Every few weeks another post arrives explaining that AI is transforming the tech industry. They all have the same shape: cloud, then software development, then chips, then cybersecurity, a section on ethics, a hopeful closing. I've written that post. You've read a dozen like it. The trouble with the genre is that it surveys everything and commits to nothing.
So here's the harder version: what actually changed in how we build software, what's mostly a sticker, and what the real costs are.
The part that actually changed my week
The most concrete shift is that typing code stopped being the bottleneck. Coding assistants — Copilot and everything that chased it — didn't replace engineers. They deleted a category of friction: boilerplate, the third CRUD endpoint that looks like the first two, the regex you always re-Google, the test scaffolding nobody enjoys writing.
I was skeptical, and I'm still skeptical about parts of it. But I'd be lying if I said my day looks like it did in 2022. The interesting part isn't that a model can write code. It's what that does to your attention: you type less and spend more time deciding whether what came back is correct — a different skill, and a more tiring one. Reviewing code you didn't write, all day, is its own kind of fatigue.
That's the honest version of "AI in software development." Not a revolution. A reallocation of where the effort goes.
The part where "AI-powered" means almost nothing
Then there's the other category: features that got an "AI-powered" label because a slide deck needed one. A ranking heuristic that's been running since 2016, rebranded. A rules engine with a language model bolted to the front so it can say the word. Some of it is genuinely useful. A lot of it is the same software with better marketing.
This matters because it makes the real changes harder to see. When everything is "AI-driven," the term stops carrying information. The useful question was never "does it use AI" — it's "what does it do that the previous version couldn't, and at what cost."
Where it genuinely earns the headline
Some of the survey is real. Anomaly detection in security and observability is a legitimately good fit — you're hunting for the unusual in enormous streams, which is exactly what these models are good at and exactly what humans are bad at doing at 3am. Cloud providers using learned models to schedule workloads and predict failures is real and mostly invisible, which is how you know it's working. And AI-assisted chip layout — search over a design space too large to explore by hand — is one of the quiet, genuinely impressive results, even if the headlines oversold "beating human engineers."
Notice what these have in common: the problem is scale and pattern, not judgment. That's the through-line for where this technology actually lands well.
The costs that don't make the slide
The ethics section of these posts is usually three bullet points and a promise to be responsible. The costs are more specific than that.
Bias isn't a footnote; it's a property of training on the world as it is rather than as you wish it were, and it surfaces in exactly the systems you'd least want it in. The data hunger is real too — a lot of "smart" features are smart because they consume an amount of user data that would have been a scandal a decade ago, and we've mostly stopped noticing. And the job question deserves more honesty than "automation creates new opportunities." Some of it does. Some of it just removes a rung from the ladder junior people used to climb. Both are true, and pretending only the comfortable one is true helps no one.
What I actually believe
The paradigm-shift framing is half right. Something real happened. But it didn't happen evenly, and the survey format flattens the one thing worth knowing: this technology is very good where the problem is scale and pattern, and much weaker where the problem is judgment and context — which is still most of the interesting work.
The companies that do well with this won't be the ones that adopted the most AI. They'll be the ones that were honest about which of their problems were the scale-and-pattern kind. The rest bought a sticker.
