A founder announces, with something close to pride, that the platform runs entirely on code an AI assistant wrote — zero lines by hand. The project doesn't survive contact with real users. Not because it lacked customers. Because nobody had checked whether the thing could hold up to the standard it needed to meet, and by the time that became obvious, there was nothing underneath to fix. The hard part hadn't been skipped. It had been automated straight past, without anyone noticing it was there.
That's not really a story about bad code. It's a story about what happens when the thing you've spent your whole life getting good at stops being the bottleneck.
Optimisation is a reflex, and it's an old one. "Make this better" and "stay alive" were, for a long time, close enough to the same instruction that there was no reason to separate them. The tribe that found the shorter route, the sharper blade, the more efficient harvest, outlasted the one that didn't. That pressure didn't let up for the length of an empire or a career — it's old enough to have written itself into how a brain works by default. We didn't choose to be optimisers. We were the ones who survived being anything else.
What never had the same pressure behind it was a separate skill: deciding what was worth doing in the first place, independent of whether you were capable of doing it well. That question barely had room to exist. If you couldn't afford it, that was your answer. If it took a year to build, the year itself forced you to sit with whether it mattered — the cost of trying was a built-in filter for the worth of the thing. Effort and judgement were bundled together so completely that most people never had to notice they were two different questions.
AI unbundles them. It performs our historical trained half — the optimising, the executing, the making-it-better — at a level that costs almost nothing and takes almost no time, as we go into the future. A functional product that once required six months and two hundred thousand dollars can now take six weeks and five. Gartner expects more than forty per cent of agentic AI projects to be shut down by 2027 for one underlying reason: the build succeeded and the case for it hadn't.
That's not quite the same unbundling delegation used to do. Handing work to a person still ran through someone's judgement — you chose who to trust, briefed them, saw what came back, could catch a mistake before it went anywhere. The checkpoints weren't designed in; they were just what happens when work moves at the speed of a human. AI doesn't remove one checkpoint. It removes the pace that created all of them, because it executes at the speed and scale of the whole organisation at once. For a founder working alone, the gap that opens is personal — deciding what's worth building. For a company, it's structural. Nobody built the layer that friction used to provide for free: who has authority over what the model touches, who answers for it when it's wrong, whether anyone could even trace how it got there, how the data feeding it is handled, what happens to roles once a system absorbs the work a team used to do. A cancelled agentic project, most of the time, isn't a failed product. It's a control plane that never got built, because nothing forced anyone to build it until the thing it was meant to govern was already running.
Here is the part that's easy to miss, because it doesn't feel like loss. When a skill goes from scarce to abundant, it stops being where the advantage lives. Nobody wins by being better than their competitors at something everyone now has for virtually free. The advantage moves to whatever is still scarce — and the thing that's scarce, suddenly and starkly, is the muscle nobody historically had to build until now: judgement about what's worth doing at all. It's a market repricing: the currency you've spent your working life accumulating just had its value cut to zero, and the new currency is one almost nobody has practised holding.
You can watch the lag in real time. Companies that lose confidence in one metric don't usually sit with the harder question underneath it — they reach for a new metric instead. Teams unsure how to measure whether their AI adoption is actually working have started tracking token consumption, as an example: how much text goes in and out of the model. An engineer who writes a precise prompt and solves the problem in one exchange looks, on that dashboard, worse than one who pastes in a bloated file and burns through ten times the tokens to get the same result. The efficient engineer is being penalised for being efficient, because efficiency was never actually being measured — token volume was standing in for it, and Goodhart's Law does the rest. It's a company following the same old habits — find the number, optimise the number — at the exact moment those habits stopped pointing anywhere useful.
The dating safety app that got breached because someone left the database on its default settings, wide open, is the same story from a different angle. Nothing there failed to look finished. The demo worked. The metric that mattered to the pitch — time to launch — was hit perfectly. The question nobody had been forced to sit with, because nothing was slow enough to force it, was whether the thing was actually safe. That question doesn't show up on a build timeline.
None of this makes optimisation a bad instinct. It's the instinct that got us here, and it isn't going anywhere. But it was never built to answer the question it's now being asked to answer, and pretending otherwise — pretending that shipping fast is itself a form of having decided the thing is good — is how you end up standing in the wreckage of something that worked perfectly and shouldn't have been built.
That's the trade actually on offer, and it isn't the easy one it's being sold as. Optimisation doesn't disappear — it just stops being where the scarcity lives. The scarcity moves to the question optimisation always let us skip: not how well something can be done, but whether it's worth doing at all. That is not a technical question. It's one of values, of purpose, of judgement — and none of those get easier just because everything around them got faster. If anything, they get harder.