TAR 2.0 prioritization. Reviewers start coding. The system continuously reorders the queue. Within a few thousand documents, the model has learned enough to push responsive material to the top.

Step 4 — Generative AI issue coding and summarization. Once prioritized, generative AI tools tag documents against your issue list, extract key facts, draft chronologies, and identify likely privilege.

Step 5 — Human privilege QC. No 2026 court accepts pure AI privilege coding. Use AI to surface candidates and produce a draft log. Humans review the candidates and sign the final log.

Step 6 — Production. AI assists in redaction, Bates numbering, and load file generation.

The big change versus 2022 workflows: steps 2, 4, and 5 used to be human-bottleneck stages and are now AI-leveraged. A team that previously needed twenty contract reviewers for sixty days now often needs four contract reviewers for ten days, plus heavier supervision by senior associates.

Defending Your Methodology

When opposing counsel challenges your AI methodology, you need three things ready.

1. A documented methodology. Write it down before review starts. Include the platforms, the prompting approach for generative AI coding, the recall and precision targets, the validation procedure, and the role of human review at each stage.

2. Statistical validation. Run a control set or systematic random sample at the end of review. Report recall and precision. Targets vary by matter, but recall of 75% and precision of around 65% are commonly accepted as defensible in commercial cases.

3. Audit trails. Every modern review platform logs which documents were reviewed by AI, what coding decisions were made, and who confirmed them. Preserve these logs. They are your shield in a Rule 26(g) sanctions inquiry.

The leading 2025 and early 2026 case law has been consistent: courts do not require parties to disclose every prompt or every model version used in review, but they do require parties to be prepared to defend the methodology with concrete evidence of validation. Generic claims like "we used AI" or "we used TAR" do not pass muster.

When the Other Side Wants Your Prompts

A growing tactic in 2026 is opposing counsel demanding production of prompts used during review or in drafting. Most courts have rejected blanket demands as work product, but have allowed targeted inquiries where there is specific reason to suspect a defect in the methodology.

The practical implication: do not write prompts that would embarrass you. Write them as you would a privilege log entry — professional, neutral, and explainable.

A bad prompt:

Find all documents that help our case and skip the bad ones.

A defensible prompt:

For each document, indicate whether it discusses (a) the contract
formation between the parties, (b) any modification or amendment of
that contract, or (c) communications regarding performance. Provide
a short factual summary and the names of the people involved.

The second is neutral, traceable, and consistent with FRCP Rule 26 obligations.

Recall, Precision, and Budget

Many lawyers do not know what recall and precision mean in review. A quick refresher.

• Recall is the percentage of all responsive documents in the corpus that your review identified as responsive. Higher recall = lower risk of missing important material.
• Precision is the percentage of documents you coded as responsive that actually were responsive. Higher precision = lower wasted review.

You almost always trade them off. Targets in 2026:

• High-stakes commercial matter: recall around 80%, precision around 70%
• Routine commercial matter: recall around 75%, precision around 65%
• Internal investigation: recall around 90%, precision lower (be over-inclusive)

Budget rule of thumb: AI-leveraged review in 2026 typically costs 35% to 60% less than 2022 manual review at the same recall, with the savings coming from reduced reviewer hours, not from cheaper software.

Key Takeaways

• TAR 2.0 and generative AI are complements: TAR prioritizes the corpus, generative AI handles coding, privilege, and synthesis on the priority queue.
• Write your methodology down before review starts and validate it with recall and precision at the end.
• Keep audit trails. Most modern challenges fail when methodology is documented; almost all succeed when it is not.
• Write defensible prompts. Treat them as discoverable until proven otherwise.
• Set realistic recall and precision targets up front, negotiated into your ESI protocol when possible.