Horizontal vs Vertical AIs

One of the seductive ideas of the last two years is that horizontal AI will flatten enterprise software. The argument runs as follows:

• Foundation models can read any document, reason across any domain, and generate any format.

• Why would an enterprise pay a premium for a specialised vertical application when a horizontal assistant, plugged into the enterprise’s data, could do the same job.

• If the intelligence is general, the applications should become general too.

• Vertical SaaS will become a transitional category; absorbed over time into a few universal platforms.

The argument sounds clean - but it is also wrong.

TLDR: Horizontal intelligence does not flatten the enterprise software landscape. On the contrary: it sharpens it. In particular, it makes domain depth more valuable. The depth dividend will disporpotionately favor vertical and specialised SaaS category owners.

To see why, it helps to be precise about what horizontal AI can and cannot do in an enterprise context.

A foundation model, straight out of the box, can:

• read a document and produce a competent summary,

• draft a response to an email,

• produce a first draft of a code change, and

• can generate a plausible-sounding analysis of a business problem.

What a foundation model cannot do (as of today):

• understand the specific vocabulary of a particular industry,

• the idiosyncrasies of its regulations,

• the shape of its transactions,

• the meaning of its codes,

• the structure of its workflows,

• the nuance of its customer expectations,

• the subtle differences between acceptable and unacceptable answers in that domain,

• …. you get the idea.

These things can be added through retrieval, fine-tuning, prompt engineering, and evaluation — but the gap is real and this can’t be solved with a better model, because it is precisely the lack of an accumulated domain knowledge and engineering around it, which results in listed failures . The horizontal tool can be pointed at the same data, and it will produce plausible (but often wrong answers) because it does not have the embedded understanding of how the domain actually works.

Vertical depth (more than a marketing claim) is a composite of several hard-won assets that are individually mundane but collectively very difficult to replicate:

1. Taxonomy and ontology. Every industry has a vocabulary, a hierarchy of concepts, and a set of relationships that experienced practitioners take for granted. Building this vocabulary into a system, cleanly and comprehensively, takes years and it cannot be bolted on afterwards. For example:

   1. A healthcare coding system knows the difference between a procedure that requires a modifier and one that does not.

   2. A reinsurance system knows what a cedant is, what a layer is, what retrocession means.

   3. A mining operations system knows what ground conditions imply for productivity and safety.

2. Labelled examples at scale. The most valuable asset a vertical SaaS company accumulates is not raw data. It is the library of “labelled examples” where domain experts have made, or corrected, specific judgements. Every labelled example is a small piece of tacit domain knowledge made explicit. Fifty thousand labelled examples in a specific vertical are more valuable than ten million unlabelled documents in a general corpus. The horizontal player simply does not have the labelled examples.

3. Evaluation and failure taxonomy. A deep vertical product knows what kinds of mistakes are acceptable, what kinds are dangerous, and what kinds are catastrophic. It has built an evaluation suite that measures performance specifically on the failure modes that matter in its domain, and a QA process that catches those failures before they reach production. Horizontal tools optimise for aggregate performance across all possible failures, which is not the same thing. In a regulated domain, missing a rare but consequential failure mode is worse than being slightly better on average. The vertical player knows this and designs for it.

4. Integration with domain-specific systems. Every industry has its own dominant systems of record, with their own quirks, their own legacy constraints, and their own political sensitivities. A vertical product that has done the work to integrate cleanly with the systems of record in its industry has a barrier to competition that horizontal tools cannot match at any reasonable engineering cost.

5. Regulatory and audit alignment. Vertical products live inside specific regulatory regimes. They understand their compliance burden, they have the audit artefacts and they speak the language of the regulator. A horizontal assistant plugged into sensitive domain data is, to a regulator, a novel object that has not yet been stress-tested. A vertical product that has passed audit in this domain for three years is, to the same regulator, a known quantity.

Each of these assets takes years to build and together, they constitute the depth that the horizontal player cannot replicate cheaply, even with a better model and vibe coding agent force.

The mistake in the “horizontal flattens everything” argument is treating the foundation model as a substitute for the vertical stack. It is not. It is a complement to it. And this is the key insight that many investors and builders currently get wrong.

When a well-designed vertical application integrates a strong foundation model, the outcomes compound:

• The vertical taxonomy guides the model’s retrieval, making its reasoning sharper.

• The labelled examples feed fine-tuning or few-shot prompting that aligns the model’s behaviour with domain expert standards.

