The standard argument for AI-assisted development runs like this: developers spend too much time on routine tasks; AI handles those tasks; developers focus on higher-value work; output increases. This argument is coherent. It also requires the bottleneck to be execution speed rather than problem clarity, strategic direction, or organizational alignment.
Most engineering organizations are not primarily constrained by how fast code gets written. They're constrained by uncertainty about what to build, by organizational friction that slows decision-making, by misalignment between what engineering understands and what the business needs, and by accumulated technical decisions that make each new feature harder than the last. None of these constraints respond to faster code generation.
There is a worldview embedded in the productivity argument that is worth naming. A significant part of the technology industry operated on the implicit assumption that writing code was the hardest thing a human could do - the apex of cognitive work. The corollary followed naturally: once AI could program, it would obviously handle the lesser tasks that everyone else was doing. Product strategy, organizational clarity, customer understanding - those things were softer, less rigorous, and therefore easier. The people doing them were just waiting to be automated.
This turned out to be wrong in a specific way. The skills that constrain most engineering organizations are not easier than coding. They are different from coding. And "different" did not yield to the same approaches that worked for code generation. The bottleneck did not move.
What AI tooling delivers in these organizations is faster production of solutions to the wrong problem. The team ships more, learns less about whether what they're shipping is right, and accumulates velocity in a direction nobody has verified. Making a development team twice as fast means shipping the wrong thing twice as often. AI coding tools are the same dynamic with better marketing.
If your engineering organization is genuinely bottlenecked on execution speed, AI tooling will help. If it's bottlenecked on anything else, you should expect to spend significant money and get better-looking metrics on a problem you didn't actually have.
Thoughts? Find me on Bluesky.