Writing code was never the slowest part of software development. It just felt that way, because it was the part with the most visible effort. Tools like GitHub Copilot, Cursor and Claude Code have made that part dramatically cheaper, and the slow part has not gone anywhere. It has simply become the only part left.
The reviewer is now the rate limiter
A pull request still needs a human who understands what the code is supposed to do, what it touches, and what it might break. That work doesn’t compress. If anything it gets harder, because AI-generated code tends to look finished. It is well formatted, sensibly named and confidently commented. The usual tells of a rushed junior developer, like messy structure and half-deleted experiments, are gone. A reviewer has to find the flaw in code that looks like it has none.
That’s a different skill from reading code a colleague wrote after a long argument in a design meeting. When a person writes a change, you can usually ask why. When a model wrote it and the author only half-read it, the honest answer to “why is this here?” is often “it seemed to work.”
Throughput is not the same as progress
Most engineering teams measure output in things that are easy to count: PRs opened, tickets closed, lines changed. An assistant inflates all of those immediately. It does nothing for the thing that matters, which is whether the system is still understandable six months from now. I’d argue that’s the real cost. Not bugs, which tests can catch, but code that nobody on the team has a mental model of.
Code that nobody understands is a loan. It gets repaid later, with interest, usually during an outage.

What should change
Teams adopting these tools should put the saved time into review rather than into more output. In practice that means smaller changes, a firm rule that the author must be able to explain every line they submit, and tests written before the generated implementation rather than after it. AI-assisted review bots can help catch obvious problems, but treating a model as the sole check on another model’s work is just circular.
The companies selling these assistants have every incentive to talk about speed, because speed demos well. Nobody puts “our users now spend more time reading” in a launch video.
A small note on juniors
Review is also how junior engineers learn. Starve it of attention and you starve them of that.