We Used to Pay a Premium for People Who Could Build Anything. AI Took That Job.

At AI Buddy, we spend our days inside the pipelines where software actually gets built, and the clearest signal from the last year is one most leaders still haven't said out loud. The single skill the industry paid the most for is the skill AI absorbed first. The scarce thing became cheap, the premium moved somewhere else, and a lot of very talented people are standing exactly where the value used to be, wondering where it went.

Think about the developer everyone wanted a few years ago. Any language, any feature. Hand them a rough sketch of an idea and they'd turn it into working software. They didn't need the problem to arrive clean. They took messy intent and translated it into something that ran, and that translation was rare enough that we paid a premium for it without blinking. The reason was structural, not sentimental. The bottleneck of the entire industry sat exactly where these people were strongest.

That premium is gone. And it left faster than most of us are comfortable admitting.

The build step went from scarce to nearly free

For thirty years the industry treated writing code as the slow, expensive middle of the lifecycle, and it poured staggering amounts of money into speeding that middle up. Better languages, richer frameworks, faster tooling. All of it aimed at one target: the cost of turning a decision into a working feature.

Then, somewhere around last November, the models got good enough that the target moved on its own. Hand a capable agent a clear, well-formed goal today and you don't get a snippet to paste and pray over. You get the thing. End to end, from intent to something that runs.

When the most expensive step in a system collapses to nearly free, the value it held doesn't evaporate. It moves. Everything uncomfortable about this shift lives in where it went.

The value moved from building to deciding what's worth building

The scarce skill isn't building the thing anymore. It's deciding what's even worth building in the first place.

And that's not a slogan, it's where the money went. GitHub's engineering team put it plainly last September: we're moving from code as the source of truth to intent as the source of truth. The goal you hand the agent, the definition of what correct actually means, is now the artifact the pipeline reads on every single generation. The code is just the output. Whoever owns the intent owns what gets built.

Andrew Ng has been making the same point from the product side. One of his teams recently proposed staffing two product people for every engineer, a ratio he admits would've sounded absurd a year ago. His logic is the same logic running underneath this whole piece. Once drafting code is cheap, the constraint moves to whoever can decide what's worth drafting at all.

The engineers feeling this most sharply are the ones whose entire professional identity was execution. Give them a defined task and they were untouchable. Now the defined task is the easy part. The market didn't stop valuing them out of cruelty. It stopped paying a premium for the translation and started paying it for the judgment that has to happen before any translation can begin.

AI can help you find the goal. It can't decide the goal for you.

This is the part the panic skips over. Give a capable model a clear goal and it's genuinely extraordinary. Give it no goal and it falls apart, and no amount of clever prompting rescues a decision that was never made.

A model optimizes for whatever target you hand it, and it holds no opinion on whether that target was worth hitting. Which goal actually drives revenue? Which bottleneck is really costing the business? Which feature will a customer pay for instead of just admire? Those are judgments about the market, the customer, and your own position, and the model carries none of that context unless you put it there.

It will absolutely help once you're inside the decision. It'll draft the options, cost them out, argue against its own suggestions, pressure test the goal you're leaning toward. What it won't do is tell you which problem is worth solving this quarter, because that answer doesn't live in the training data. It lives in your understanding of your own business.

We see this every week in our own pipelines. Hand the agent a vague goal and you don't get a clarifying question the way you would from a senior engineer. You get a confident, wrong build, because the ambiguity ships as a hallucination instead of a question. "The page should load fast" is a wish the model will quietly interpret however it likes. "Page load under two seconds at the 95th percentile on a slow connection" is a decision it can execute and a pipeline can verify. The gap between those two sentences isn't engineering. It's judgment, and judgment just became the scarce input.

Why the panic points the wrong way

The panic is real, and I'm not going to wave it away with optimism. When the skill you spent a decade sharpening becomes the cheap part of the process, fear is a rational first response.

But most engineers I talk to are mourning execution as if execution was the whole job, when it was only ever the visible part. The invisible part, deciding what deserves to be built and defining it precisely enough to be trusted, was always the higher skill. It was just buried under the labor of typing. Now the labor is gone and the skill is standing there in the open.

This is where I part ways with the dominant narrative, and where AI Buddy planted its flag early. Our mission is AI for employment, not replacement. When building gets cheap, the contraction-minded leader asks who can be cut. The expansion-minded leader asks what we can now build that was impossible last year. Which market just opened? What can we offer because the thing that used to take a quarter now takes a week? Jensen Huang said much the same at GTC 2026: leaders using AI as an excuse to shrink their workforce have simply run out of ideas.

The engineer who moves up into choosing the problem, defining the intent, and verifying that what got built is what the business actually needed doesn't get replaced. That engineer becomes worth more than they were when their edge was speed, because they now sit exactly where the new bottleneck lives.

What this means if you run an engineering team

Stop pouring money into making the build faster. It's already fast, and optimizing it further is optimizing a step that no longer costs you anything. The pressure moved to the two ends most teams aren't staffed for: the decision going in and the verification coming out.

Promote the people who are good at deciding what's worth building and defining it with precision, and give them the room to actually do it. That judgment is now the highest paid work in the building, whether or not your org chart has caught up. The engineer who can look at a business problem and shape it into a goal clear enough for an agent to execute and a pipeline to verify? That's your best builder, pointed one level up.

That's the work we do at AI Buddy, helping engineering organizations restructure around the decision and the verification instead of the vanished bottleneck in the middle, so the speed AI brought shows up as real outcomes rather than confidently built software nobody asked for.

The developer who could build anything was never valuable because they could type. They were valuable because they could turn intent into reality, and that part hasn't changed. What changed is that the reality got cheap and the intent became everything. The teams who understand this are about to build things the rest of the market can't picture yet.