The Vertical Software Reckoning

Language models make some familiar software advantages cheaper to reproduce. The durable advantages are the ones rooted in data, regulation, transactions, and participation.

February 2026

I recently read Nicolas Bustamante’s account of ten years building vertical software. He organizes the traditional defenses of companies such as Bloomberg and LexisNexis into ten moats, then asks what happens to each one when language models can read documents, express domain rules, and operate software.

The useful part of this framework is that it avoids treating every vertical software company alike. AI changes which parts of a product are difficult to reproduce, so its effect differs across the category.

This scorecard is the spine of Bustamante’s argument. The five moats on the left were expensive ways of packaging knowledge and access. The five on the right depend on assets, obligations, or participation that a model cannot produce on demand.

What used to be expensive

A vertical product often earned its place by gathering several kinds of friction into one system. It offered an interface that experienced users learned, encoded industry rules in software, normalized awkward public data, and assembled many small tools into a bundle. Recreating the product required engineers who understood both the technology and the domain.

Language models reduce some of those costs at once.

Learned interfaces. A user who once had to remember product specific commands can describe the desired result in ordinary language. The old interface may still be efficient for experts, but familiarity alone is a weaker switching cost when an agent can operate several systems on the user’s behalf.

Encoded business logic. Domain rules no longer have to begin as application code. A knowledgeable practitioner can describe a method, examples, and exceptions in text that both people and models can read. Reliable implementation still requires tests and engineering, especially around consequential decisions, but the first translation from expertise to working software is much cheaper.

Access to public information. Parsing filings, court documents, forms, and reports used to require a collection of custom extractors. Models can now do a surprising amount of that work directly. A company whose main advantage was turning public documents into searchable fields has less room between the raw source and a capable competitor.

Bundling. A suite once won because each added module saved the customer from integrating another vendor. An agent can become the integration layer. It can gather data from one service, apply a method from another, and present the result through one conversation or workflow.

Scarce technical talent. Vertical software used to require a rare combination of domain knowledge and engineers able to encode it. Models let practitioners express more of their method directly and let smaller teams build the surrounding product. Engineering still matters, but the domain expert no longer has to communicate every rule through a long product and development chain.

None of these advantages disappears completely. A well designed interface, tested business rules, clean data, and integrated workflows remain valuable. The change is that competitors can reproduce more of them without rebuilding the original company’s entire software stack.

What remains difficult to copy

The other moats are tied to assets or relationships that a model cannot infer from public text.

Proprietary data. A live record assembled through customer activity, private transactions, or expensive collection remains scarce. Models may increase its value because more products can use it. The important distinction is whether the data is truly unavailable elsewhere or merely inconvenient to parse.

Regulatory position. Certification, validated processes, contractual responsibility, and the cost of switching a critical system do not vanish when a new interface appears. An agent may improve the experience around a regulated product while leaving the certified system underneath it.

Network effects. A communication network is valuable because the other participants are present. Natural language does not recreate those relationships. It may give the network a new interface, but the network remains the asset.

Transaction infrastructure. Software that moves money, originates a loan, adjudicates a claim, or executes a trade occupies a real position in the transaction. An agent can initiate the action. It still needs the trusted rails that complete it.

Systems of record. The authoritative history of a customer, patient, asset, or account is hard to replace because other processes depend on it. Agents can read across several systems and create a more convenient working view. That may weaken the system’s control over the user interface without immediately replacing its role as the official record.

The customer relationship may move upward

This suggests a change that is subtler than wholesale replacement. The vertical application can remain in the stack while losing the place where the user begins. If an agent understands the request, chooses the data sources, and coordinates the workflow, the agent owns more of the customer relationship. The underlying products become services it calls.

That is good news for a provider with unique data or essential transaction rails. More agents can bring it demand. It is uncomfortable for a provider whose pricing depended on forcing users through a proprietary interface to reach information available elsewhere.

A practical way to evaluate a product

I would start with five questions:

1. What does this product know that a capable competitor cannot obtain?
2. Which obligations or approvals make replacing it genuinely difficult?
3. Does it participate in the transaction, or only describe the transaction?
4. Does its value increase because other people or organizations use it?
5. If an agent supplied a better interface, what indispensable role would remain underneath?

The answers are more informative than asking whether the company has an AI feature. A chat box can be copied. Exclusive data, trusted execution, and a network of participants are harder.

The category is not disappearing

Vertical software still benefits from deep knowledge of a customer’s work. In fact, models make that knowledge easier to turn into a product. The result should be more competition and more specialized tools, not one general model replacing every application.

The distinction I would watch is whether a company owns something scarce or merely packages something that used to be difficult. AI is rapidly reducing the price of packaging. It has done much less to reduce the price of trust, exclusive information, regulatory standing, or a place in the flow of money.