How to Succeed with an AI Wrapper (Part 1): Proprietary Data

Proprietary Data Wins (Whether We Like It or Not) 

There is an inconvenient truth about AI in law that people keep trying to talk their way around. Proprietary data wins. It always has. It still does. It probably always will. That feels unfair, anti-competitive, and structurally hostile to startups, but law is not a fairness-based market. It is a risk-based one. Lawyers do not buy tools to express values about innovation; they buy tools to reduce uncertainty and increase coverage. Nothing reduces uncertainty like having access to more things. 

Why the Incumbents Are So Hard to Dislodge 

This is why incumbents like Lexis and Westlaw are so difficult to dislodge. They did not become dominant because of interface design or clever workflows. They became dominant because they accumulated vast, licensed, curated datasets over decades: case law, legislation, commentary, looseleafs, practice notes, and editorial overlays built by humans who understood doctrine and consequence. That accumulation is brutally hard to replicate. It is slow, expensive, and requires patient capital, institutional governance, and relationships with authors and publishers who are not interested in licensing their crown jewels cheaply, if at all. Every major publisher wants to be part of the AI boom themselves. Very few are willing to hand their most valuable assets to a third party and watch someone else capture the upside. 

Commentary Is the Real Prize 

Raw law is not enough. Judgments, legislation, and public regulatory material are necessary but not sufficient. What lawyers actually rely on is interpretation, synthesis, and judgment layered on top of that raw material. That layer lives in commentary, and commentary is difficult to license, difficult to maintain, and expensive to produce. Good commentary requires subject-matter experts, editorial discipline, constant updating, and accountability when the law changes or an argument turns out to be wrong. Anyone who has tried to build or license this kind of dataset knows how quickly the costs escalate. 

Writing Your Own Dataset Is Possible, but Brutal 

In theory, a wrapper company could write its own proprietary dataset. That is the cleanest moat available. In practice, it is a brutal undertaking. You are no longer a startup; you are a publisher. You need authors, editors, review processes, version control, liability management, and a long-term commitment to maintenance. You need to fund something that will not show returns quickly, while competing against incumbents who have been doing this for decades. The upside, if you succeed, is real. A genuinely proprietary dataset is one of the few things that can protect a wrapper from being obliterated by the next base-model release. But most startups do not have the appetite or capital to play that game. 

The Better Toaster Problem 

This leads to what I think of as the “better toaster” problem. I have seen this firsthand. Large firms rigorously tested Ailira’s semantic search over public ATO material, legislation, and case law. The feedback was consistent and honest. It was faster, smarter, and more flexible. It was genuinely impressive. And yet what they wanted was not a better toaster. They already had a toaster. What they wanted was access to everything: commentary, secondary materials, cross-jurisdictional coverage, and the reassurance that if something existed, it was probably in there somewhere. Even when a tool is objectively better at a narrow task, that is often not enough to displace an incumbent in law. Lawyers optimise for coverage and reassurance, not elegance. 

Why This Fails at Renewal Time 

Right now, many firms are buying “AI” and ignoring the toaster question altogether. They are paying for intelligence divorced from data depth. It demos well and feels modern, but it does not survive contact with renewal cycles. When the excitement fades, someone will ask what the tool actually provides that the existing stack does not. If the answer is “it talks nicely, but it does not give us more authority, more sources, or more coverage,” the contract quietly dies. This is the graveyard most wrappers are walking toward. 

The Uncomfortable Takeaway 

The uncomfortable takeaway is simple. If you are building an AI wrapper in law and you do not control proprietary datasets, you are running uphill with ankle weights. You are competing against incumbents whose advantage has nothing to do with models and everything to do with accumulated authority. That does not make success impossible, but it does mean you need to be honest about whether you are actually building a moat, or just renting one until the ground shifts underneath you.