The most dangerous sentence in a startup is: it worked when I ran it.
I have said it. I have believed it while saying it. I have sat across from someone whose money I wanted, opened my laptop, typed a careful sequence of keystrokes I had practiced the way a card sharp practices the false shuffle, and watched the thing do the thing. The room leaned in. Somebody nodded. And in the warm glow of that nod I made the oldest mistake in building software, which is to confuse a performance with a possession.
A demo is a magic trick. I mean that as a compliment to magicians and a warning to founders. The trick is real — the coin does vanish, the card is your card, the model does answer the question correctly and even a little beautifully. But the trick lives inside a sealed environment that the magician built and controls absolutely. The lighting is right. The deck is the deck he chose. The volunteer is, often enough, in on it. Nothing in the room is allowed to be hostile to the outcome. A demo works for the same reason: you have removed every variable that wants it to fail.
A product is the same trick performed on a Tuesday, in the rain, by someone who has never seen you, on a deck of cards a stranger brought from home, half of them sticky.
The gap nobody budgets for
Between those two events sits a distance most teams never measure because they cannot see it from where they're standing. The demo is the visible nine-tenths of the iceberg, the part that photographs well. The product is everything under the waterline: the retries, the malformed inputs, the user who pastes forty thousand words into a box you sized for forty, the network that dies mid-request, the edge case that is not an edge for the one customer who lives there full-time.
You can feel the gap as a ratio. In my experience the demo is maybe a tenth of the work and nine-tenths of the applause. The product is nine-tenths of the work and, on a good day, no applause at all — just the quiet absence of complaints. Founders raise on the applause and then discover they have to fund the silence. That mismatch is not a detail. It is the shape of the grave a lot of companies dig for themselves.
This year the gap has gotten wider and harder to see at once, which is a bad combination. We are living through a flood of genuinely astonishing demonstrations. Drop a paragraph into a model and it writes you a function. Ask it a question in plain English and it answers in plain English, fluently, instantly, in front of investors who have never seen anything like it. The demos are not fake. That's the trap. They are real, and they are reproducible exactly once, under the exact conditions you chose, which is to say they have the precise reliability of a magic trick and we keep mistaking it for the reliability of a bridge.
I have shipped software that worked ninety-five percent of the time and felt, in the demo, like it worked one hundred. Ninety-five percent sounds like an A. Then you do the arithmetic of a real day. A user touches the thing twenty times before lunch. At ninety-five percent reliability, one of those twenty hands them a failure, and people do not remember the nineteen successes. They remember the one time the trick broke and the dove was clearly just a pigeon you'd been hiding up your sleeve the whole time. Trust does not average. It takes the minimum.
The last mile is most of the trip
There's a comforting lie in engineering that the demo is "ninety percent done" and the rest is polish. Polish is the wrong word. What's left is not polish; it's the part that makes the thing true rather than merely possible. Handling the input you didn't imagine. Failing in a way that doesn't lose the user's work. Being fast enough on the customer's slow laptop, not your fast one. Doing it again tomorrow after you changed three other things. None of that demos well, because none of it is visible until it's missing.
The economics here are genuinely strange and worth saying plainly: it is now almost trivial to produce the appearance of a working product and almost as hard as it ever was to produce a working one. The cost of the trick has collapsed. The cost of the bridge has not. Anyone can summon a convincing first ninety percent in an afternoon, which means the first ninety percent is no longer where companies are won or lost. The moat, such as it is, has migrated entirely into the last mile — into reliability, into the boring discipline of doing the same trick ten thousand times without flinching. I suspect we are about to watch a lot of well-funded demos fail to become businesses, and the postmortems will all say the same thing in different words: it worked when we ran it.
So how do you tell, early, whether you have a trick or a product? You stop running the demo yourself. Hand the laptop to the most distracted, least cooperative person you can find and say nothing. Watch them break it within ninety seconds, not because they're clumsy but because they're real, and reality is the one volunteer who was never in on the act. Whatever survives that is the beginning of a product. Everything before it was theater — useful theater, the kind that raises money and recruits believers, but theater.
I don't say any of this to make light of demos. A good demo is how an idea earns the right to become real. It is the spark, the proof that the universe permits the thing to happen at all. But the spark is not the fire, and you cannot heat a house with the moment a match is struck. The work — the actual work, the work nobody claps for — is keeping it lit.
A demo proves the thing can happen. A product proves it happens. The whole game, the only game, is the distance between can and does — and that distance is not measured in features. It's measured in Tuesdays.