Durable artificial intelligence investment is not flowing toward the cleanest markets. It is flowing toward the ones where the work is genuinely unpleasant to automate, where buyers have been burned by generic tools before, and a wrong answer carries real operational cost. Sapphire Ventures published a detailed piece this week by partners Cathy Gao and Aditya Reddy laying out exactly why. I follow both authors. The analysis is written for founders evaluating markets to enter. But the argument it makes has a buyer consequence that the document does not spell out, and that is what I want to address here.
Founders care whether a market is defensible to build in. CIOs and CTOs need to care whether it is defensible to exit. Those turn out to be the same structural problem, just read from opposite ends of the contract.
A personal note before the analysis. In the early 2020s I worked with a Sapphire portfolio company that was already using machine learning to analyze and predict revenue and retention, well before generative AI had a brand name or a cultural moment. The return on investment was real and measurable. What made it work was not the model. It was how deeply the system had learned the specific shape of that business's customer data. That experience is why the Sapphire thesis on workflow grit as a moat reads as confirmation rather than novelty to me. I had seen the pattern. I just did not have a name for it yet.
If you are not already following Cathy Gao and Aditya Reddy at Sapphire, I would encourage it. The same goes for LLR Partners, where Emily Oakes publishes research worth tracking closely, and for Bessemer Venture Partners, whose published memos consistently surface structural shifts before they become consensus. These are not marketing newsletters. They are working documents about where enterprise software is actually heading.
Operational Complexity Is the Moat, Not a Limitation of It
Clean workflows get commoditized fast. A narrow, well-defined task is easy to demo and easy to replicate, and once intelligence becomes portable enough, a foundation model provider simply absorbs it into a general-purpose product. The operational specificity that makes a workflow painful to automate is precisely what makes a successful automation durable. That is the structural argument at the center of the Sapphire piece, and it holds.
Salient builds AI voice agents for auto lending debt collection. Each call operates under the Fair Debt Collection Practices Act, the Telephone Consumer Protection Act, and Regulation F simultaneously, where a single procedural error can trigger regulatory action at the state or federal level. The system has to negotiate payment terms in real time, track call frequency limits, and hand off to a human agent when the situation falls outside a defined boundary. Per figures cited in Sapphire's analysis (unaudited), a human collections call costs between $4 and $12. An AI-handled call costs a fraction of that. The economics are not subtle. But the execution surface is large enough that no competitor can shortcut it by licensing a better base model.
Charta Health works in medical billing, automating pre-bill chart review across payer rules, Current Procedural Terminology codes, and denial patterns that shift by specialty and geography. HappyRobot and peers in freight logistics handle carrier and shipper coordination across voice calls, emails, and portal updates, dozens of manual touchpoints per load, in an industry Sapphire estimates spends over a trillion dollars annually on non-physical operational costs (unaudited).
None of those descriptions sound like a venture-scale market from a product pitch. That invisibility is structural, not accidental.
The operational messiness that hides these markets from casual competitors is the same force that makes exit costly for buyers once they have chosen. Grit cuts both ways.
Your Software Budget Is Not the Right Number to Look At
When enterprises evaluate vertical AI vendors, the default frame is software spend: what does the category cost today, and is this vendor's price reasonable against that baseline. In operationally messy markets, that is the wrong denominator entirely. The real money is in services and labor, the people doing the work, the outsourced providers handling overflow, the contractors filling gaps that the core system cannot.
EliseAI started with leasing automation for property managers, a market that looked bounded when sized by software spend. Once the product shifted from assisting leasing work to replacing it, the addressable revenue per customer changed. The system then expanded into maintenance coordination, collections, and AI-guided property tours across the tenant lifecycle. Sapphire cites figures (vendor-cited, unaudited) showing EliseAI now serves one in eight U.S. apartments, with operators spending well into the millions annually for the platform. The company has since moved into healthcare, targeting $600 billion in annual administrative costs using the same approach.
A property management operator might spend $30,000 a year on leasing software and $300,000 on leasing staff. Once the product is doing the work rather than supporting it, the vendor's addressable spend per account is no longer competing against the software line item. It is competing for the labor budget. That is a potential 30-times expansion in what the vendor can charge, with the same customer, without adding a single new account.
Enterprise procurement teams rarely model that trajectory when signing the initial agreement. They should.
Fragmentation Protects the Incumbent, Not the Market
The U.S. tax and accounting market runs to $145 billion by Sapphire's figures (unaudited) and is spread across roughly 46,000 CPA firms, 86 percent of which have fewer than 10 employees. Blue J, an AI-powered tax research platform, has found traction across that long tail and the large national practices simultaneously, now serving more than 2,800 organizations with usage reported at over 700 percent year-over-year growth (vendor-cited, unaudited). The fragmentation that makes the market unattractive for a horizontal vendor to chase is the same fragmentation that gives a purpose-built system years to compound operational context before anyone credible shows up to compete.
Horizontal AI vendors need concentrated, high-value customers to make their go-to-market economics work. When the revenue is distributed across thousands of small operators running different systems on inconsistent data, a generalized player cannot justify the sales investment. A vertical system built to work market by market can. And because the buyers in these markets rarely develop engineering capacity over time, the structural conditions that created the opening tend to persist. The window does not close just because AI becomes more capable.
This is also the real constraint on Anthropic, OpenAI, and similar foundation model providers as a competitive threat to vertical systems. They are simultaneously advancing model capability, managing token-based revenue economics, and facing pressure from enterprise customers as agent adoption scales. Building high-quality, operationally specific applications across dozens of complex verticals at the same time is a genuine organizational problem, not a hypothetical one. Vertical AI wins in these markets by outexecuting on focus.
The Switching Cost Nobody Prices In
When a vertical AI system has been running inside a business for two years, it has encoded something that a contract termination cannot transfer: the specific shape of how that workflow actually runs. The exception handling and approval hierarchies that took months to get right, the edge cases that required a workaround in month three and quietly became standard operating procedure by month eight. A competitor arriving with a better base model cannot replicate that without starting the encoding process from zero.
Removing the system means rebuilding the staff capacity that was eliminated, reconstructing the process documentation that was never written down because the system handled it, and absorbing an operational disruption that ripples well beyond the technology team. That is not a contractual switching cost. It is structural, and it compounds with time.
Most vendor evaluations do not model this. The RFP asks about integration complexity and contract pricing. The question that rarely gets asked in year one is what exit looks like in year three, who owns the workflow context the system has accumulated, and what operational continuity actually requires if the vendor is acquired, pivots, or fails.
By the time that question becomes urgent, the answer is usually expensive.
Before signing a multi-year agreement with a vertical AI vendor in a fragmented, operationally complex market, get specific answers to three questions your RFP almost certainly did not ask. What workflow exceptions has this system already encoded that a new entrant could not replicate in under a year? Is the vendor's pricing trajectory tied to software seats or to the labor budget it is displacing, and does your contract account for that shift? And if this vendor is acquired or changes strategic direction, what does operational continuity actually require, and who owns the context the system has built?
