The revolution will eat its young
A whole generation of AI applications is genuinely useful and should never become companies. Confusing those two properties is how the money gets lost.
The most useful AI applications I see are embarrassingly low-concept. A thing that summarizes the weekly calls into the format one specific team actually reads. A triage layer over one inbox. A report that used to take an analyst every Friday. Nothing about them would survive a pitch meeting, and they produce real money: hours back, errors down, decisions faster. The trap is the next thought, which arrives on schedule: this is valuable, so we should productize it.
Valuable and productizable are different properties
That inference smuggles in an assumption from the old economy: that value plus scarcity of builders equals a business. The second half collapsed. A product is value that's hard to replicate, and replication cost everywhere is heading toward an afternoon of prompting. Your prospective customer doesn't have to buy the thing that took you a weekend. They can have their own enthusiast build a worse version that fits them better, and a worse tool that fits is usually the rational purchase at price zero.
Worse, the low-concept layer is exactly the altitude the platforms are climbing through. Every model release ships with someone's startup as a feature. The summarizers, the doc-chat tools, the email drafters: each was a real company in 2023 and a checkbox in 2025. The pattern isn't bad luck. If your product's entire substance is "the model, pointed at a common need," then the model's next version is your competitor, built by the company you pay for inference.
If the model is your product, the next model is your replacement.
The taste defense, and why it's half right
The standard rebuttal is taste. The model is a commodity but the curation isn't: the opinionated workflow, the right defaults, knowing what to leave out. This is half right, and the half matters. Taste is a real advantage and a terrible moat. It differentiates you exactly until the platform copies your choices, which is cheap, legal, and routine, or until the model gets good enough to exercise the judgment itself, at which point taste has been trained on. Taste works as a moat only when it's compounding somewhere defensible: proprietary data the taste generates, a workflow position that's painful to leave, distribution the copier can't reach. Taste plus nothing is a head start, and head starts end.
What to do with valuable non-products
The conclusion isn't to stop building these things. It's to stop forcing them to be companies. For an operator, the internal tool that saves your team thousands of hours doesn't need a pricing page to be a triumph. Its return is leverage, not equity, and capturing value as leverage is allowed. Some of the best AI investments right now have an audience of forty people.
For a founder, the test is brutal but simple: describe your product without mentioning the model's capabilities. Whatever's left, the data, the distribution, the workflow lock-in, the regulatory soak, is the actual company. If nothing's left, you haven't found a startup. You've found a feature, and possibly a great consulting engagement, but the revolution eats its young, and it's hungriest for the ones who mistook being early for being defensible.