More capable and more varied models
After the 2023 explosion, models gained context length, multimodality (text, image, audio, video) and reasoning ability: by late 2024, models designed to “think” before answering appeared. Open models progressed quickly and offered an alternative to proprietary offers, with European players such as Mistral.
From assistants to agents
Assistants no longer just answer: they chain actions (search, read, write, call a tool). Integration standards are emerging to connect models to company data and applications.
A regulatory framework taking shape
The EU AI Act came into force in August 2024 and applies in stages: prohibition of certain practices, obligations for general-purpose models, then requirements for high-risk systems. Organisations must map their uses and define roles.
From experimentation to industrialisation
The debate is moving from “what can AI do?” to “what value, at what cost, with what risks?”: energy sobriety, sovereignty, quality evaluation, team training.
Key takeaways
- Choose measurable use cases rather than multiplying experiments.
- Set a framework: usage policy, data, compliance.
- Train: adoption depends on skills as much as on tools.