Tech
AI Made the Typing Faster. Your Delivery Didn’t Speed Up — Because the Bottleneck Was Never the Code.
Your team got an AI coding assistant six months ago. Pull requests fly out faster than ever. So look at the number that actually matters: how long does a change still take to reach production? If you’re honest, it’s the same as last year — maybe worse. You poured rocket fuel into one cylinder of an engine and wondered why the car didn’t go faster.
The number nobody wants to check
Through June 2026 this has become the loudest argument on Hacker News and in engineering blogs — the year AI is finally asked to show the money. The thread that crystallized it was blunt: the bottleneck was never the code. InfoQ reported the same finding from inside Agoda, a company with thousands of engineers: AI coding assistants did not speed up delivery, because writing code was never the slow part. A change still takes six to twelve months to travel from a commit to a live system. What eats those months is infrastructure provisioning, testing, sign-off, change management, deployment scheduling — the parts no autocomplete touches. Typing was maybe ten percent of the journey. You optimized the ten percent.
The reflex here is to argue the tool isn’t the point — that AI also helps you think, design, explore. Fair. But that is a different claim from the one being sold, which was speed: ship more, ship sooner. Measure that claim and it dissolves. A developer who used to spend two days writing a feature now spends two hours; the feature still waits a quarter for the security review, the load test, the release window, the ticket that has to be approved by someone three time zones away. The hours you saved drain into a queue that didn’t get any shorter. The savings are real and local; the system’s output is set somewhere else entirely.
Why a faster node doesn’t make a faster system
Kevin Kelly spent Out of Control arguing that the interesting things in the world are systems, not parts — beehives, economies, ecologies, software at scale. A system has its own behavior, and that behavior is governed by how the parts connect, not by the speed of any single one. Make one bee twice as fast and the hive does not gather twice the honey; the colony’s output is set by foraging routes, comb space, the queen’s pace. Software delivery is exactly that kind of creature. It is a pipeline of coupled stages, and a pipeline moves at the speed of its slowest stretch — its constraint. The constraint, this whole time, was downstream of the keyboard. So when AI made the keyboard fly, the rest of the pipe didn’t notice. Worse: a faster source can flood a narrow channel and make the jam visibly worse, which is precisely what teams are now watching happen. Think of a highway where you widen the on-ramp but leave the same number of lanes. Cars merge in faster; the traffic doesn’t clear faster; it stacks up sooner. The ramp feels like progress to everyone standing on it. The drive home is exactly as long.
The bottleneck didn’t vanish — it moved upstream
Here’s the part that gets missed. AI didn’t remove the constraint; it relocated it. Pull the weight off typing and the strain travels to the two stages that resist automation hardest: deciding what to build, and trusting what got built. The spec — figuring out the right thing, in enough detail that a machine or a junior could execute it — turns out to be the actual work, and it needs human judgment. And verification turns out to be the other wall. Addy Osmani’s analysis put a number on it: median code-review time has climbed roughly 441.5%, with first-review and average review times both about doubling. Read that slowly. When a model can emit a thousand lines in a minute, the question stops being how fast can we write and becomes how fast can a trusted human grow confident this is correct. That confidence is the scarce resource now, and no copilot ships it. A reviewer staring at a large, machine-written diff is in a worse spot than before: more to read, less context on why each line exists, and the nagging sense that the author didn’t fully hold it in their head either. So review slows, the queue lengthens, and the change that was “done” in two hours sits for a week waiting for a human to vouch for it.
What this means for you
If your delivery hasn’t sped up, the instinct is to buy a better model. Resist it. There’s a line going around — you spent six figures on AI and the real bottleneck is a four-minute CI pipeline nobody bothered to fix. That sentence is the whole memo. Before you upgrade the part that was already fast, go find your actual constraint: walk one change from idea to production and time every wait, every queue, every approval. Then aim every improvement at the longest bar on that chart — usually a review queue, a flaky test suite, a manual sign-off, an environment that takes a day to provision. Speeding up anything else feels productive and changes nothing. The discipline is unglamorous and it is the entire game: fix the bottleneck you have, not the one your tools happen to be good at. And when you do clear the longest bar, don’t celebrate yet — the constraint just moves to the next-longest one, and the work begins again. That isn’t failure; it’s what managing a living system looks like. Kelly’s point lands here: you don’t finish a system, you keep finding where it’s pinched.
A pipeline moves at the speed of its constraint, not its fastest stage. AI made the fastest stage faster — so the system saw nothing, and the strain just piled up at the human gates of spec and review.
Stop optimizing the 10% that was never slow. Find the longest bar, and aim there.
Source: Kevin Kelly, Out of Control (a system’s behavior is set by how its parts connect and by its constraints, not by the speed of any single node). News peg: June 2026 debate across Hacker News (“the bottleneck was never the code”) and engineering blogs (frederickvanbrabant.com, May 15, 2026) on AI’s “show-me-the-money” year; InfoQ’s report on Agoda finding AI coding assistants did not speed delivery because coding was never the bottleneck; changes still taking 6–12 months from commit to production; and Addy Osmani’s analysis that median code-review time rose ~441.5%. Popular-science interpretation; figures are as reported.