Your developers code faster, the customer still waits just as much. The speed of AI stops at review and testing. It’s a system problem, not a tool problem.
I followed a request one day, from start to finish. The dev team processed it in four business days. The customer had waited forty-five days. The other forty-one, the request was sleeping in queues: prioritization, review, validation, behind other work. The team’s indicators remained green from start to finish.
This is not an anecdote. This has become the reality of almost every organization that has deployed AI in their development teams. Code has never been produced so quickly. The profession does not receive its functionalities any faster. And someone up there starts asking where the KING is.
I have been repairing tech delivery systems for fifteen years. Banking, industry, energy, public sector. All this, well before the first prompt. What AI exposes today, I already measured by hand ten years ago. The only difference now is that she has made the problem impossible to ignore.
The numbers confirm what I saw. Faros AI’s study of 22,000 developers shows median review time up 441%, and 31% more pull requests merging without any review. The DORA 2025 report states it bluntly: the adoption of AI increases throughput, and at the same time instability. His conclusion is in one sentence: succeeding with AI is a system problem, not a tool problem. The AI delivered on its promise perfectly. It accelerated the writing of the code, the only step that was already fast. The bottleneck just shifted.
So far, everyone is in agreement. The observation has been hanging around everywhere for a year. The interesting question is no longer the diagnosis. It’s what you make of it.
The two clocks
There are two clocks, and most development teams only watch one.
The first measures your team’s speed on an element, within its own process. The second measures customer expectations, from request to delivery. AI improves the first. The customer only lives in the second.
Let’s go back to the request I followed. Four days on the first clock. Forty-five on the second. The developer dashboard was reading the wrong one. Here’s the trap, and it’s mechanical: your velocity can increase while the customer waits longer. Your metrics say you’re winning while the customer disagrees. However, he is the one who pays.
The bottleneck rises
There is good news though. The bottleneck of the magazine is now attacked at the level of the AI tool, and by the very people who created this speed.
Take the case of OpenAI. Almost all of their engineers now use Codex, compared to just over half a few months earlier. And since they adopted it, they are merging 70% more pull requests every week. The human magazine could not absorb this volume. So Codex himself proofreads almost every PR before a person sees it. Anthropic placed a multi-agent review system in front of the flow. Other publishers build their own, dividing the work between coordination, implementation and verification agents.
Write down what they did. They didn’t buy one more tool. They redesigned their review system around the new speed. The creator of the tool faced his own constraint first, and he had the discipline to treat it as a system problem.
Now look what’s left when the agents are done. Judgment: what to build, what makes sense, what compromise to accept. This was the real constraint under the review process from the beginning. The machines take pattern recognition and verification. They don’t make the decision.
And that’s where most analyzes stop, concluding that we need to recruit more experienced profiles. This is a diagnostic error. Judgment is not a skill to be bought in the market. This is your new bottleneck. And this one cannot be recruited, it must be piloted. We design the system around it: we allocate it, we protect it, we concentrate it where it counts. This is exactly the work that Lean has been doing for seventy years. The constraint has changed place. The method for treating it is always the same.
Industrialize, but what?
The watchword for 2026 for AI is to move from experimentation to scale. The word is right. The target is wrong.
What needs to be industrialized is neither licenses nor POCs. It’s the delivery system itself. And this work is concrete, known, and predates AI:
- Map the entire flow from request to delivery, not just the coding stage
- Find the real constraints, most often the review, the tests or a validation door
- Limit outstanding amounts, so that queues stop growing
- Make each wait visible, so the team sees where the time is going
- Pulling workflow through the system instead of pushing more and more at the blind entrance
There’s nothing exotic about it. This is the work that halved the deployment time of a cloud team, and reduced the defects by 91% for a banking dev team, returning on schedule within months. The tools are new, but the way to manage flow through the system is not.
There is an important nuance to note in the market today. Integrators announce deadlines divided by three or four thanks to AI. Look at the scope: these are projects that they control from start to finish, with a production chain rebuilt around the tool. They don’t contradict the rule, they confirm it. The gain comes from the redesigned system, not the tool. The question that remains open is that of your internal teams, your legacy, with your specific constraints.
Is this your situation?
Read these five sentences, count the ones that are true for you:
- Your developers produce more code, but the business does not receive its features faster
- Pull requests pile up waiting for review
- Your velocity increases and your lead time does not decrease
- Management funded the AI tooling, and now someone is asking for ROI
- No one knows how to point out the precise step where time is wasted
If three of these statements are true, then your AI tools are not the problem. The system around it, yes.
The AI did its job. Writing code has never been faster. The next win won’t come from one more license or one more agent. It will come from a redesigned delivery system so that speed reaches the customer. The AI didn’t break it. She tensed him up just enough that you could finally see him.
