Every vertical AI app founder has the same answer to the same objection. The objection is "can't Claude just do this?" The answer is "no, a horizontal tool can't go deep enough for your industry, so buy ours, which does." I too gave that answer to VCs a few months ago. I no longer believe it.
Picture the shape
Draw how a firm will actually use AI. There's a wide horizontal bar across the top: the general assistant that reads the inbox, drafts the memo, interrogates the spreadsheet, sits in on the meeting. Every person in the building needs that, in every department, whether or not you ever buy anything specialised.
Then there's a bar dropping straight down from it: real depth in the one or two processes that are your actual business. Reviewing a contract against your own positions. Screening a candidate the way your firm actually screens. The future isn't horizontal or vertical. It's T-shaped: one general assistant everyone stands on, and firm-specific depth in the few processes where judgement and regulation actually bite.
Vertical app builders sell you the down-stroke and tell you the cross-bar won't reach it. There are two holes in that.
You're buying the horizontal tool anyway
No firm is going to run its whole operation through a contract-review app. It won't answer the HR question, and it won't draft the board paper. So the general assistant arrives anyway, at least for everyone doing knowledge work, and you're paying for the horizontal tool regardless.
Which means the vertical app isn't the thing you buy instead of the horizontal one. It's a second subscription and a second place your data lives, sitting next to a tool you were always going to pay for. The question stops being "vertical or horizontal" and becomes "do I need a separate product for the deep part, or can the tool I already have go there?" A few months ago I'd have told you it can't. Then the tooling caught up.
The deep part is now something you write, not something you buy
Anthropic calls them Skills. A skill is a folder of plain-language instructions, and optionally a few scripts, that Claude pulls in only when the task in front of it calls for that expertise. That last part matters: it's progressive disclosure. The model doesn't lug every specialism around at once. It loads your contract-review playbook when it's reviewing a contract, and leaves it on the shelf the rest of the time. The down-stroke of the T is now something you write on top of the horizontal tool, not a separate product you have to buy.
The part that should change how you buy is who holds the pen. You write the skill in words you can read. You keep it in files you control, and you change it the morning a regulation changes, or the afternoon a lawyer tells you the last version waved through a liability cap it should have flagged. A standalone app buries that same logic in a codebase you'll never see, charges you a seat to rent it, and ships the fix on its roadmap rather than yours. Same expertise. One version you own and improve, one you lease and wait on.
Owning the skill is only half of it. You also have to know whether it works, which means evals: a bank of real cases with known-good answers that you run every version of the skill against before you trust it. Those evals, and the corrections that come out of them, are the loop that turns a plausible-looking playbook into one you'd stake a client file on. Own that loop and every correction makes your version better. Rent it, and the compounding happens on someone else's side of the wall. It's the same argument I made about the learning loop you own.
This doesn't make every vertical app pointless. Some are real businesses rather than wrappers: Harvey and Legora in legal, FurtherAI in insurance, Abridge in clinical notes. They've built depth you couldn't reproduce with a page of instructions, and often real infrastructure underneath it: integrations into your systems of record, permissions, audit logs, workflow state that has to persist between people. If that's what you're buying, buy it. But if the value is really your fallback positions and house drafting standard, ask why that has to live inside someone else's codebase and someone else's roadmap.
And if you're the one building it, and after all that you still think you're special: ship the depth into the tools people already use. Build it headless, as an MCP server the horizontal assistant can call, not as one more app your customers have to switch into and out of. Every firm I talk to is already tired of juggling AI logins. Nobody is asking for another window to keep open.
We told VCs the opposite
I'm not neutral about this. Earlier this year Airclerk was building a standalone vertical app for commercial insurance, and when we raised it with investors, the story we told was the one I've just pulled apart: Claude is impressive, but it can't be trusted to go deep on the hard parts of commercial insurance, so someone has to build the specialist layer on top, and that someone is us. We believed it.
Then we started using skills to make Claude do the work the app was going to do. It did it. Anyone could open the skill and read, in plain English, exactly what it did and why. We threw the codebase away in May, figuratively speaking; the parts worth keeping became Claude skills and MCP tools. The longer version of that decision is here. The uncomfortable part is simpler than the strategy note makes it sound: the argument we'd been using against Claude stopped being true in our own hands.
What it means if you're buying
When a vendor tells you their vertical product exists because the horizontal AI can't go deep, ask two things. Am I going to buy the horizontal tool anyway, for everything else? Almost always yes. And can the deep behaviour you're selling me be written as a skill I own and change, rather than a product I rent? Increasingly, also yes.
Buy the horizontal foundation once. Grow the vertical depth yourself, in the open, in language your own people can audit. That's the bet we're making with our own money now, instead of describing it to someone with theirs.