There comes a point in any technological cycle when the media shifts from ‘this is amazing’ to ‘this is overhyped’ to ‘ah, this is what it really is’. With generative AI, we’ve been in the second phase for a while, but I think we’re very close to the third.
AIs are maturing. And maturation means they are no longer evaluated solely by benchmarks, but by how they fit into real processes, in real organisations, with real people behind them.
What is somewhat overrated
The narrative of ‘the world’s smartest AI’ changes every four or six weeks and offers little to those who have to make real decisions.
ChatGPT isn’t overrated as a tool, but it is as a ‘universal winner’. It’s the most versatile, but not the best in every category. And sometimes people use it for tasks where another tool would clearly be more suitable, simply because it’s the one they know.
Copilot isn’t overrated either if you understand that its value lies in integration. It is overrated when someone expects it to be the most advanced AI on the market. That’s not what it’s about.
What is clearly underrated
NotebookLM, for most professionals working with documentation, is far from being adopted to its full potential. People who try it rarely go back to working without it.
Qwen, in technical environments, is receiving far more specialist attention than it does in public discourse. Its capabilities in coding and mathematical reasoning are impressive.
Mistral, in the European context of technological sovereignty, is hugely underestimated outside specialist circles. And with the progressive implementation of the EU AI Act, that advantage is set to become increasingly significant.
Europe: the most overlooked variable in discussions about AI
The EU AI Act, in force since August 2024, is forcing companies worldwide to rethink their AI architectures. It classifies systems by risk level: from prohibited uses, such as mass surveillance or subliminal manipulation, to high-risk applications in areas such as healthcare, employment or education.
But beyond regulation, there is a fundamental question that more and more organisations are asking: where is my data, who processes it and under which laws? This is not merely a technical issue. It is a matter of trust.
Anthropic began offering EU data residency for Claude via its API in August 2025. Mistral offers it natively. Google and Microsoft have European cloud options. But the regulatory framework will continue to tighten, and organisations that have not considered this factor will face unpleasant surprises.
Where is this heading in two years?
If you were to force me to make a prediction on what the landscape will look like in 2028:
ChatGPT will remain a major centre of gravity due to its versatility and ecosystem, but it will lose market share in the professional segment to Claude, which already leads in that area.
Gemini could grow significantly if Google turns its integration advantage into a more visible experience for the non-technical user. The infrastructure is there. The UX is missing.
The open-source bloc — Llama, Mistral, Qwen — isn’t going to disappear but will become more relevant. No major economy will want to depend entirely on a handful of closed models.
And the biggest change of all will not be a new model: it will be the normalisation of agents. When automating a complex process ceases to be an IT project and becomes something anyone can do from their usual app, the way we work will change permanently.
