Vertical AI defensibility begins inside the workflow
A vertical AI company is not defensible because it has put a chat box on an industry-specific website. It earns the claim by taking responsibility for a consequential job—and fitting into the real, often untidy process around it.
That is the difference between a market label and a product. Vertical SaaS is software built for a particular industry; vertical AI applies AI to industry-specific work. In its 2023 analysis, Index Ventures argues that specialized products can address precise business problems better than generalized tools, and that AI applications can benefit from understanding end-user workflows and industry-specific training data.
The investor’s first question should be less “How clever is the demo?” and more “What changes on Tuesday morning if this works?” Look for a recurring job that is slow, costly, error-prone, or difficult to staff. Then follow the product into production: the messy inputs, exceptions, permissions, integrations, and human sign-off that a demo tends to leave offscreen.
A narrow wedge can be exactly right. But a wedge is an entry point, not a moat. Index describes AI as a potential opening in vertical markets where incumbent AI features remain nascent, while noting that established vertical software leaders bring distribution, relationships, and proprietary data advantages.
The question is whether the startup uses that opening to become part of the customer’s operating routine—or remains a clever feature the customer can swap out.
The business model has to follow the work
AI companies get paid when customers value the outcome enough to keep paying for it. The sentence is simple; making it true is the work.
A vertical AI product might charge a subscription, price by usage, or tie fees to completed work or measurable value. Those are options, not evidence of a business model. A product handling a defined volume of tasks may suit usage pricing; a broader system may fit a recurring subscription. The right measure is the one a buyer can understand, approve, and connect to a result.
Index raises a specific concern about seat-based pricing: if AI reduces the number of end-user seats a customer needs, vendors relying only on seat revenue may have to realign pricing with the value generated.
That is not an argument against subscriptions. It is a reminder to inspect what the invoice measures. Is the customer buying access, transactions processed, time saved, lower costs, or a dependable result? As editorial diligence, ask whether the claimed value is visible to finance and procurement as well as the product team. Ask, too, what it costs to deliver: model and infrastructure expenses, human review, and implementation all count. A pilot proves little about durable economics until those costs and the path to repeat use are clear.
A model is not a moat
Models matter. They are also not, by themselves, a reason a customer will stay.
A useful way to assess defensibility is to look beyond the model at the surrounding system: the data the company can use, the product that turns capability into workflow, the speed at which customers see value, and the relationship through which the product reaches and serves them. Index’s analysis emphasizes proprietary industry data, the ability to train or adapt models against relevant datasets, application design, fast time-to-value, and direct, high-quality customer relationships. It also points to the opportunity to generate proprietary, industrywide datasets.
Those ingredients are not a moat by declaration. They are hypotheses to test. As editorial diligence, ask what data the company can lawfully access, what improves as customers use the product, and whether the feedback loop can operate without turning each deployment into a bespoke services project. Ask how the company handles customer data obligations and whether improvements can be retained and applied appropriately. These are questions for a particular company—not sourced claims about what every market permits or requires.
The distinction matters. Data access is not the same as a compounding data advantage. If the product collects information but cannot use it to improve the outcome, the “data moat” is just a noun. If every customer requires a different implementation, that may be a deliberate services-led entry strategy; it may also constrain software economics. The investor’s job is to find out which.
Distribution is part of the product
The best model does not win in a vacuum. The product has to reach the people who need it, fit the way they work, and earn the confidence of those accountable for adopting it.
Index argues that direct customer relationships and distribution matter alongside product, and notes that established vertical SaaS companies may benefit from existing relationships and data. Its analysis leaves room for either AI-native companies or incumbents to win, depending on data, product utility, and distribution.
For a specific company, diligence should map the route from first champion to sustained rollout: who uses the product, who approves it, and what can block adoption? In an enterprise deployment, a persuasive demo is an introduction, not a sale. The product still has to work for the people doing the job and satisfy the people responsible for security, compliance, and budget. That is an editorial lens for examining a deal, not a claim about a universal buying process.
The competitive comparison should be equally grounded. Can an incumbent add a similar feature? Perhaps. But “can copy the feature” is not the whole question. Does the startup own a deeper workflow, improve outcomes more quickly, or build a customer relationship that is hard to displace? And does the incumbent’s installed base help it adapt—or make change harder? Index notes both the advantages of established vertical software and the innovator’s dilemma that can arise when new entrants offer different ergonomics and pricing.
What would make the moat real?
Before calling a vertical AI company defensible, look for evidence, not a category label, in five places:
- Workflow depth: The product handles a consequential, recurring job, including the exceptions that make the job difficult.
- Customer value: Buyers can identify a meaningful improvement and have a reason to keep paying for it.
- Data advantage: The company has a lawful, durable route to relevant data and a credible mechanism for improving the product with use.
- Distribution: The company can reach customers and turn an initial use case into sustained adoption.
- Economics: Pricing reflects value customers recognize, while delivery costs leave room for a viable business.
This is an editorial diligence framework, not a claim that any one attribute guarantees defensibility. It draws on Index’s discussion of workflow knowledge, industry data, customer utility, and relationships.
The bar should be high. A vertical label is not a moat. Neither is a model, a pilot, or a large market on a slide. Show the workflow the company owns, the outcome customers value, the mechanism by which the product gets better, and a credible path to the next customer. Until then, defensibility is a thesis. With that evidence, it becomes an investment case.
No-text editorial illustration of vertical AI embedded in an enterprise workflow