Vertical AI Is Eating Horizontal SaaS. Here Is What the Evidence Shows

Vertical AI companies reaching meaningful revenue share three characteristics: permissioned access to industry data, pricing tied to outcomes rather than seats, and workflow depth that took years to encode. None of those is model quality. Founders in this category consistently report that plumbing, not intelligence, decides enterprise deals, and that is the most useful finding available about where AI value accrues.

What the claim actually is

The claim is not that horizontal software disappears. It is narrower and more useful than that.

Horizontal products are built to be configurable across many industries, which means they encode almost no industry knowledge. When a general model can do the reasoning, the configurability that was the horizontal product's advantage becomes a cost the customer pays in setup and maintenance.

Vertical products encode the industry knowledge directly. The workflow, the terminology, the edge cases and the regulatory constraints are built in rather than configured, and none of that is something a model provides on its own.

That difference was worth less when software was mostly a data structure and a user interface. It is worth considerably more when the model handles the reasoning and the remaining value is context.

The three patterns that repeat

Permissioned data access

Industry data is frequently not public and not purchasable. It sits behind contracts, professional relationships or regulatory permission. A vertical company that has spent years securing that access holds something a general platform cannot acquire quickly.

The legal case illustrates both sides. Verified legal content was the structural weakness of the AI-native challenger until it partnered for access, a dynamic examined in the analysis of legal AI competition. The incumbent's owned database was its strongest asset and its only durable one.

Outcome-linked pricing

Vertical products increasingly price per resolved case, per processed document or per completed task rather than per seat. That aligns the invoice with a budget line the buyer already understands, and it sidesteps the problem where a product that reduces headcount also reduces its own revenue.

Encoded workflow depth

The unglamorous one, and the most reliable. Years of accumulated rules about what happens in an unusual case, which jurisdiction applies, what triggers escalation, and which fields are mandatory in which circumstances.

A platform can build this. The return on doing so is poor relative to its alternatives, which is a weaker protection than impossibility and a real one.

Why integration depth beats model quality

This is the most counterintuitive finding in the category and the most consistently reported.

Founders building vertical AI report that deals are decided by whether the product connects to the systems the customer already runs, handles their data formats, and fits their approval process. Model choice rarely appears in the deciding criteria.

There is a structural reason. The model is a commodity input available to every competitor on similar terms, and it changes every few months. Integration into a specific customer's environment is neither commoditised nor transferable.

The implication for founders is uncomfortable. The work that determines whether you win is unglamorous, non-differentiating in a demo, and impossible to describe compellingly in a pitch deck. It is also the work that decides the outcome.

What the evidence supports

Two findings are well supported and a third is frequently claimed without support.

Supported: the category is attracting serious capital. Legal AI venture funding reached approximately $4.3 billion across 180-plus deals in 2025, against roughly $2.1 billion across the five years from 2019 to 2023.

Supported: vertical products command premium valuations. Harvey reached an $11 billion valuation on roughly $190 million ARR, a multiple far above typical vertical software.

Not supported: that vertical is taking share from horizontal. In legal, both the AI-native challenger and the incumbent grew simultaneously. The market expanded. Work previously done by junior staff or not done at all moved to software, and both categories of vendor sold into that expansion.

That distinction matters. Vertical AI eating horizontal SaaS is a compelling headline and the observable evidence currently shows vertical AI eating labour budgets instead, which is a larger opportunity and a different argument.

Where horizontal software still wins

The vertical case is usually made without this section, which weakens it.

Genuinely universal workflows. Communication, document storage, identity and payments look the same in every industry. Verticalising them adds cost and subtracts nothing.

Buyers who value consolidation. An organisation running 291 applications is not looking for ten more industry-specific tools. One adequate system that covers several jobs frequently beats several excellent ones, as covered in the analysis of software sprawl.

Markets too small to support a vertical vendor. Plenty of industries cannot sustain a dedicated software company at venture scale. Those buyers configure a horizontal product and always will.

Speed of capability. Horizontal platforms ship frontier capability faster because they amortise the work across every customer. A vertical vendor is always slightly behind on raw capability and ahead on fit.

Vertical or horizontal, as a buying decision

The last row is the one that changes the maths entirely, and it is examined below.

How the pricing differs, and why it matters

Vertical AI prices against the cost of the work rather than the cost of the software it replaces, and that is the whole commercial argument.

A horizontal tool competes with other software, so it is priced against software budgets. A vertical AI product performing work previously done by a qualified professional competes with a salary, and salary budgets are considerably larger.

That is why vertical AI pricing can look extreme relative to software norms while remaining reasonable to the buyer. A price that is absurd per seat is modest per case resolved.

The risk in this position is that it invites scrutiny. A buyer paying salary-scale prices measures outcomes with salary-scale rigour, which means vertical AI vendors face harder proof requirements than horizontal ones ever did.

What this means if you are choosing between them

For a buyer, the vertical versus horizontal question has a reasonably clean answer, and it is not the one either vendor gives you.

Buy vertical where the work is regulated, where getting it wrong is expensive, and where the edge cases are numerous enough that configuring a general tool would take months. The premium is buying encoded knowledge you would otherwise build.

Buy horizontal where the workflow genuinely is universal, where you already run too many applications, and where adequate coverage of several jobs beats excellence at one. That describes most of what most organisations buy.

The mistake to avoid is buying vertical for prestige. A specialist tool for a workflow that was never specialist adds cost, a vendor relationship and an integration to maintain, in exchange for terminology that matches your industry.

What to watch over the next two years

Two developments would settle whether this trend is structural or a phase of the current cycle.

The first is whether vertical products retain pricing power once general models improve further. If a frontier model plus a thin verification layer reaches parity on the core work, the premium compresses quickly.

The second is whether the data partnerships holding several vertical positions together survive commercial renegotiation. Those arrangements are contracts rather than assets, and their terms are not public.

Where this argument is weak

Three problems.

The evidence base is heavily weighted toward legal, which is the most favourable vertical available. Legal work is document-heavy, judgement-based, billed by the hour and performed by expensive professionals. Very few industries share all four properties, and conclusions drawn from legal generalise badly.

The permissioned data advantage may be less durable than it appears. Data partnerships are commercial agreements rather than owned assets, and an agreement that closes a competitive gap today can be renegotiated or ended.

And the strongest counter-argument is simply that horizontal platforms have distribution. A platform reaching a buyer through an existing relationship does not need a better product, it needs an adequate one, which is the pattern examined in the analysis of category absorption.

Frequently asked questions

Where to start this week

If you are building vertical AI, one exercise separates a real position from a claimed one.

List what your product knows that a competent engineer with a frontier model could not encode in three months. Be specific: which edge cases, which jurisdictions, which validation rules, which data you can reach and they cannot.

If the list is short, that is the roadmap. If it is long, that is the pitch, and it is a better one than anything about your model choice.

Buyers should run the same list from the outside and compare answers with the vendor's version.

Related reading