The Market bifurcated between vertical specialists with proprietary Data and horizontal tools heading toward zero

The products doing the most sophisticated work with AI right now are, paradoxically, the least threatened by it. And the simplest, friendliest, most user-loved tools on the market are the most exposed. That inversion tells you nearly everything you need to know about where software value is heading.

For years, the spectrum from horizontal to vertical was a smooth gradient. A CRM sat at one end. A fleet maintenance system for a specific aircraft type sat at the other. Everything in between occupied some point on the continuum, and the conventional wisdom was that horizontal tools had the larger addressable market while vertical tools had stickier customers. Both could prosper. Both had defensible positions.

That continuum just snapped in half.

The Data Moat

AI did not create a new kind of competition. It created a new kind of test. The test is simple: can your product's core value be replicated by an AI agent that has access to a general-purpose language model and an API? If the answer is yes, your product is now competing against a marginal cost that approaches zero. If the answer is no, you may be more defensible than you were before AI arrived. But most founders have not yet asked themselves which side of that test they sit on.

The products most likely to survive are the ones AI cannot replace because AI does not have their data.

At Boeing, I worked on products embedded in aviation maintenance workflows. These were not elegant consumer experiences. The interfaces were dense, technical, built for specialists who understood regulatory frameworks, fleet-specific maintenance histories, and safety-critical decision trees. The data inside those systems was accumulated over years of operation, tied to specific aircraft, specific routes, specific regulatory jurisdictions. No general-purpose AI model could replicate it because the data did not exist anywhere else. It lived inside the product, generated by the product's use, and was inseparable from the domain it served.

I remember a conversation with a maintenance engineer at a customer site who described their system as "ugly but irreplaceable." He was not being sentimental. He was being precise. The system's value was not in how it looked or how it felt to use. It was in what it knew. Decades of fleet-specific maintenance records, regulatory compliance histories, failure pattern data. An AI agent could be trained to understand aviation maintenance in general. But it could not replicate the specific data that this system had accumulated for this airline's specific fleet.

That is the data moat. Not a theoretical concept. A structural reality.

The products with proprietary, domain-specific data generated through years of customer operations are becoming more defensible, not less, as AI capabilities expand. AI needs data to be useful. But the data these systems hold is precisely the data AI does not have.

The Interface Vulnerability

But the other side of the split is brutal. Products whose primary value lies in providing a well-designed interface for tasks that AI can now perform directly through an API are facing an existential question. Not a competitive question. An existential one.

At Freshworks, I saw this positioning from the inside. The CRM competed heavily on user experience. Ease of onboarding. Clean interface. Intuitive workflows. Those qualities won deals. Sales teams would demo the product next to clunkier enterprise competitors and the reaction was immediate: this feels better. That feeling translated into purchase decisions, especially at the SMB and mid-market level where buyers had less patience for complex deployments.

But the interface was the product's moat. Not its data. The customer data lived in the customer's own account, portable, exportable, not structurally unique. What made the product sticky was that people liked using it. When switching costs were high and moving to a competitor meant retraining an entire team, that liking was worth a great deal. But what happens when an AI agent can interact with any CRM through an API, performing the same tasks a human sales rep performed, without ever seeing the interface?

When AI can operate any interface via API, the value of the interface converges toward zero.

That is the interface vulnerability.

The very quality that made horizontal tools popular, their ease of use for human operators, becomes irrelevant when the operator is not human. An AI agent does not prefer one CRM's interface over another. It does not experience frustration with a clunky workflow. It interacts through APIs, and at the API level, most CRMs look functionally identical.

The Split in Practice

The bifurcation is not theoretical. It is visible in real numbers.

Products deeply embedded in deterministic, high-stakes workflows are holding their valuations. But products providing a UI layer for tasks that AI can perform via API are seeing their multiples compress. The market is pricing the split before most product teams have acknowledged it exists.

I have watched this play out from both sides of the divide. At Boeing, nobody was worried about AI replacing the maintenance management platform. The suggestion would have been absurd. But absurd is not the same as irrelevant. It was absurd specifically because the regulatory context made it impossible. Every maintenance action on a commercial aircraft is governed by regulatory requirements so specific that a general-purpose AI, no matter how capable, cannot make the decisions those systems support without the domain data those systems contain.

But at Freshworks, the conversation would have been different. A CRM that wins on interface quality is a CRM that loses when the interface stops mattering. And the interface stops mattering precisely when the entity using it does not experience interfaces the way humans do. Nobody at Freshworks was having that conversation when I was there. The per-seat model was growing. The product was winning deals on user experience. But the foundation of that competitive advantage, the assumption that a human would always be the one interacting with the software, was never examined. Because it did not need to be. Until now.

Where the Value Lives

It is not about being vertical or horizontal in the traditional sense. It is about whether your product contains something that AI cannot access from the outside.

Proprietary data. Regulatory context. Domain-specific intelligence generated through years of customer operations. Those are the moats that hold. But they are moats you cannot build in a hurry.

The interface vulnerability does not mean that all horizontal tools will disappear. Some will adapt. Some will find new sources of value beyond the interface layer, perhaps by becoming the data layer themselves, or by embedding so deeply into customer workflows that the switching cost becomes structural rather than experiential. But the ones that remain primarily an interface for tasks AI can perform directly are on a trajectory that the market has already begun to price.

The data moat is not a strategy you can adopt in a quarter. It is the result of years of operating in a domain, accumulating knowledge that exists nowhere else, building products so deeply embedded in specific workflows that removing them would mean losing the data itself.

That is what defensibility looks like when capability is no longer scarce.

Some products were always going to survive this shift. They just did not know it yet, because the quality that made them resilient, their depth rather than their breadth, was the same quality that made them unfashionable during the years when growth meant going horizontal.