There is an article sitting in my repository right now, scheduled to publish on June 8th. The title is “Vertical or Horizontal: The Open Question I Can’t Answer.” The content, in summary, is me admitting after a year of using both general-purpose AI tools and specialized legal AI tools that I genuinely could not tell which one was better for lawyers. I wrote it three weeks ago. I scheduled it. I felt fine about leaving the question open.
Three days ago, Anthropic launched Claude For Legal.
The article isn’t published yet. It is already obsolete.
I want to write about why—not because the article was wrong, exactly, but because the framing was. There was a question I should have been asking, and I wasn’t asking it. The Anthropic announcement made the question visible. This piece is my attempt to think out loud about what I now see.
How I found out about it
I want to be specific about how I came across the news, because the detail matters.
I was scrolling through a short-video platform—the kind of casual evening browsing where you watch three minutes of content and move on. Two videos in a row showed up about the same event. The first was from a legal-industry commentator I follow occasionally, describing Anthropic’s announcement as an “earthquake.” The second was from a general AI commentator, framing the same event as the moment BigLaw lost its monopoly.
The fact that both communities were talking about the same news, from completely different angles, in the same evening, told me this was not an ordinary product launch. It was the kind of event that crosses domain boundaries. When AI people and legal people are both saying “this changes everything”—in the same week, in the same week’s videos, in their own languages—that’s not marketing noise. That’s the actual ground moving.
I went to read the details. The headline number was twelve. Twelve role-specific legal plugins—Commercial, Employment, Privacy, Product, Corporate, AI Governance. Twenty-plus connectors to the software lawyers already use—DocuSign, Westlaw, LexisNexis, iManage, NetDocuments, Everlaw. Microsoft 365 integration baked in, so Claude shows up inside Word, Outlook, Excel, PowerPoint. Free access for self-represented litigants through partnerships with Free Law Project and Courtroom5.
My first reaction was uncomplicated. It was: this is going to end a lot of careers.
What I actually meant by “end careers”
I want to be careful about this, because “AI will replace lawyers” is a cliché, and clichés are usually wrong.
What I actually meant in that first reaction was something more specific. I was not thinking about senior litigators, or M&A partners, or boutique tax specialists. I was thinking about the bottom half of the profession—the lawyers whose work is repetitive, high-volume, and low in technical barrier. The ones who do the same kind of contract review, the same kind of basic research, the same kind of simple drafting, week after week. The lawyers whose business model depends on volume processing of standardized matters.
For those lawyers, Claude For Legal is not just another tool. It is a complete workflow product that does most of what they do, cheaper, faster, and—this is the key part—without the production friction that has historically protected them from displacement. The polarization of the profession that I have written about before just accelerated by something like five years in a single Tuesday announcement.
But that was the surface reaction. The deeper realization came when I read past the headlines and looked at what Anthropic had actually built.
The two details that made me sit up
Among everything in the announcement, two specific design choices stopped me.
The first was completeness. This is not a product. It is an ecosystem. The twelve plugins are not standalone applications—they are entry points into a system where the carrier (Microsoft 365), the use cases (specific practice areas), the ecosystem integrations (Westlaw, DocuSign, iManage), and the underlying intelligence (Claude) all work as one connected thing. A single specialized legal tool, no matter how good, cannot replicate this. Harvey can be excellent at contract review and still not be inside the lawyer’s email client. Spellbook can be excellent at redlining and still not connect to Westlaw the way Claude now does. The completeness is the product. Each individual capability is replicable. The integration of all of them, attached to a general-purpose model with a broader ecosystem, is not.
The second was grounding. Anthropic was very explicit about this in the launch: the new connectors are designed so that Claude draws from verified, live sources—the actual Westlaw database, the actual CourtListener archive, the actual document repositories—rather than generating answers from training data. The phrase Anthropic used was that an AI reading a real document behaves differently from one synthesizing text from training data. For lawyers, this matters more than almost any other technical detail, because the single biggest fear about AI in legal practice is hallucinated citations. The fear has been justified. It has produced the sanctions we’ve all read about. If grounding works as advertised, the failure mode that has been holding back serious lawyer adoption for two years is significantly mitigated—not solved, but mitigated.
