Most 'Legal AI' Is Just a Foundation Model Behind a Brand. Here's Why That's Not Enough.

Alex Su posted something last week that got 80,000 views: "At this point I can't tell if Harvey is in the business of selling technology to lawyers or equity to VCs."

That got me thinking. Not about Harvey specifically, but about what most "legal AI" actually is under the hood. Strip away the brand, the enterprise pricing, the sales deck. What's left? A foundation model. The same Claude or GPT anyone can use, pointed at your data room.

These models extract text fine. But lawyers don't think like regular people, and foundation models think like regular people. When an amendment uploaded a week after the parent agreement supersedes a termination clause, a foundation model doesn't know. It has no concept of what an amendment is. No one engineered it to understand that.

We spent years doing exactly that. We call it Legal-Native Intelligence. A purpose-engineered reasoning layer, calibrated against thousands of real M&A transactions, that gives foundation models the legal eye they lack on their own.

Extraction Is a Commodity. Understanding Is Not.

Any LLM can pull words from a PDF. That's table stakes. The real question: does your AI understand how legal documents actually work?

Any attorney who has been doing this long enough can tell whether an agreement is well-drafted or amateur hour with a quick skim of the headers and formatting. They don't need to read every clause. Pattern recognition built from years of practice tells them instantly: this was drafted by someone who knows what they're doing, or it wasn't.

That's legal intuition. Foundation models don't have it. We engineered it.

Legal-Native Intelligence is a purpose-engineered reasoning layer, calibrated against thousands of real transactions, that encodes that same legal intuition into infrastructure. Document relationship detection, amendment chain resolution, cross-reference linking, defined term propagation. Not a prompt. Not a wrapper. Years of engineering distilled into systems that give models the pattern recognition lawyers develop over careers.

We wrote about this in The F1 Engine Problem. The most powerful engine in the world is not useful without the right chassis. In legal AI, that chassis is Legal-Native Intelligence.

The Amendment Problem

Here is a scenario that plays out in every data room.

A Master Services Agreement is uploaded on day one. A week later, Amendment No. 3 shows up as a separate, unlinked file. The amendment states: "Section 4.2 is hereby deleted in its entirety and replaced with the following..."

Every other AI treats these as two unrelated documents. It extracts the original MSA provisions. It confidently reports a 30-day termination clause that was superseded two years ago. The output looks clean. It is wrong.

Mage recognizes the amendment as a modification to the parent agreement, even when uploaded separately and unlinked. It reads them together. It extracts the current effective terms: 90-day termination per Amendment No. 3. It flags the change.

This is not a prompt engineering problem. You cannot instruct a generic model to reliably do this. It requires understanding what an amendment is as a legal concept, how it modifies a parent agreement, and what "deleted in its entirety and replaced" means for the enforceability of the original provision.

That understanding has to be purpose-engineered. We built it.

Beyond Amendments: The Legal Eye

Amendment chain resolution is one example. Legal-Native Intelligence handles the full spectrum of document relationships that lawyers navigate instinctively but foundation models miss entirely.

Document family recognition. A commercial lease, its guaranty of lease, and an estoppel certificate are not three unrelated documents. They are one family. When reviewing the lease, you need the guarantor's obligations and the estoppel's representations in context. Legal-Native Intelligence recognizes these relationships and presents them together.

Cross-reference resolution. "Subject to Section 12.1" appears in a termination clause. A foundation model extracts the termination clause and moves on. Legal-Native Intelligence follows the reference, reads Section 12.1, and surfaces the actual constraint on termination rights. The extracted provision is complete, not truncated at the cross-reference.

Defined term propagation. "Material Adverse Effect" is defined in Section 1.1. That definition controls the entire agreement. Every representation, every closing condition, every indemnification threshold references it. Legal-Native Intelligence traces this defined term across every provision where it appears, ensuring the extracted meaning reflects the actual contractual definition rather than the model's generic understanding of the words.

Why Prompts Can't Fix This

You can tell a model "look for amendments." You cannot make it understand what an amendment does to a parent agreement's enforceability. You can tell it to "follow cross-references." You cannot make it understand that "subject to Section 12.1" fundamentally changes the meaning of the clause it appears in.

Lawyers develop this intuition through years of practice. Thousands of agreements. Hundreds of transactions. The patterns become automatic: see an amendment reference, check the chain. See a defined term, trace it to the definition. See a cross-reference, follow it.

We distilled that intuition into infrastructure. Systems calibrated against thousands of real transactions that encode how lawyers actually reason about document relationships. You cannot replicate this with a system prompt. The reasoning has to be built into the architecture.

Legal-Native Intelligence + Model Fusion Technology

Legal-Native Intelligence is the purpose-engineered reasoning layer that gives models a legal eye. Model Fusion Technology is the statistical rigor that ensures accuracy.

Here is how they work together: multiple frontier models analyze the same document, each operating through legal-native infrastructure calibrated against real transactions. The best outputs are fused together. The result is extraction that is both legally sophisticated and statistically verified.

Legal-Native Intelligence ensures the models understand amendment chains, cross-references, and document families. Model Fusion ensures the extracted values are accurate through multi-model consensus. This is why our accuracy is materially higher than anyone else's. Not because we use a better model. Because we instrumented the models with legal reasoning and then verified the output through statistical consensus.

The Stakes

The difference between "Termination: 30 days written notice" and "Termination: 90 days written notice, per Amendment No. 3" is not academic.

One is the superseded provision from the original agreement. The other is the current effective term. Report the wrong one, and your client plans for a 30-day exit that doesn't exist. That could mean post-closing liability, a failed transition, or a deal term that unravels.

Every provision in a data room exists in a web of amendments, cross-references, and defined terms. Foundation models see isolated documents. Legal-Native Intelligence sees the web.

---