Eric Harmon is the CEO of Reveal, a global provider of leading AI-powered eDiscovery and investigation platforms.getty​Anthropic's launch of legal plugins for Claude sparked predictable headlines about AI disrupting the legal industry. Thomson Reuters stock dropped 16% in a single day. LexisNexis' parent company RELX fell 14%. The market narrative was clear: Foundation models are coming for legal tech.Foundation models are genuinely transformative for legal work. They're excellent productivity tools that will change how lawyers draft documents, conduct research and summarize materials. But here's what the headlines missed: General-purpose AI built on public internet data will not (at least for the foreseeable future) deliver the same transformation in eDiscovery. Why not? It's about fundamental architectural differences in how these systems work and what legal defensibility actually requires.​I run an eDiscovery company, so I've spent a lot of time thinking hard about where foundation models excel and where they face real limitations. The companies that will thrive understand which problems foundation models solve brilliantly and which problems require purpose-built approaches.​Where Foundation Models Excel​Foundation models are exceptional at discrete, bounded tasks, such as drafting a contract, summarizing a deposition, researching a legal question or generating an analysis. The way these models work makes them particularly good at accelerating individual tasks. They're trained on massive amounts of text to understand language patterns and generate helpful responses. For work that requires speed, creativity and good-enough accuracy, they're genuinely transformative.​Legal professionals will use them extensively, and I believe they should. The productivity gains are real.​The eDiscovery Challenge​Discovery operates differently. We're talking about millions of documents—terabytes of data and decisions that must be defensible in court. A single matter might involve processing communications from 50 custodians across five years of email, Slack, Teams and collaboration platforms. You're not summarizing one document. You're identifying patterns across datasets too large for any human to read.​Foundation models can help with analysis, but the core challenge isn't analytical; it's infrastructural. You need systems built to ingest terabytes of data, process millions of documents and maintain defensibility throughout. Better language models don't solve infrastructure problems.​More critically, the way foundation models work creates limitations that matter in eDiscovery. It is also true that in eDiscovery, one is frequently looking for the needle in a stack of needles that is the critical smoking gun in high-stakes litigation. Foundation models are better at “suggesting,” taking the routine work out of a task or completing the basics of a task to reduce error rate. This is all highly useful, but it's not exactly what needs to happen in eDiscovery. ​The Defensibility Problem​In most business contexts, 90% accuracy is genuinely useful. A human reviews the output, makes corrections and moves on. The productivity gain is real even if the AI isn't perfect.​eDiscovery requires a different standard. When opposing counsel or a court questions your process, you must defend every document classification, privilege determination and redaction—not just the accuracy but also the reasoning behind each decision.​Foundation models are optimized for helpfulness and fluency, not legal precision. They generally don't show their work or provide the transparency legal defensibility requires. This isn't something that gets solved by making them bigger or training them on more data.​Even as foundation models improve dramatically, they won't meet courtroom defensibility standards without validated processes, human oversight at critical junctures and audit trails documenting every decision. Legal work requires architectural approaches that these models weren't designed to provide.​How Purpose-Built AI Changes The Equation​AI designed specifically for eDiscovery workflows operates differently. It's calibrated for legal precision, trained on legal data to understand legal concepts and embedded in processes that provide the audit trails and validation courtroom work demands. It's also architected to show its reasoning, not just its conclusions. ​When you build AI with specific workflows in mind, you can design for defensibility from the ground up. The AI learns from how attorneys code documents, adapts based on case strategy and suggests relevance rankings that improve as review progresses. It documents every decision in ways that are designed to hold up under scrutiny.​AI Is Transformative In Legal, Including eDiscovery​​None of this diminishes the impact foundation models will have on legal work. They are genuinely transformative for productivity tasks and will fundamentally change how lawyers operate.AI is equally transformative in eDiscovery, but in a different way. Here, the impact comes from purpose-built systems designed for the specific demands of discovery work—systems that operate at massive scale, deliver courtroom-ready defensibility and embed intelligence directly into the workflows legal teams already rely on.Foundation models and purpose-built legal AI serve complementary roles. The former excels at accelerating individual tasks and generating insights, such as drafting, research and summarization. The latter provides the infrastructure, precision and defensibility needed to make those insights usable in high-stakes legal contexts.Discovery at scale, regulatory compliance and complex litigation operate under fundamentally different constraints. These domains require systems built for large-scale data processing, auditable decision-making and seamless integration into legal workflows—capabilities that sit outside the core design of general-purpose models.Legal professionals need both. As different problems demand different architectures, effectiveness ultimately comes down to choosing the right tool for the task.​Making It Work​The biggest challenge organizations face isn't technical but cultural. Legal teams need clear guidance on when to use foundation models versus specialized tools. Without that clarity, some attorneys may over-rely on ChatGPT while others avoid AI entirely.​Start by mapping specific tasks to appropriate tools. Document drafting and initial research work well with foundation models. Discovery collections, privilege reviews and productions require purpose-built platforms with audit trails. Make these distinctions explicit in workflows and training.​Organizations getting this right invest in education before technology. Run workshops comparing what each tool does well, establish approval processes for high-stakes work requiring defensibility and designate internal champions who can guide colleagues toward the right tool for each task.​Don't treat foundation model subscriptions and purpose-built legal platforms as competing budget items. They serve different functions. Frame them as complementary tools.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?