The Surprising Gap Between Writing Code and Shipping Software
If you have spent any time around programmers lately, you have probably heard the excitement about AI coding assistants. Tools that can spit out hundreds of lines of working code in seconds sound like a dream. But anyone who has actually shipped software with these tools knows the catch: that code still has to be checked, and checking it takes real time. A new study now puts hard numbers behind that frustration, and the results are a little humbling for anyone who assumed AI would instantly multiply software output.
The research looked at how engineering teams actually work rather than how they feel about their tools. Harvard researchers Fiona Chen and James Stratton analyzed data covering roughly 300 million work events, such as code commits and pull requests, drawn from more than 700,000 employees at over 700 software development firms. The window ran from 2021 through March 2026, which means the findings reflect the real AI boom rather than a lab demo.
Why Human Review Became the Traffic Jam
The headline finding is that code review has become a bottleneck. Writing code got faster, but the efficiency gained at that stage was soaked up by everything that comes after it. As review queues grew longer, pull requests were more likely to bounce back for revisions, and reviewers left more comments per submission. In other words, the factory floor got quicker, but the inspection station did not keep up.
This matters because it cuts against a popular assumption that AI would quickly shrink engineering headcount or sharply lift the number of finished products. The researchers found little evidence that firms increased software output or reduced employment as a result of adopting these tools. Speed at the keyboard is one thing. Confidence that the code is safe, correct, and maintainable is another, and that confidence still comes from people.
What This Means for the Tools You Use and Buy
For consumers, this story is more than an inside-baseball debate about developer workflows. Every app, smart device, game, and online service you rely on is built by teams like these. When review becomes the choke point, release schedules stretch, bug fixes take longer, and the gap between a promised feature and a working one can widen. The next generation of tools will likely focus less on raw code generation and more on helping reviewers catch errors, test automatically, and understand changes faster.
That shift is already shaping the market. Expect to see review-focused features, automated testing platforms, and AI assistants designed to explain code rather than just write it. For shoppers and tech adopters, the lesson is simple: the most valuable AI productivity tools may not be the flashiest code generators, but the ones that make human judgment faster and more reliable. Keep an eye on products that promise quality and verification alongside speed, because that is where the real gains are likely to show up next.
