Defensible eDiscovery Productions using AI

The next eDiscovery production fight won’t be about your documents — it will be about your AI process.

Generative AI has moved from the review room into the case law. In 2025 and 2026 decisions, courts are grappling with AI-assisted review, AI-generated business records, and generative AI in discovery workflows. ESI protocols are evolving to accommodate these issues — and what parties must disclose, validate, and defend is changing with them.

The consequences are concrete. Choose the wrong form of production under FRCP 34(b)(2)(E), and you fail the reasonably usable standard. Sign a Rule 26(g) certification without validating completeness, and you invite sanctions exposure. Produce AI-generated content without preserving its metadata, and you face challenges to completeness, reliability, and transparency. If you skip Rule 502(d) clawback protections, your privilege problems multiply.

DWR can help negotiate ESI agreements that address new AI-related issues, validate TAR and AI review workflows, and avoid production-format mistakes. Our process focuses on assembling and defending the ESI production package — defensible keyword and filter criteria generation, quality-control documentation, redaction and privilege checks, and Rule 26(g) discipline that survive scrutiny.

Rule 26(g) requires that every discovery request, response, or objection be signed by an attorney (or the party if unrepresented), certifying after a reasonable inquiry that it is consistent with the rules, not for an improper purpose, and not unduly burdensome; violations require mandatory sanctions.

DWR eDiscovery will provide high-touch support, feedback and consulting when using our AI tools.

• ESI Protocol Negotiation
  Support when negotiating TAR/CAL and AI protocols whose transparency and validation terms hold up as ESI protocols evolve.
• Form of Production
  Advice on how to support native, TIFF, or PDF under FRCP 34(b)(2)(E) and the reasonably usable standard without production-format mistakes.
• Emerging AI Issues
  Discussions on how collaboration platforms and GenAI review tools change what courts expect from review and production workflows in 2025 and 2026 decisions.
• Rule 26(g) Certification
  Validating production completeness and documenting quality-control measures before certifying — the discipline that limits sanctions exposure.
• Production Specifications
  Building production specifications and load file integrity checks that account for AI-generated content and its metadata.
• Privilege and Clawbacks
  Structuring effective Rule 502(d) clawback protections and redaction quality control to protect privilege in AI-assisted review.

DWR Customer feedback: For two years, we have used AI for DWR customers as a professional service, with a minimum fee of $2,000, to cover the initial tokenization and concept/topic exploration and associated reports. Customers use the reports to provide feedback. After attorneys understand the corpus in terms of probative and non-probative topics, we often use AI to generate complex Boolean searches and filters to mark and issue-code all potentially probative documents, and issue-code by topic. Task-by-task deliverables are based on topic summaries and topic signal strength. Common tasks include iterating on topics and subtopics, summarizing by topic and subtopic, timelines by topic and/or people, and document summaries in a specific, well-defined topic area or for individual documents within a well-defined timeline.

DWR AI assist, process-specific AI tools. Over the last 18 months, DWR has worked with two sets of AI specialists to deliver AI tools and a DWR interface for a number of narrowly defined tasks and objectives. DWR is conducting a soft launch with select customers to tweak user interfaces. Expect a formal product launch before the end of 2026.

Some high-level Design constraints for the Rollout of DWR AI automation tools:

• Confidentiality: LLMS must not learn from client data.
• Context and specificity: AI cannot learn across matters (Matters are silos)
  • • • • • Defensibility: The results of your AI analysis and work product are both auditable and defensible.
            We rely on the most conservative case law to date to inform our defensibility constraint. This is largely achieved by (i) disallowing AI creativity (the AI output won't generate fictitious results), (2) allowing AI to translate results into complicated Boolean searches and filters that can be audited by opposing parties and the courts if challenged.
• Value: The solution will save more money than doing specific tasks in more traditional ways.