I have spent fifteen years doing due diligence on companies before clients invest in them or acquire them. The skill that matters most in this work is not technical analysis. It is reading a company’s own materials—pitch decks, marketing pages, investor presentations—with a particular kind of skepticism. The question I ask, over and over, is: which of the claims this company is making about itself are real, and which are claims they need to make in order to justify their valuation?
I have been doing this exercise, informally, with the specialized legal AI vendors—Harvey, CoCounsel, Spellbook, and others—for about a year. Reading their public materials. Listening to their conference talks. Talking to lawyers who have used them. Following their funding announcements.
And after Anthropic’s Claude For Legal announcement two weeks ago, I think I can finally say something about what I’ve found.
The vertical legal AI companies have built their entire positioning around three competitive moats. As far as I can tell, none of them are real.
I want to walk through each of the three, because I think this matters not just for these specific companies, but for how lawyers should evaluate which tools to bet their workflow on.
Moat One: Industry Expertise
The first claim, made in nearly identical language by every vertical legal AI company, is: “we are built by lawyers, for lawyers. We understand legal work in a way general AI tools cannot.”
The framing assumes there is a meaningful gap between “AI that understands legal work” and “AI that does not.” There is such a gap. The question is whether the gap is one a customer can actually detect, and whether they would pay for it if they could.
My honest assessment is that the gap exists, but it is shrinking, and—more importantly—it is shrinking in a way that makes it harder for non-experts to perceive.
A senior litigator might be able to tell that a vertical legal AI’s output is, in some subtle way, more lawyer-like than a general AI’s output. The phrasing is better. The structure is more familiar. The instinct for which arguments to feature first is sharper. These are real differences.
But these differences are perceptible only to people who already know what good legal output looks like. To an in-house counsel reviewing a draft, the difference is detectable. To a small-business client receiving a contract, it is not. To the general counsel deciding which AI tool to buy, it is partially detectable—but increasingly hard to articulate in terms that justify a four-times price premium.
This is the structural problem with industry expertise as a moat: it depends on the customer being able to recognize the expertise. And as AI quality across the board rises, the recognizable differences narrow. The vertical product is still better, in absolute terms, on some dimensions. But the perceptible better-ness is getting smaller. And customers do not pay premiums for differences they cannot perceive.
The moat depends on customer perception, not on actual capability. It is, in a real sense, a moat made of customer ignorance—and customer ignorance is not a sustainable defensive position.
Moat Two: Information Asymmetry
The second claim is that vertical legal AI tools have access to information that general AI tools do not. This claim has two parts.
The first part is database access. Westlaw integration. LexisNexis integration. Document management system connectors. These were, for two years, the things vertical legal AI tools had that ChatGPT and Claude did not.
This claim collapsed two weeks ago. Anthropic released more than twenty connectors covering exactly these systems. The exclusivity is gone. The integration is now a feature of the general platform, not a moat of the vertical platform.
The second part is behavioral data. This is the more interesting claim. The argument goes: lawyers using our tool generate usage patterns, edits, preferences, corrections. Over time, this behavioral data lets our tool become better at predicting what lawyers actually want. The general AI tools do not have this layer of legal-specific behavioral data.
This claim is partially true today and almost entirely false in the medium term, for a structural reason that the vertical companies have not yet absorbed.
The reason is that lawyers are now using general AI tools extensively. When a lawyer uses Claude to draft a contract, the lawyer’s edits, preferences, and corrections become part of the data the general platform sees. The “behavioral data advantage” the vertical tools had was based on lawyers using their tools rather than general tools. That premise is no longer true. Lawyers are using both. And the lawyers who get the most out of AI are precisely the ones experimenting with everything—which means the most valuable behavioral data is being generated on the general platforms now.
The vertical tools have a head start on behavioral data. The general platforms have a flatter learning curve and a vastly larger user base. The head start narrows every quarter.
The information asymmetry, in other words, is not durable. It was a real advantage two years ago. It is a temporary lead today.
Moat Three: Training Data Advantage
The third claim, often the most technical-sounding, is: our models are specially trained on legal data. We have proprietary access to case law, contracts, legal commentary, and similar material that general AI models were not trained on. Our outputs are better calibrated to legal work.
This was once a real advantage. It is increasingly not, for two reasons.
The first reason is that the general models have already absorbed enormous amounts of legal training data. Claude, GPT, and Gemini have been trained on case law, on legal commentary, on contracts, on regulatory texts. The data was publicly available; the general platforms ingested it. The proprietary legal training advantage that vertical companies had was real, but it was a marginal advantage on top of the general model’s already-broad legal exposure.
The second reason is that grounding via connectors is replacing training as the dominant approach. When Anthropic gives Claude a Westlaw connector, Claude no longer has to rely on training-time knowledge of case law. It can query Westlaw directly, in real time, for current case law. This is structurally better than training-based knowledge in two ways: it avoids hallucination (the cases exist; they are being retrieved, not generated), and it stays current (today’s cases are accessible, not just cases that existed when the training data was assembled).
A vertical AI company that bet heavily on training-data advantage is being out-competed by a general AI company that bet on real-time grounding. The bet has been settled. Real-time grounding won.
The training-data moat was, in retrospect, betting on the wrong substrate. Training data is static. Grounding is dynamic. In a fast-moving field, dynamic beats static.
What’s Left
Walk through the three moats and what remains?
Industry expertise is real but increasingly imperceptible to customers. Customers will not pay for perceptual differences. The moat dries up.
Information asymmetry is collapsing as general platforms get the same integrations and absorb more lawyer behavior. The moat shrinks every quarter.
Training data has been outflanked by real-time grounding. The moat is built on the wrong substrate.
What is left that vertical AI companies can defend?
Two things, in my honest assessment.
The first is depth within a narrow workflow. A vertical tool that has spent years refining one specific kind of legal task—say, contract redlining with deep templates and playbooks—may have accumulated workflow depth that the general platforms have not yet replicated. This is a real but specific defensible position. It is not a category-wide moat. It is a niche.
The second is customer relationships and switching costs. A vertical company that has spent three years training a particular law firm’s employees on its tool has accumulated organizational inertia. The firm will not casually switch. This is a real but temporary moat. Inertia decays. New hires arrive who do not have the training. General platforms keep getting better. The inertia advantage has a clock on it.
Neither of these defenses scales. Neither of them justifies the valuations the vertical legal AI companies have raised at. Both of them suggest that the future of vertical legal AI is much smaller than the current investment narrative implies.
Why This Matters for Lawyers
If you are a lawyer evaluating which tools to use, the analysis above suggests a different question than the one most vendors want you to ask.
The vendor’s preferred question is: “which AI tool is best for legal work?”
The better question is: “which AI tool will still exist, with its current quality and access, in three years?”
The answer to the first question might be a vertical specialist on some specific task today. The answer to the second question is much more likely to be the general platforms—because the general platforms are absorbing the vertical advantages, not the other way around.
Betting your firm’s workflow on a vertical tool means betting that the company will survive the next three years of pressure. Some will. Most will not, at their current valuations. The ones that survive will look very different from how they look today.
The honest version of “which AI should I use?” is: build your workflow on the general platform, supplement with vertical tools where the workflow depth justifies it, and assume that the supplements may need to be replaced.
That is uncomfortable advice for vertical AI vendors. It is also, I think, accurate advice for lawyers.
This is part of an ongoing series of reactions to Anthropic’s Claude For Legal launch. Earlier pieces examined the announcement itself and what the connectors actually enable.
Email [email protected] if you work at a vertical legal AI company and want to push back on this analysis. I will engage seriously with disagreement.