Does AI accelerate development?
AI tools are considered productivity boosters – but do they deliver on their promise? And what does all this have to do with gambling in casinos?
We asked our colleagues how they use AI in their everyday work as developers and what advice they would give to juniors on working with AI.
Subjectively faster – objectively slower?
Expectations for AI in software development are high: automated code completion, fast bug fixes, and more efficient development teams are among the common promises. However, a new study by the research organization METR shows that for experienced developers, the use of AI tools can actually lead to productivity losses.
In an experiment with 16 experienced open-source developers, one group worked with AI support and the other without. The result: the AI group took 19% longer on average – even though they subjectively felt they were working faster. Where does this feeling come from? We asked our colleagues about this contradiction between subjective speed and actual progress.
AI alternately gives you "wow" moments and total failures. I think developers often use AI like a slot machine, where you pull the lever until the right code comes out. It's easy to forget how long you've been sitting in the casino.
Many developers tend to overestimate the productivity gains of AI – an effect that is exacerbated by the current AI hype. While AI tools can generate code at impressive speeds, the result is not always optimal. "Wow" moments and total failures often alternate, and it is easy to get lost in constant reworking.
The researchers in the METR study also see this as the cause of the discrepancy between expectation and reality: interaction with AI, especially dialogue-based communication, is often time-consuming and cognitively demanding. Added to this are necessary checks and adjustments to the generated code, as well as sometimes suboptimal solutions from the models.
Another factor is that many experienced developers already work extremely efficiently, so additional speed through AI is not guaranteed in such cases.
Where AI helps – and where it doesn't
Not every task is equally suited to AI support. AI can bring real added value to some activities, but in other areas it reaches its limits. We investigated in which situations it is worth using—and when it is better to do without it.
If you are not very familiar with something, AI can be a great help. Because topics change in everyday development work and you cannot be an specialist in everything, no one is always an expert or always a beginner.
Added value ✅
- Writing, proofreading, reading and reproducing documentation
- Anything that can be found by googling
- Repetitive tasks, refactoring, writing test mocks
- "Simple programming tasks" that are easy to check
Limits 😌
- Complex technologies that are highly application-specific or not widely used
- Complex existing code bases
- Complex debugging across modules and libraries
- Major changes (involving multiple steps across many files)
Recommendations for less experienced developers
AI can be a valuable tool, especially for beginners—but it can also become a trap if suggestions are accepted uncritically. Our experienced colleagues provide recommendations on what to look out for when getting started with AI-supported development, which skills remain important, and why reflection is more important than automation.
It is important to recognize when your mind has actually switched off—when you are just hoping for an answer in AI mode that somehow works, but that you no longer understand yourself. That is precisely when it is worth taking a step back, being honest with yourself, and identifying the unclear points—and really explaining them to yourself (with the help of AI, if necessary).
Always have AI explain generated code to you until you understand it. It's best to write this down in your global system prompt right away. If you adopt code without understanding it, you'll forever remain a junior who can only deliver until the LLM inevitably screws up. We'll need fewer and fewer developers like this in the future.
Conclusion – Tools instead of miracle cures
The voices from the team show that AI can support developers—but it cannot replace them. Productivity does not come from tools alone, but from experience, contextual understanding, and critical thinking. Those who use AI in a targeted manner save time. Those who rely too heavily on it risk the opposite.