The Best AI Businesses Are the Boring Ones
The AI work that demos well is rarely the AI work that compounds. Repetitive, rule-bound, universally disliked workflows have the highest margins precisely…
The AI businesses that get attention are the ones that demo well: the copilots, the chat interfaces, the generative tools that produce something impressive on a keynote stage. The AI businesses that compound are the ones nobody wants to talk about. Customs paperwork. Insurance audits. Document classification. Compliance checklists. The work is repetitive, rule-bound, and universally disliked, which is exactly why it has the best unit economics in software. Agents thrive on rules and repetition. The margin is highest where the glamour is lowest, and most builders walk right past it looking for something that photographs better. There is a quiet inverse relationship between how impressive an AI product looks and how good its economics are. The impressive products cluster in the open, generative, hard-to-constrain problems, the ones where the output is creative and the demo is dazzling. Those problems are crowded, because everyone can see them, and they are commoditized, because the underlying capability is a model that every competitor can also call. You end up competing on a thin layer of product on top of a capability nobody owns. The boring products cluster where the work is narrow, rule-bound, and repetitive. Those problems are uncrowded, because nobody wants to build customs paperwork automation, and they are defensible, because the value is in the accumulated handling of edge cases and the integration into a specific operation, not in the model. The dull problem and the durable margin are the same problem. The market's attention and the market's economics are pointed in opposite directions. The workflows people most despise share a profile: high volume, low variance, governed by rules, punishing when done wrong, mind-numbing when done right. That profile is misery for a human and ideal for an agent. Rules are exactly what an agent can encode. Repetition is exactly what an agent does without fatigue or drift in attention. The very properties that make the work unbearable for a person, sameness, volume, rule-density, are the properties that make it automatable to a high standard. The human cost of the work and the machine's suitability for it are correlated, which is why the most hated jobs are often the most automatable ones. The same correlation explains why these businesses produce real auto-resolution rates rather than demo-grade ones. A bounded, rule-governed problem can actually be solved to ninety-plus percent autonomy, because the space of cases is finite and the edge cases are enumerable. The open, impressive problems cannot make the same claim, which is why so many of them stall at the prototype. The practical move is to invert the instinct. When a founder asks what to build, the pull is toward the impressive thing, the thing that will demo well and pitch well. That pull leads directly into the most crowded, least defensible part of the market. The discipline is to ask a different question: what is the workflow in this industry that everyone hates, that runs on rules, that nobody would ever choose to build because it is too dull to put on a slide. That workflow, more often than not, is where the defensible margin is hiding, precisely because the dullness kept competitors away. Boring is not a constraint to apologize for. It is a moat. The unglamorous workflow is uncrowded because it is unglamorous, and uncrowded plus rule-bound plus high-volume is the recipe for the best economics in the category. We have built repeatedly in exactly these spaces, document classification, deal screening, and the compliance-shaped workflows that no one volunteers for, because that is where autonomous systems actually reach production and actually compound.