Each of these details, individually, would be notable. Together, they describe something I had not fully imagined: a general-purpose AI platform that has acquired the specific capabilities that vertical legal AI products were supposed to monopolize.
Why my original framing was wrong
In the article I wrote three weeks ago, I framed the question as “vertical or horizontal”—as if these were two stable categories competing on a level playing field. As if specialized legal AI tools had certain inherent advantages (workflow design, database integration, hallucination mitigation), and general-purpose tools had different inherent advantages (flexibility, rate of improvement, cost). The question was which set of advantages mattered more.
What I missed was that the general-purpose platforms were going to absorb the vertical advantages. I was treating these as separate species. They are not. They are stages.
The Anthropic announcement makes this absorption visible. The twenty connectors give Claude database access that previously only vertical tools had. The twelve plugins give Claude workflow specialization that previously only vertical tools had. The Microsoft 365 integration gives Claude in-workflow deployment that previously only vertical tools had. None of these were technically impossible for Anthropic to build. They just hadn’t been built yet. Now they have.
This is not, by the way, unique to Anthropic. Claude Code has been evolving from a developer tool inside a terminal toward a more general-purpose, multi-device agent. OpenAI’s Codex is on a similar trajectory. The platforms are growing horizontally and vertically at the same time—covering more environments, integrating with more workflows, picking up more specialized capabilities. The horizontal product is becoming the vertical product. The distinction is dissolving.
The deeper reason horizontal will win
I want to articulate something that I think is the real reason this is happening, because I do not think most analysts have framed it this way.
It is not a technical question. It is a human-nature question.
Product design has a brutally simple rule: a product cannot win by being anti-human. Humans are lazy. Humans are forgetful. Humans want to do everything in one place, with one interface, with one set of credentials. Any product that requires the user to leave their current environment, learn a new system, and maintain a separate workflow loses against any product that just shows up where the user already is.
For two years, vertical legal AI tools have been asking lawyers to leave their email, leave their Word, leave their normal workflow, and use a specialized application to do legal work. The product was excellent. The friction was high. The friction was the actual problem.
When Anthropic puts Claude inside Word, inside Outlook, inside the document repositories lawyers already use—it solves the friction problem. The lawyer doesn’t have to do anything different. The AI is already there. This is not a feature. This is a structural change in the competition.
The historical advantages of vertical AI—industry-specific data training, deep workflow knowledge, domain-specific guardrails—looked like moats. They were not moats. They were temporary leads. In the AI era, where models are general-purpose and capabilities scale fast across the underlying foundation, the vertical advantages are not durable. They are catchable, and they have been caught. The human-nature advantage—being where the user already is, integrated with everything else the user already uses—is the durable one. That advantage favors horizontal.
This is what I should have seen when I was writing the original article. I was looking at the products. I should have been looking at the friction.
Where this leaves me
I am not, to be clear, saying that vertical legal AI tools are about to disappear. Harvey, CoCounsel, Spellbook, and the others have real businesses, real customers, real revenue, and intelligent teams. Some of them will adapt. Some of them will be acquired. Some of them will carve out defensible niches in specific high-end practice areas where the workflow specialization is so deep, and the customer relationship so embedded, that even an excellent general platform cannot replicate it cleanly.
But the default has flipped. Six months ago, the burden of proof was on horizontal AI to show why it could compete with vertical AI for serious legal work. Today, the burden of proof is on vertical AI to show why it should exist as a separate product category at all. That is a different conversation. And it is the conversation I was not having in my original article.
So I am going to leave the original article scheduled. It will publish on June 8th as planned. The record of my prior uncertainty will exist, in public, in its original form. But this piece, in some ways, is the more honest version of the same thinking. The open question I said I couldn’t answer turned out to be answerable. I just had to wait three weeks for the company that was going to answer it to make the announcement.
The lesson for me, going forward, is to pay more attention to the structure of competition than to the comparative features of products. The products were a snapshot. The structure determined where the snapshot was going.
I’ll try to do better next time.
The original article on this question—“Vertical or Horizontal: The Open Question I Can’t Answer”—is still scheduled to publish on June 8th. If you want to see how my thinking looked before this week, that piece is the record. I’m leaving it untouched on purpose.
Email [email protected] if you’ve been following Anthropic’s legal expansion and have views on how the vertical players should respond. I’m collecting reactions for a follow-up.