OpenAI, xAI, Anthropic et al. are the sharks of the AI realm: massive, muscular, and moving fast. If you’re building an AI startup today, chances are you're swimming right alongside them, hooked into their APIs, drafting off their wake, and hoping not to get swallowed whole.

In short, you're a suckerfish. Or, if you prefer a more dignified title, a remora.

That’s not an insult. In the early days of an epochal shift like this one, symbiosis is a feature, not a bug. Startups win by positioning themselves in the slipstream, focusing on application layers, niche workflows, and underserved users while the sharks burn billions on core research.

The problem is, huge numbers of remoras are very much alike. If your AI startup is built entirely on piping API calls into a slick interface, you may find yourself dis-intermediated the moment the platform adds your feature natively. The challenge isn’t just building on AI. It’s building around it, in ways that are adaptive, sticky, and user-specific. 

Today, many AI startups are inference engines. They intake a prompt, daisychain some foundational models together, and output something useful. Alas, useful is not the same as defensible. To build a moat, AI startups need to pipe in something else that foundational models cannot deliver: tacit knowledge. They need to understand how users work, not just what they ask for. The real unlocks of vertical AI will be found in the tricks, shortcuts, and unspoken workflows that power real productivity. The winners will train systems that don’t just respond, but mirror, amplify, and adapt to those human patterns.

The future isn't a lone GPT clone sitting in a chat window. It is a chorus of well-orchestrated agents: one quietly monitoring how you draft a report, another flagging when you veer from best practices, a third suggesting a better phrasing based on team norms. While each agent is narrow-minded individually, together they form a composite intelligence. Over time, this swarm becomes your edge not because it knows more than GPT, but because it knows you better than GPT ever will.

Defensible AI startups don’t just ship features. They ship loops. They observe how users interact, where they hesitate, override, or ignore the assistant. They adapt micro-models or routing logic based on usage data. They reflect progress back to the user in ways that build trust. 

All of this compounds over time. The more the system learns, the more useful it becomes, and the harder it is to rip out. For example, a truly sticky radiology AI wouldn’t just transcribe images; it would mirror how radiologists annotate edge cases. That tacit knowledge is captured in feedback loops, not in the base model.

If you can close that loop faster and more tightly than a general-purpose platform, you’ve built something powerful. Of course, data scarcity, agent brittleness, and evaluation remain open challenges, but that’s precisely why domain-specific loops are so invaluable.

What the Sharks Can't (or Won’t) Do

The biggest AI companies struggle with the kind of intimate workflows that vertical AI companies are uniquely positioned to serve. Foundation models must build for everyone; vertical AI startups can build for someone.

That opportunity is not new. For all the architectural novelty AI introduces, the underlying fundamentals remain stubbornly familiar. Winning products do not begin as clever applications of novel technology. They begin as adaptive solutions to deeply understood problems in specific fields. They succeed because they are embedded inside operational pain, business logic, and the inarticulate friction of how work really gets done. Hype aside, the AI layer doesn’t change this key tenet of the software landscape. It simply raises the bar. 

This is the work of serious software companies: to observe real workflows, to mirror and refine them, and eventually to abstract them into infrastructure that no longer feels optional, to the point that a noun (google, slack) becomes a verb. The assistive layer becomes invisible because it flawlessly reflects how the user thinks. It conforms to (and reshapes) the work as it actually unfolds.

The largest AI companies will continue to improve the core primitives (language, vision, code, speech), but they will not solve for the specificity of all use cases, despite what AGI’s most fervent evangelists might suggest. Sharks are generalist by design. Their incentives and architectures are built around breadth. Remoras, by contrast, can privilege depth. They can choose a problem worth solving, one constrained by domain logic, shaped by tacit knowledge, and riddled with inefficiencies that resist horizontal solutions. 

That’s why we tend to get most excited not by companies that introduce themselves as "AI startups" per se, but by those that arrive with a real problem in hand and persuasively describe how AI helps them solve it better. These are not abstractions in search of traction. They are applications in motion, pulling in AI where it counts — sometimes at the core, often at the edge. 

The test isn’t whether you’ve added AI. It’s whether, without your product, the workflow remains broken. That’s the difference between a remora with staying power and one that gets shaken off at the shark’s first turn.