I came to AI from corporate development, not up an engineering ladder. For a long time my work was the bridge between technology and the business: translating what a system costs, what it can do, and what it returns — in both directions. Building was what happened afterwards, done by someone else.
That has inverted, and the reason is economic, not technical.
The threshold moved
Every idea has a threshold: the point at which it becomes worth the effort to build. When that threshold sat at a quarter of engineering time, most ideas never got built. Not because they were bad, but because they never cleared the bar.
Agents lowered the bar. What used to cost a quarter now takes a weekend. For me that has a pleasant consequence: the part a CTO eventually gives away has become economical again. I write code I would have delegated three years ago.
That is the harmless half of the story.
The threshold was also a filter
Whatever sat in the backlog for seven months and never made it into a sprint was usually sitting there for good reason. Nobody had to make that call — the price made it. The question “is this worth a quarter of engineering?” prioritised without anyone having to prioritise. It sorted out the things nobody had asked for.
Remove the cost, and the filter disappears with it.
Which is exactly what is happening: products are being built that no customer asked for. Not out of stupidity, but because the only reason not to build them is gone. Cheap is not a justification. It is merely the absence of an obstacle — and anyone who mistakes a removed obstacle for an argument will now produce, very quickly, a great deal that nobody wanted.
The effort was never the problem. It was the excuse that spared us the question of priority.
Judgement is the new taste
Someone now has to make the decision the price used to make. Judgement — the ability to pick the right thing and let the rest die — is the work that became scarce. Not the building.
Writing code, organising information, pulling data into a usable shape: that I hand to agents. They are faster than I am, and getting faster. Two things I do not hand over.
Strategy. A model’s judgement is a compression of what already happened. It pulls toward the average of its training data — and in a decision that actually matters, the average is precisely wrong. A strategy that recommends the same thing it recommends to everyone else is not a strategy.
This is not a hunch. A study in the Harvard Business Review put six leading models against seven classic strategic trade-offs, across more than 15,000 simulations. They converge: 96 percent chose differentiation, 93 percent chose augmentation — almost regardless of the situation described. A model that advises everyone identically cannot be the one deciding what gets built. I took the study apart here.
Architecture. Here the mistake costs more than in code, because it surfaces later and stays longer. The dangerous part is not that models are wrong, but that they are plausibly wrong. A bad proposal looks like a good one until you have paid for it for two years.
What follows
The scarce thing is no longer the building. It is knowing what to build, and what it must never become. That stays human work, supported by AI research rather than delegated to it.
The cheap part is priced in. Anyone still selling it as an advantage is selling something everybody now has. The advantage lies in what is left once the price no longer decides for you.
I am not claiming this holds forever. Models will get better, judgement included. But as long as a model is fundamentally a compression of the past, it will land on the average. And decisions that matter never live at the average.
