Before I write a line of code for an AI product, I run the idea through one question. Does it do something that was impossible twenty-four months ago? If the answer is no, I am not looking at a company. I am looking at a feature wearing a company’s clothes, and whoever already owns the workflow will add it in a quarter and win. That is the whole filter. Everything else I check comes after it clears.
A wrapper is any product whose real value is a thin layer of interface over a general model. You take a task people already do, pipe it through the model, put a subscription around the output. That can be genuinely useful and still not be a business, because useful and defensible are not the same thing. When the value lives in the model and not in you, the company that already has the customers and the data copies your layer faster than you can pull away from it. They do not even have to win the user over. They already have the account.
I did not arrive at this from the sidelines. I founded an AI company while leading design, built the brand, the interface, and the agent pipeline end to end, and shipped it to people who pay real money. Everything here is what I learned watching my own ideas pass or fail this test, not from reading think pieces about moats. The filter got sharper every time I ignored it and paid for it.
The twenty-four month rule
The line I care about is capability, not polish. A model getting cheaper, faster, or more pleasant to talk to is not a new capability. It is a nicer version of something that already worked. If your product only became possible because inference got a little cheaper this year, you are standing on a feature, and the price of that feature is falling toward zero on someone else’s roadmap while you sleep.
The reason I anchor on twenty-four months is that the frontier moves fast enough to make it a real test. METR, a research group that measures what AI agents can actually do, found that the length of task a frontier model can finish on its own with 50 percent reliability has been doubling roughly every seven months (opens in new tab). Work that was flatly impossible two years ago is routine now. So the real question is not whether a model can help with your idea. It is whether your idea sits on a capability that crossed from impossible to possible inside that window, or on one that was already there and just got a coat of paint.
If it does not do something that was impossible two years ago, it is a feature, and the incumbent wins.
Where the wedge is
Once an idea clears the capability test, I look for the wedge. The wedge is the narrow, specific thing you can do that the incumbent cannot or will not, and it is almost never a better model. It is a workflow their own business punishes them for shipping. The big player with the distribution has a pricing model to protect, a support org not to overload, a brand promise that makes some experiences off-limits. The wedge lives in the gap between what the technology now allows and what the incumbent is structurally willing to do.
This matters because distribution is a real moat and the numbers are not close. In Menlo Ventures’ 2024 survey of enterprise AI buyers (opens in new tab), 64 percent said they still prefer buying from established vendors they already trust. If your only advantage is that your version is a bit better, you are asking a buyer to leave a vendor they already pay, to save a little, on a feature that vendor is probably already building. That is a losing pitch. The wedge has to be somewhere the incumbent will not follow you, not somewhere they are one release behind.
The only-AI-can-do test
The next test kills the most ideas on my own list. Could a team of people have done this at all, at any price, before? If the answer is yes, then what you are really selling is the same outcome for less money, and cheaper labor is a race the company with distribution and scale wins on price. If the answer is no, if the thing simply could not exist because no human process could produce it, now you have something worth building.
At Story Genie, the company I founded, agents write and illustrate a full personalized hardcover in about sixty seconds, and the customer sees the entire finished book before they pay a cent. That last part is the only-AI-can-do part. No human studio could ever show you a complete, custom, illustrated book about your specific child before you decide to buy, because producing it by hand would cost more than the sale itself. The model did not make an old process cheaper. It made an experience possible that could not have existed at any price. That is the difference I am hunting for.
The tell that you are on the wrong side of this line is retention, and the retention data on AI products is brutal. Sequoia found that generative AI apps have a median daily-to-monthly active user ratio of about 14 percent, against 60 to 65 percent for the best consumer companies (opens in new tab). People try these products and leave, because a feature earns a trial and only a product earns a habit. If the thing you built is genuinely impossible without AI and solves a problem people have on a schedule, they come back. If it is a wrapper, they churn, and no amount of growth spend fills a hole in the bottom of the bucket.
A feature earns you a trial. Only a product people cannot get any other way earns you a habit.
The real kill criteria
Here is the list I actually kill ideas on, before any code. I kill it if the incumbent could ship the same thing as a checkbox in their existing product, because they will, and their version comes pre-installed for everyone who already pays them. I kill it if the only moat I can name is the prompt, because a prompt is copied in an afternoon. I kill it if the value evaporates the day a better model ships, because then I am betting against the exact thing my product rides on, and that bet always loses.
The last criterion is the hardest and the most useful. I make myself name, in one plain sentence, the single thing this does that was impossible twenty-four months ago. If I cannot say it without hedging, without listing three small conveniences that add up to nothing, the idea is dead. A real answer sounds like a capability. A fake answer sounds like a feature list. I have talked myself out of good-looking ideas on that one sentence alone, and every time I skipped the exercise I ended up rebuilding on a foundation that belonged to someone bigger than me.
Who this is for, and when the filter is wrong
This filter is for founders trying to build a durable company, the kind that has to survive the incumbent noticing it. If that is you, run every idea through it and be ruthless, because the market is already full of well-designed wrappers quietly waiting to be absorbed. The demo working is not the test. Surviving the copy is the test.
It is the wrong filter in two cases, and I want to be clear about them. If you are building an internal tool, a wrapper is often exactly right, because you are not defending a market, you are giving your own team their hours back, and a thin layer over a model that deletes a weekly chore is a fine use of a Tuesday. I wrote about that kind of work in why execution got cheap and ideas got expensive. And if you already own the distribution, if you are the incumbent adding AI to a product people already use, then being a feature is the point. You are not trying to become a company. You already are one.
There is one more case, and it is the one that keeps me clear-eyed about my own rule. Sometimes the new capability is so large that the incumbent does not win, they get erased. In that same Menlo report, Chegg lost 85 percent of its market cap and Stack Overflow’s web traffic halved, both to AI products that did something their old models could not survive. Those incumbents had the distribution and lost anyway, because what changed was not a feature they could bolt on. It was the ground under their whole business. So the rule is not never bet against the incumbent. It is only bet against them when you are doing the thing they structurally cannot copy, and be clear-eyed with yourself about which one you actually have.
That clarity is the entire discipline. It is easy to fall for your own AI idea, because the demo works, the model is genuinely amazing, and the landing page looks like a company. The filter exists to slow that feeling down and ask the one question the excitement wants to skip. Does this do something that was impossible two years ago, in a place the incumbent cannot follow. If yes, build it with everything you have. If no, you found a feature, and the kindest thing you can do for yourself is admit it before you spend a year proving it.