Satya Nadella said up to 30% of Microsoft’s code is now written by AI . If the software is writing itself, what happens to the people who design the systems it runs on?

I’m rarely optimistic (the curse of architecture, I guess), but I feel safe.

The sharper version of the argument is this: AI doesn’t need to become the architect. It just needs to make architectural mistakes cheap enough to ignore. If you can generate five design options in an afternoon and test them all, the upfront reasoning architects do starts to look like expensive ceremony. That logic has a flaw: it assumes mistakes are recoverable. In a greenfield SaaS app, maybe. In a regulated financial system, a distributed platform with ten teams, or anything touching data residency and compliance, a wrong call compounds in ways no iteration budget absorbs. Speed of recovery doesn’t make the original error free.

Architecture is forecasting behaviour, not just validating it. Spotting how a pattern will break under load three years from now, in a context nobody specified. That’s the part no checklist covers.

AI won’t replace architects, especially in large enterprises or complex systems. When you can’t satisfy all constraints, deciding which one to drop is not a data problem. It takes intuition and craft. Neither is trainable from a dataset.

Architecture is about designing something that fits a purpose. And to get that right, we go through endless rounds of refining requirements and making sense of them. AI will (and already does) make people lazy. We see it in the declining quality of code and content. Generation is the easy part. Qualifying what comes out, deciding what to keep, is where the judgment lives. Speed without that just accelerates garbage.

Frankly, no client is going to throw a vague “I need an application to handle my process and cost me less than now” at an AI and get a well-structured, future-proofed solution without hallucinations, or the weeks of review and orchestration effort that follow.

The tool earns its place. It’s already helping architects move faster, cutting through multiple technical domains, gathering insights, and automating the grunt work. That’s exactly what I’m doing right now. It’s making me more efficient, letting me focus on actual architecture instead of chasing documents and aligning contracts.

It makes me faster, which is useful precisely because I already know what direction to move in. The judgment that determines that direction is the part that doesn’t compress.

Large system migrations make the gap concrete. Aligning security, compliance, cost, and scalability is an exercise in competing priorities. AI can surface the options; it can’t weigh them. AI can assist, but it doesn’t have the strategic thinking to balance those forces effectively. It doesn’t see beyond immediate outputs, and it doesn’t think long-term.

The answer is in the question. If the software is writing itself, someone still has to decide what it writes, why, and what happens when it gets it wrong. That’s the job. It’s not going anywhere.