Vertical AI for M&A: The Missing Layer

A very common question has been “why can’t the frontier models just do this?” It is a fair question and the answer matters. General AI models give us reasoning, but they do not give us M&A judgment. That is the gap.

Today, most legal AI products are built directly on top of general models, which are not trained on how M&A lawyers actually work. They do not understand the workflow, the judgment, the sequencing, the issue-spotting, or the practical patterns that associates learn by doing deal after deal. That is why many products are designed as copilots or single-session agents. They help the lawyer move faster on specific tasks, but nothing carries over and the lawyer still has to make sense of the output, fix it, interpret it, connect the dots, and turn it into actual deal work.

The missing piece is the middle layer.

For M&A, that middle layer is a layer that encodes how M&A work is actually done. Not just what purchase agreements say, but how experienced lawyers review them, what they look for, what they prioritize, what they ignore, how they connect issues across the transaction, and how one finding changes another piece of work, exactly as an experienced associate would.

You can encode a large part of that initial judgment. Associates develop judgment by seeing the same patterns repeatedly. They learn that when a purchase agreement says X, they should immediately check Y. If diligence reveals A, they know to revisit B. If a covenant is drafted one way, they know the corresponding representation may also need attention. Much of that initial judgment is pattern recognition built through repetition. That is exactly the part that can be encoded and the first layer of practical M&A judgment can be built into the system. And once that layer exists, it improves. Every time the system encounters something new, you add the principle. If X is present, the answer is usually Y. But if X, A, and B are present together, then the answer may be Z. Over time, the layer compounds. It keeps accumulating the judgment that lawyers usually develop through correction and experience, except it does not forget, does not get fatigued, and does not leave the firm.

That is basically how an associate develops. A first-year M&A associate is not trusted on day one. But after a few years, they would have seen enough. They would have made mistakes, been corrected, and have learned how seniors think. They would start to recognize patterns. At some point, you trust them more because their judgment would have developed.

The same idea applies here, except the learning curve can be compressed.

A lateral fifth-year M&A associate does not join a new firm and relearn M&A from scratch. They already know how deals work. What they learn is how that specific firm does M&A. The firm’s style, preferences, forms, playbooks, risk tolerance, and way of working. That is the right way to think about AI-native M&A products too.

The intelligence layer would be equivalent to the experience a fifth-year associate brings when they laterally join a new firm. They already understand M&A execution. The firm-specific product layer then teaches them how this particular firm negotiates, drafts, allocates risk, and runs deals. The answer is not simply that every law firm should build its own AI product on top of the same general models. If they do that, they will keep running into the same problem. The model still does not understand M&A execution deeply enough. The first step is building the middle layer. That is how this can scale.

The challenge of course is that judgment is not one decision. It is thousands of small decisions. Every line in a purchase agreement, every diligence finding, every issue raised or ignored, every drafting choice, every follow-up question, every connection between one workstream and another reflects judgment. Building an M&A intelligence layer means encoding those decisions one principle at a time. There is no shortcut. The work is slow, unglamorous, and has to be done by people who have actually practiced M&A and have the experience to make judgment calls in a structured way. The middle layer takes longer to build, but once it exists, it is very hard to replicate and it gets better after each use.

That is what we are building at UseABI. An M&A intelligence layer, and then a product on top of it. The goal is to take work that currently takes hours, sometimes hundreds of hours, and move it into minutes, by encoding the parts of judgment that are repeatable, teachable, and already learned through experience today.