• The eval suite catches failures that the horizontal tool would ship unnoticed.

• The integration into systems of record allows the model to act, not just advise.

• The regulatory alignment means the application can be deployed where horizontal tools cannot.

The net effect is that the same underlying model produces meaningfully better outcomes in a well-designed vertical application than in a horizontal one. The customer is not choosing between a clever assistant and a clumsy incumbent. They are choosing between a general-purpose assistant operating in unfamiliar terrain and a specialised system whose every surface has been designed for the domain.

In most serious enterprise use cases, the vertical application wins this comparison, not by a small margin, but decisively. And it wins more decisively as the workflow becomes more consequential, because the tolerance for “plausible but wrong” shrinks as the stakes rise.

This is the depth dividend. Vertical depth, combined with horizontal AI, is more powerful than either alone, and much more powerful than horizontal AI alone.

Not every vertical rewards depth equally. The depth dividend is largest where several conditions align.

• The industry has high-consequence decisions where being wrong has specific and costly failure modes. Healthcare, legal, financial services, insurance, defence, aviation, energy. In these industries, a 10% improvement in accuracy on consequential decisions is worth enormous amounts of money, and a horizontal tool that is 90% as good as a specialist is not 90% as valuable. It is maybe 1% as valuable, or often less.

• The industry has rich, structured, regulated vocabulary. Legal codes, medical codes, financial instruments, compliance taxonomies. Where the vocabulary is precise and adjudicable, a vertical product can build evaluations that directly measure correctness. Horizontal tools cannot.

• The industry has deep integration into legacy systems that are expensive and painful to replace. For instance, core banking platforms, EHR, insurance policy administration systems, production master planning software in manufacturing etc.

• The industry has sensitivity to provenance, audit, and accountability. Regulated financial advice. Clinical care. Critical infrastructure. Legal judgements. In these settings, a horizontal tool that cannot explain how it reached its recommendation is not permitted into the workflow, regardless of how capable it is. A vertical product that has built the audit trail and the explanation layer can go where horizontal tools cannot.

For builders, the implication is to choose vertical as an intentional strategy, not as a niche hedge. Vertical is not smaller than horizontal. It is often more defensible, more capital-efficient, and more durable. The category leaders in vertical SaaS, across the last two decades, have produced better equity outcomes than most horizontal players when measured on capital in to market cap out. The AI era intensifies this pattern, because the depth dividend compounds with every round of domain engineering.

The temptation to avoid vertical strategies is usually explained as a total addressable market problem. It is almost always a strategic depth problem. Founders who go vertical have to know the domain deeply. Founders who go horizontal only need to know their go-to-market. The easier path is horizontal. The more durable outcomes are vertical. This has always been true. It is more true now.

For buyers in a vertical industry, the strategic implication is to stop expecting horizontal tools to solve domain-specific problems, and to buy, or build, vertical capability deliberately. The buyer who tries to lift a horizontal assistant into a deep vertical workflow, and who is surprised when the results are disappointing, is buying the wrong tool for the job. The buyer who finds, or funds, a specialist capable of the depth, and who is willing to invest in the integration, gets a significantly better outcome.

For investors, the depth dividend argues for a re-weighting of attention. The last decade’s bias toward horizontal plays reflected a real thesis, that distribution was the scarce asset. AI changes the scarce asset. In an era where anyone can build a plausible horizontal tool, depth in a specific industry becomes the scarce asset. The investors who rediscover this sooner will allocate better.

Horizontal AI platforms will be big. They will have massive IPOs, will capture most of the easy surface area of enterprise productivity, and will, no doubt, win a lot of deals.

The question is where the durable, premium-margin enterprise software value will concentrate. And the answer, based on the mechanics rather than the marketing, is in the companies that did the unglamorous, decade-long work of going deep into specific industries. The horizontal platforms will be valuable. The vertical specialists, in many domains, will be more valuable per dollar invested, because they occupy positions that cannot be replicated cheaply from above.

The industry narrative has not caught up with this yet, but it will. Somewhere in the next two to three years, a handful of vertical AI-native companies will print outcomes that force the reprice. The builders and investors who saw it earlier will be positioned for it.

In the next piece in this series, I want to take on an implication that cuts across both horizontal and vertical plays. After twenty years of buy-not-build, enterprises are about to build again. The reasons, the shape of it, and the categories where this matters most are worth going through carefully.