Why GenAI isn’t enough for legal work

Author: Laurence Baker
Read time: 3 minutes
Published date: July 6, 2026

Large language models can summarize and draft, but they can't replicate firm-specific standards, client history, or the judgment that legal work requires.

The next generation of legal AI won't just write—it will need to understand.

In a recent article, we explored the surge in legal AI startups and the growing trend of tools built on general-purpose language models. These platforms promise to make lawyers more efficient, but often miss the mark when it comes to the realities of legal practice.

Why? Because law isn't just about documents. It's about context. Judgment.

And no general-purpose AI delivers that out of the box.

Legal work demands more than language fluency

Large language models (LLMs) are good at many things. They can summarize, translate, draft, and even answer complex questions. But they don't understand the why behind legal decisions, or the subtleties that differentiate a solid decision from a risky one.

Legal work is layered with firm-specific standards, jurisdictional considerations, evolving regulatory obligations, and client-by-client preferences. Even something as simple as a non-disclosure agreement (NDA) can carry different requirements depending on the industry, counterparty, or desired outcome.

This kind of context doesn't live in public training data. It lives in internal notes, playbooks, past decisions, client history, or a partner’s brain. And unless the AI has access to all that—and knows how to interpret it—it's just guessing.

The myth of the plug-and-play legal AI

A growing number of vendors offer tools that promise quick wins: just upload your contracts or connect your data, and the AI will start working for you.

The reality is far messier.

To get AI tools to truly reflect a firm's standards and approach, someone needs to teach them. That means:

• Curating high-quality internal documents
• Defining what "good" looks like in specific contexts
• Providing feedback on AI outputs
• Iterating again and again as edge cases emerge

This process is often called fine-tuning, and it's essential for accuracy. But here's the problem: most vendors put that burden on the client.

Why this rarely works in practice

Legal teams are not data teams. They don't have the time—or the technical support—to tag documents, evaluate AI decisions, and manage iterative training loops. And even if they did, most firms don't have their data in the right format. It's scattered across emails, Word docs, deal folders, legacy systems—and, in many cases, physical file boxes.

The result? Tools that never quite click. Lawyers try them once or twice, hit inconsistencies, and revert to the manual process they trust. Adoption stalls. Return on investment evaporates.

Even worse, the firm may have invested time and effort into customizing the tool without getting value back.

The return on investment problem in legal AI

If an AI tool requires hours of lawyer input before it delivers a usable result, it's not really saving time. And if the outputs still need to be double-checked, edited, and explained, it's not reducing risk either.

That's the trap many legal teams fall into. The promise of AI feels compelling—especially when budgets are tight and workloads are high. But unless the model is built around the firm, not the other way around, it rarely delivers meaningful results.

It's not just about what the AI can do. It's about what it can do without asking the lawyer to do more.

A different approach: embedded, data-rich AI

At Carta Law, the approach is different.

Instead of asking lawyers to train a tool, Carta Law has built a Matter Intelligence Engine, its internal AI management tool, directly into existing workflows. Because Carta Law already handles high-volume, repeatable legal work, its data has been structured from the start. That means the Engine and the agents that sit below it are already trained on each of our service workflows, processes, and market standards, and already understand the context.

The Engine can instantly pull client-specific guidance, flag issues based on historical preferences, and offer suggestions that reflect real-world decisions—not just legal theory.

There's no separate tool to log into. No new interface to learn. Just instant access to the right information, in the place lawyers are already working.

Less input, more outcome

This is the future of legal AI: not over-engineered dashboards and portals or AI drafting, but embedded intelligence that makes lawyers faster, more confident, and more consistent—without slowing them down.

When AI is built around the firm, it becomes an enabler. When it's bolted on, it becomes a chore.

That's why the most effective legal AI is invisible. It doesn't ask lawyers to change how they work. It just makes their work better.