P2P

50 AVOIDING LOCK-IN WITH MODEL-AGNOSTIC DESIGN As language models (LLMs) proliferate—from GPT-4 to Claude to emerging open-source alternatives—the ability to remain LLM-agnostic is becoming a key architectural principle. Rather than committing to a single model provider, some platforms are being built to swap in whatever model best fits the task—general drafting, specialized clause analysis, multilingual review, or something else entirely. This model-agnosticism protects against pricing shifts, service outages, or changing licensing policies. It also future-proofs legal teams as new, better- performing models emerge. In particularly high-security environments, some firms may even run small, fine-tuned legal models entirely on- premises, using their data to train "SLMs" (small language models) for specific workflows. These smaller models offer the benefits of AI while maintaining full data residency and control. BALANCING AI PERFORMANCE WITH GOVERNANCE The growing debate around safety, compliance, and the international governance of AI systems is not hypothetical—it is already here. Take the recent release of models like DeepSeek, which have garnered praise for outperforming better-known large language models (LLMs) in long-form summarization and legal reasoning. At the same time, concerns persist about data handling, transparency, and policy frameworks. Legal teams adopting cutting-edge models must be able to do so within strict boundaries. Hosting models in jurisdiction-aware environments— where all data processing happens on systems aligned with local privacy laws—allows firms to test new AI capabilities without compromising on regulatory obligations. This careful balance of innovation and oversight is no longer optional; it is essential for any AI adoption strategy in high- stakes legal work. BEYOND TEXT GENERATION: AI BUILT FOR CORPORATE TRANSACTIONS AND M&A Transactional legal work— especially in M&A—is fundamentally complex, nonlinear, and highly contextual. It is not just about reading contracts; it is about understanding how a web of documents interrelates, how risks cascade across structures, and how subtle legal interactions shape the dynamics of deals. AI tools built for this domain must go far beyond surface-level text generation. Next-generation legal AI platforms are beginning to mirror the way deal lawyers think. Instead of reviewing documents in isolation, these systems can conduct folder-level diligence, ingesting entire deal rooms—hundreds or even thousands of documents— and identifying how different agreements intersect, override, or create conflicts. Whether it is a shareholder agreement that modifies the terms of a prior SAFE or a side letter that introduces hidden preferences, the AI surfaces dependencies and inconsistencies that matter at the negotiating table. These systems go beyond extracting clauses—they apply deductive reasoning 67 I L T A N E T . O R G A little over a year ago, the legal industry started talking about generative AI (Gen AI) and its never-ending possibilities for the legal profession. In a remarkably short time, that vision became a reality, especially regarding e-discovery. However, this has created a whole new set of challenges for practitioners. How do they evaluate the defensibility of these solutions? How do they know they can trust the results? How do they stay on top of the ever-changing technological advances? How do they protect their data and their client's data? The answers are not simple. This article will identify a set of best practices and discovery questions that can be used to evaluate Gen AI e-discovery solutions for legal professionals. These best practices and questions were compiled from an analysis of Relativity's Gen AI-powered e-discovery tool, aiR for Review. What is the tool's intended purpose, and which model(s) does it rely upon? Understanding the use cases the solution is built for and what it claims to do is essential. Is it a replacement for first-pass review, a means for better QC, a way to issue code documents, a summarization tool, etc.? Then, evaluate whether it does what it purports to do – request a demonstration and, if possible, obtain access so you can try it out to see firsthand whether it genuinely works. If you are considering multiple technologies, be careful not to compare a solution whose primary purpose is summarization to a solution whose main purpose is for first-pass review – that will not be a fair comparison. Many Gen AI models are in the marketplace, each with different strengths and weaknesses. Find out which model or models are used by the product. Some products use more than one model, depending on the task. A company should be able to tell you which model is used for which task, following a core Gen AI principle of transparency. For example, aiR for Review currently uses GPT-4 for document analysis. GPT- 4 scored in the top 10% on the Uniform Bar Exam and is highly performant. Other solutions may be using GPT-3.5, which failed the Uniform Bar Exam and may not yield as performant results, depending on the task. How does the product keep data private and secure? Data privacy and security should be at the forefront of any technological solution, whether it incorporates Gen AI or not. You need to know where your client's data is going, who has access to it, and whether it is being stored or retained by another party. Gen AI includes creating a series of prompts where you ask the model a question or make a classification decision, and then it returns an output. In e-discovery use cases, the data itself may be sent to the model, along with the instructions. It is essential to understand where the model is housed and who has access to the data going through it. It is also crucial to I L T A N E T . O R G 66 P E E R T O P E E R : I L T A ' S Q U A R T E R L Y M A G A Z I N E | S P R I N G 2 0 2 4 B Y C R I S T I N T R AY L O R & P H I L W E L D O N Building and Applying Responsible and Defensible Generative AI Solutions for E-Discovery "You need to know where your client's data is going, who has access to it, and whether it is being stored or retained by another party." MORE ONLINE Read this article from the Spring 2024 issue. https://epubs.iltanet.org/i/1521210-spring24/65 ó

• Cover