In almost every conversation about generative AI within the legal space, an “ah-ha” moment arises where someone admits they’re using it but aren’t entirely sure how it works behind the scenes. The reality is that AI tools arrived fast, the marketing arrived faster and practical guidance lagged behind both.
In a recent Repario educational content webinar, Paul McVoy, EVP of Technology Innovation at Repario, and Alex Khoury, Director of eDiscovery and Partner at Smith, Gambrell & Russell, walked through AI foundations to bring clarity to a complex world as it evolves discussing:
- What generative AI actually is
- How it functions inside legal workflows
- Why the human element remains non-negotiable
HOW WE GOT HERE
Legal technology has been chasing the same problem for decades: data volumes that grow faster than the tools to manage them. Paper review gave way to coding sheets, then graduated to scanned images, then expanded to databases and eventually incorporated technology-assisted review (TAR).
Each generation solved part of the problem yet left the door open for compounded confusion and the need for clarity. During the webinar, Khoury put it plainly: “Generative AI is really the first tool where I feel like we have a chance to win the fight and be able to successfully and efficiently manage this ever-growing volume of documents.”
What also feels different is adoption speed. While TAR required years of court approval, judicial education and practitioner training before it gained traction, Generative AI arrived already embedded in people’s personal lives. The lawyers who once needed to be persuaded to use predictive coding are now asking their eDiscovery providers how to use AI more, negating the option to ignore and not use it.
CONCEPTS AND TERMS WORTH KNOWING
- Transformer: The transformer changed how AI processes information. Earlier systems like TAR processed data serially: one document, one decision, in order. Transformers process in parallel, analyzing parts of a sentence simultaneously against each other. This is why the same prompt can produce slightly different outputs at different times.
- RAG: RAG (Retrieval Augmented Generation) is the difference between novelty and utility. RAG ties the model’s outputs to a known document set — your case data, your client files — so results can be grounded and traced. As McVoy explained during the webinar, “For us in our world, it was a necessary step because we really needed to have backup for answers.” Khoury agreed, “RAG is a must-have in the legal setting. It’s what separates a Gen AI tool from being a novelty to being an actual working tool in your professional environment, because it ties the work product.”
- Agentic AI: Agentic AI is coming, yet practitioners should approach it carefully. Agentic systems do not just answer questions, they take autonomous action based on a prompt. The upside is significant, particularly for repetitive discovery tasks like validating production completeness or checking document counts. The risk is also real. A widely reported incident involved an agentic system overriding its own security guardrails and deleting an entire database. As McVoy described it, “Agentic AI is kind of like our teenagers where they’re not asking for permission, they’re asking for forgiveness. And until we learn how to train it better, it’s one of those tools, like Alex said, that has risk, but potential, just like our teenagers.”
- Tokens, Chunks and Context Windows: A token is the smallest unit an LLM processes. Tokens also establish how Gen AI usage is priced. Chunks are how data is grouped for processing. Context Windows determine how much data the model can consider at once. When a multi-hundred-page deposition transcript is chunked for summarization, a key sentence may end up separated from the context that makes it important. The AI does not necessarily misunderstand it; it analyzes it in isolation. Solutions like overlapping chunks and RAG-based verification are being built to address this. Understanding the mechanism helps practitioners interpret unusual results instead of losing confidence in the tool entirely.
HOW GENAI WORKS IN LEGAL AND WHERE THE LINES ARE
One of the more useful distinctions McVoy drew is the difference between two categories of legal AI use: attorney work product and discovery work.
- Attorney Work Product — drafting briefs, legal research and legal analysis involves generating something new that will go to courts, clients or opposing parties. The hallucination risk is higher, the validation burden is higher and the consequences of failure are more direct.
- Discovery Work — document categorization, responsive coding, PII extraction and privilege analysis are different. The task is classifying and organizing what already exists, not creating something new. The hallucination risk is lower. The workflow is more analogous to what eDiscovery practitioners have been doing with TAR for years, just executing with a different engine.
Teams that use tools interchangeably across both categories, applying the same workflow to brief drafting and document categorization alike, aren’t getting the most out of either. The use cases are fundamentally different, and so are the risks.
There is also the matter of Shadow AI.
Shadow AI is the unauthorized use of personal, unapproved AI tools by employees to process work-related content, including confidential or privileged information, outside of firm-sanctioned platforms and data protections. When an associate drops a draft brief or client documents into a personal, free version of a generative AI tool, they are, by default, consenting to that tool using their prompts, conversations, and outputs to train its model. Behavioral policy is the primary guardrail most firms have in place. It is not enough on its own.
WHY HUMANS STILL LEAD
The storyline of the recent webinar was that generative AI is a tool, not a decision-maker.
McVoy was clear on this point: “We’re not relying on AI to give us the final answer. It can propose a thought, an answer, a set of documents to be produced, but those will be validated, just like we’ve always done.”
Khoury offered a practical test: before using any AI tool, think through how you will validate the output. “If you look at the output and it sounds right or is about what you expected, you’re validating your own preconceived notions. Have a plan for validation beforehand.”
Transparency is the other pillar.
As AI-assisted workflows become the standard, the meet-and-confer needs common language. Teams that can articulate what tools they are using, how they are using them and how outputs are validated will move cases forward more efficiently.
The world is also changing fast.
Meaningful case law on AI in discovery surfaced the same week McVoy and Khoury conducted the recent webinar with a California judge ruling out the discoverability of prompts. In short, ongoing education is not optional – it’s the only way to stay ahead of a market moving faster than legal technology has in the past.
In short, refusing to engage with emerging technology is not a strategy.
Refusing to engage isn’t a strategy. Adopting AI without guardrails has already put law firms in front of judges. What works is what McVoy and Khoury kept coming back to: AI-powered, human-led. Use the tools to get through the volume. Keep humans in control of what goes out the door.
This conversation is ongoing. Repario is building out an ongoing series of AI educational content, going deeper into how these tools work in real legal workflows, and introducing what we’ve been working on to help teams actually put this into practice. Stay tuned.
