From Promise to Practice: How Legal Teams Are Actually Using Generative AI in Discovery

If it feels like you’ve attended 57 versions of the same AI webinar this year, you’re not alone. In this article and its companion webinar, we go beyond theory and directly address the practical applications of AI in eDiscovery. 

Innovative Driven experts Wayland Radin and Andy McClary hosted the session, alongside industry experts Julia Beskin, Partner at Schulte Roth & Zabel LLP, and Joe Leonard, Associate Attorney at Morris James LLP. 

You can also watch the complete webinar, segmented by topic for easy reference:

Why AI Again (and Again)? 

The legal industry has long relied on machine learning (think predictive coding and TAR), but while these systems are very cost-effective, they operate in a black box. You upload your data, train the model, and get a score. But the “why” behind the decision? That’s typically a mystery. 

Generative AI changes the dynamic. It allows for creativity, transparency, and control. Rather than needing to look at specific training documents, legal teams can prompt a model to determine relevance, extract key quotes, translate text, or even summarize documents. And with outputs that include rationale, not just results, teams gain more insight and greater flexibility. 

“Generative AI is a force multiplier, not a magic button.”

While generative AI hasn’t been “judicially accepted” for identifying relevance the way several foundational cases have done for TAR, there are still many use cases that case teams can explore to garner the efficiency of AI without stepping too far into uncharted territory. Importantly, that doesn’t mean it’s not accepted; it means parties have been meeting and conferring and agreeing to it for the most part, rather than finding themselves in a situation where they need a court or a special master to intervene. 

Another significant factor is that, while the potential efficiency gains are hard to ignore, cost remains a critical factor. While TAR tools are now essentially free, the per-document cost of generative AI can scale quickly depending on the data volume and tool used. 

That’s why the most successful implementations start with a clear goal and a well-scoped project. As one panelist put it:”Generative AI is a force multiplier, not a magic button.”

A Practical Framework: Building an AI-powered Review Workflow

When incorporating generative AI into discovery workflows, the panel emphasized a few foundational questions: 

1. What are we trying to achieve? 
2. What does a successful output look like?
3. What level of validation is required?

Whether you’re summarizing deposition transcripts, extracting PII, or translating documents, clearly defined outcomes lead to better design, better prompts, and better results. And while many large firms are still testing the waters with internal tools, some smaller firms have fully embraced external solutions with caution. Data security and privacy remain top of mind, especially when it comes to using open-source or publicly accessible tools. Enterprise-level contracts, closed systems, and internal LLMs are increasingly common ways to balance innovation with client confidentiality.

Case Studies: Where Legal Teams Are Seeing Results 

Our panel shared a range of real-world use cases where generative AI has delivered meaningful value: 

The tools may be new, but the standards for defensibility haven’t changed.

1. Privilege Log Drafting: Utilizing AI-assisted drafting and targeted human quality control, teams achieved up to 40% cost savings (compared to 100% manual logging) and 60% time savings, while maintaining full human oversight. The entries were also more organic and less templated than traditional dropdown-generated logs. These use cases were scalable across reviews from 10,000 to 500,000 documents. 
2. Inbound Production Review: In one internal investigation, a team used generative AI to classify over 7,000 documents. The AI flagged all documents relevant to known issues and successfully excluded non-relevant records. One key highlight was the count of five prompts in the review, proving there are no “zero-shot” prompts. The result: a significantly smaller, more focused set for human review.  
3. OCR Rescue and Unitization: For poorly scanned or overly inclusive PDFs, generative AI can identify handwritten notes and reunitize thousands of pages, automating a previously highly manual and expensive process. 
4. ECA Summarization: Rather than performing a linear review of documents, teams used prompts to extract key points, citations, and supporting or weakening language. The result: cutting weeks of work down to hours. 

Across all these use cases, the common thread was validation. Whether through sampling, statistical analysis, or cross-referencing with known data points, validation remains essential. The tools may be new, but the standards for defensibility haven’t changed.

Legal and Judicial Considerations: What’s Next? 

Generative AI in legal discovery is moving toward greater judicial awareness, but we haven’t had our “Da Silva Moore moment”- yet. No definitive ruling exists regarding what must be disclosed when generative AI is used in document review. But the likely path forward will echo TAR: 

1. The use of general AI will not be prohibited if done responsibly. 
2. Disclosure may include validation metrics but not prompts, which are increasingly seen as protected work product when attorney generated. 
3. Courts will expect collaboration between parties before stepping in. 

Importantly, there is growing pressure from insurers and in-house counsel to explain how generative AI was used to improve efficiency.  

Final Thoughts: The Road Ahead 

Generative AI is no longer hypothetical. It’s part of the tools legal teams are using today, both on the litigation support side and in the hands of counsel. But success requires intentionality. The most effective implementations begin with thoughtful planning, realistic expectations, and a commitment to validation at every step. 

The bottom line? Start small. Take the time to explore the tools. Build confidence in how outputs are generated, reviewed, and validated. Soon, using Gen AI won’t just be a competitive advantage, it’ll be an expectation.