Most engineering leaders adopted coding assistants expecting a throughput story. Two years in, the throughput is real and the story is more complicated. Teams are producing more pull requests, larger diffs, and more speculative branches than before. What they are not producing is more reviewers, more test infrastructure, or more architectural judgment. The constraint did not disappear. It moved.
This matters because the economics of software delivery were never dominated by typing speed. They were dominated by the cost of understanding a change well enough to accept responsibility for it. When a senior engineer writes a service, the review is partly redundant because the author already carries the model of the system in their head. When an assistant writes it, that model has to be reconstructed by the reviewer from scratch, against code that is syntactically confident and often subtly wrong in ways that compile cleanly. The reviewer is now doing the hardest part of the work with the least context.
The organizational symptom is a review queue that grows faster than the team can drain it, followed by a quiet collapse in review standards. Nobody decides to approve things they have not fully understood. It happens because the queue is twenty items deep, the diffs are eight hundred lines, and the author is waiting. The failure shows up months later as production incidents in code that passed review, which is the most expensive place to discover that a control has stopped working.
The teams handling this well have treated it as a platform problem rather than a discipline problem. They invest in what makes verification cheap: fast and trustworthy test suites, tight service boundaries so a change has a knowable blast radius, contract tests at integration points, and observability good enough that a regression announces itself rather than waiting to be found. They also set expectations about diff size and about which parts of the system an assistant is allowed near unsupervised. Payment logic, authorization, and data retention are not the same risk class as a reporting endpoint, and the review protocol should reflect that.
There is a governance dimension too. If an assistant contributed materially to a change, that provenance belongs in the record, because it changes what an auditor, an acquirer, or your own incident responder needs to check.
Before you buy more seats, measure your review latency and your test suite runtime. Those two numbers, not lines of code generated, tell you whether AI-assisted engineering is actually making your organization faster.
