Vertical or horizontal: where AI applications will win

Two kinds of AI company

Horizontal AI sells a general capability to anyone who will pay: assistants, coding tools, enterprise search, agent platforms. Microsoft 365 Copilot, priced at $30 per user per month when Microsoft announced it in July 2023, is the clearest example. Glean in enterprise search, which raised a $150 million Series F at a $7.2 billion valuation, and Cursor in coding are the startup versions.

Vertical AI sells a specific job inside one industry. Harvey does legal work. Abridge turns doctor and patient conversations into clinical notes. EvenUp builds personal injury claims. The buyer is a law firm, a health system or a claims practice, not a general knowledge worker.

By spend, horizontal is ahead today. Menlo Ventures estimated that in 2025 enterprises spent $8.4 billion on horizontal AI applications (copilots alone were $7.2 billion), $7.3 billion on departmental tools and $3.5 billion on vertical AI. Healthcare took $1.5 billion of the vertical total, 43%, and tripled in a year. Bigger is not the same as durable, though, and durability is what an investor should be judging.

The case for horizontal, and its problem

Horizontal markets are huge, and the best horizontal products grow faster than anything I have seen in software. Cursor announced a $2.3 billion Series D at a $29.3 billion post money valuation in November 2025 and said annual recurring revenue had passed $1 billion. Dealroom reported in June 2026 that its annualized revenue had reached about $4 billion, with about 75% coming from enterprise customers.

The problem is that horizontal products sit directly in the path of the model labs and the suite incumbents. Jasper is the cautionary case. It had raised at a $1.5 billion valuation, and then ChatGPT arrived as what one report called a formidable low cost competitor practically overnight. Jasper laid off staff in July 2023, cut its 2023 ARR forecast by at least 30% and cut its internal valuation by 20%. Chegg is not an AI company, but it shows the same force: its shares fell more than 40% in May 2023 after its CEO said student interest in ChatGPT was hurting new customer growth.

Bundling is the quieter version of the same risk. When Microsoft sells Copilot as an add on to Microsoft 365, a standalone horizontal tool has to beat a line item on an invoice the customer already pays. That is a hard sale even with a better product.

Coding is the exception, and I think it is worth being precise about why. Developers adopt tools from the bottom up, outcomes are easy to measure, and the product sits close enough to the model that the company can train its own. Cursor says its models generate more code than almost any other LLM. In my view, coding behaves less like a horizontal market and more like a vertical one: a specific user, a specific workflow, a measurable result. It still competes with GitHub Copilot, OpenAI and Google, which is why I would price it as a contested market, not a monopoly.

The case for vertical

Bessemer published a vertical AI thesis in September 2024 that I mostly agree with. Its partners argued that vertical AI goes after high cost, repetitive language tasks in fields like law, healthcare and finance, and reported that LLM native vertical companies had reached about 80% of the average contract value of traditional vertical SaaS, while growing around 400% a year at roughly 65% gross margins. They also pointed out that business and professional services are 13% of US GDP, against about 1% for software. My shorthand: vertical AI sells work, not seats.

Harvey is the leading example. In September 2026 it raised $550 million at a $15.5 billion valuation, co-led by Lightspeed and Diffusion, and said 80% of the Am Law 100 use it. The Next Web reported about $400 million of annual recurring revenue and more than 3,000 organizations as customers. Harvey has also started post-training its own open weight model, which tells me the strongest vertical companies plan to own more of the stack, not less.

In healthcare, Abridge raised a $300 million Series E led by Andreessen Horowitz, with Khosla Ventures, at a $5.3 billion valuation, and works with more than 150 health systems, including 26,000 clinicians at Kaiser Permanente. In legal claims, EvenUp raised $150 million at a valuation above $2 billion in October 2025, led by Bessemer. It says it supports more than 2,000 law firms, has helped resolve more than 200,000 cases, and runs a model trained on hundreds of thousands of cases and millions of medical records.

What makes an AI application durable

I look at six things. The first is workflow depth. A note that lands in the medical record or a demand letter that goes to an insurer is far harder to rip out than a chat window. The second is a proprietary data loop. EvenUp trains on case files and medical records that no model lab can scrape from the public web.

The third is distribution. Horizontal tools fight the incumbent suites for the IT budget. Vertical tools have slower sales cycles, but once 80% of the top law firms use one product, that adoption becomes its own sales argument. The fourth is trust and liability. In medicine and law someone answers for an error, and buyers pay a premium to vendors who are built to carry that weight.

The fifth is pricing power. A vertical product priced against the hours of a paralegal or the time of a physician has room that a $30 seat does not. The sixth is encroachment risk, and here the falling cost curve cuts both ways. Cheaper inference improves application margins, but it also makes it cheap for a lab or an incumbent to ship the same feature.

Vertical is not immune. Epic, the electronic health record vendor, announced an AI assistant called Art with ambient scribe features at its user group meeting in August 2025, built with Microsoft. One health tech analyst wrote that there is no rational reason for an EHR not to offer a scribe. In vertical markets, the encroacher is usually the system of record, not the model lab. That is the risk I watch most closely.

How I underwrite a horizontal AI company

The first question I ask is what happens when OpenAI, Anthropic, Google or Microsoft ships this as a free feature. If the honest answer is that customers would switch, the company is a feature with a funding round. The second is who owns the budget line. If the buyer already pays a suite vendor, I want proof the product wins head to head in production, beyond a pilot.

The numbers I care about are daily active users as a share of paid seats, net revenue retention, gross margin after inference costs, and model costs as a share of revenue over time. My bar is net revenue retention well above 100% and gross margin that rises as model prices fall. If margin stays flat while token prices drop 10x, the company is passing all of the savings to customers because it has no pricing power.

The failure modes are familiar by now: the lab ships the feature, a suite bundles it, seats shrink as customers consolidate tools, and a model provider raises prices or changes terms on a company that cannot switch models.

How I underwrite a vertical AI company

I start with the job. What work does the product do, what does that work cost the customer in labor today, and what share of that cost can the company charge for. Then I ask who carries liability when the output is wrong, how deeply the product writes into the system of record, and whether every new customer makes the product better for the next one.

The numbers I care about are contract value relative to the labor it replaces, gross revenue retention, time to first value, expansion into adjacent workflows, and gross margin with human review included. Bessemer's 65% gross margin figure is a useful reference. A vertical company that sits far below it usually has services hiding inside the software.

The failure modes are different from horizontal ones. The market can be too small once the first workflow is saturated. The system of record can copy the feature, as Epic is doing with ambient notes. Human review can refuse to come out of the loop, pulling margins toward a services business. And regulation can move faster than the product.

Where I land

My bet is that vertical AI builds more of the durable companies, and horizontal AI builds a few very large ones that are hard to pick and expensive to own. The line that matters is not vertical against horizontal as labels. It is whether a company owns a specific workflow, carries responsibility for the result, and prices against the work it replaces. Most vertical companies are built that way by design. Most horizontal companies are not, and coding is the main exception because it behaves like a vertical.