Tech • AI • Robotics
The moment the developer profession changed
7/10TL;DR
At Clever Cloud, intensive use of AI coding agents has pushed senior developers to become far more productive, but the gains came only after the company rebuilt testing, security and workflow practices around the tools.
KEY POINTS
A productivity paradox in AI
A February study by Tony Truand drew attention by reporting that 6,000 executives saw no measurable productivity impact from large language models. That finding clashes with experiences in parts of software engineering, where some teams report a sharp acceleration since late 2025, especially with coding orchestrators such as Claude Code and Codex. The gap suggests that generic chatbot deployment and deep workflow integration are producing very different outcomes.
From skepticism to adoption
Quentin Adam, head of Clever Cloud, said early code generation systems often produced unreliable output, including code that did not compile. He became less skeptical as models improved through better training and fast feedback loops, particularly once agentic systems began breaking tasks into ordered steps and working with local codebases. The turning point came when AI produced code he judged better and faster than his own, saving hours on a single task.
Why strict languages work better
At Clever Cloud, AI performed best when specifications were clear and the programming environment was highly constrained. Adam argued that models work more reliably with strict compilers and strong typing, because compiler errors create high-quality feedback the model can immediately use to fix code. He cited Rust as especially effective: the language is stable, memory-safe and backed by a compiler that gives detailed, actionable diagnostics.
A cloud provider testing AI on critical systems
Clever Cloud, founded more than 15 years ago, develops software for cloud infrastructure, from low-level operating system layers to managed services and internal business systems. The company also operates its own database stack built on FoundationDB. That makes it a demanding test case for AI-assisted development, since many of its systems are critical infrastructure where failures are costly and conventional assumptions held that AI-generated code would be too risky.
The rollout was cultural before it was technical
Adoption did not happen automatically. Adam described strong resistance from developers, including senior technical staff, and said persuasion often had to happen one by one. Once early holdouts saw AI complete work equivalent to five weeks in about two hours, attitudes shifted. The company then abandoned the idea of imposing a single enterprise tool, instead allowing staff to expense whichever AI products they found useful, while avoiding long annual commitments because the market was changing too quickly.
Senior developers benefited most
Contrary to the expectation that juniors would gain the most, Adam said senior engineers adapted fastest and saw the biggest productivity jump. Managers and tech leads who previously spent most of their time coordinating teams regained the ability to prototype and code directly. He said some senior staff described themselves as more productive than at any point in their careers, with cleaner code and faster formalization of architectural ideas.
Testing became the real center of work
The company’s biggest change was not simply generating more code, but surrounding code generation with heavy automation in unit tests, integration tests, security checks and simulations. Tasks once considered too expensive, such as building mocks or broad regression coverage, became worthwhile because they compressed the feedback loop. In that model, the durable asset is less the hand-written code itself than the specifications, architecture, test suite and validation environment.
AI is used for “inhuman” tasks
Adam argued that AI should not be treated as a junior developer but as a non-human system suited to non-human tasks. One example is meeting preparation across a monolithic codebase touched by roughly 30 people each week. AI can compare branches, scan failing tests, analyze pull requests, issues, messages and email, then identify tensions or architectural decisions that actually require discussion. He said work that would take three or four days manually can be done in about 20 minutes.
Security and legacy rewrites are central targets
Every incoming commit now goes through automated pentesting, and multiple models can be run against internal systems to expose vulnerabilities that may never have appeared in logs. Adam said that since opening those systems to AI-driven probing, the company has found numerous previously unnoticed flaws. He has also set a goal for 2026: to run only code produced in 2026, effectively eliminating legacy accumulated over 15 years, with AI used to rewrite or migrate old components far faster than before.
Limits remain, but they are uneven
Adam said gains vary sharply by domain. Some highly specialized areas, such as kernel-level work or custom network operating systems for switches, still produce weak results because training data is scarce and models default to more common software assumptions. In more standard development work, however, he described gains ranging from incremental improvements to cases that can be benchmarked at 10 times faster.
A wider sovereignty opportunity in Europe
Adam linked the AI shift to a broader reopening of competition in European digital infrastructure. Falling software costs and growing concerns over dependence on US providers have strengthened arguments for sovereign alternatives. He pointed to a recent European cloud procurement worth 180 million euros in which a consortium including OVHcloud, POST Luxembourg and Clever Cloud won part of the contract under a sovereignty framework designed to favor genuinely European operators.
CONCLUSION
The experience at Clever Cloud suggests that AI’s impact on productivity depends less on access to models than on rebuilding engineering, testing and governance around them. If that pattern spreads, the shift could reshape both software work and the balance of power in European tech.
Explain this