What is generative AI? Definition and business implications

Traditional AI was limited to classifying, predicting, or detecting (fraud detection, credit scoring, image classification). Generative AI instead produces new content (text, image, sound, video, code) from a natural-language instruction. It does not replace predictive AI; it opens a parallel field of uses, wherever the need is to produce a deliverable rather than make a decision.

It relies on conditional prediction: given your instruction, it generates the most probable continuation element by element, based on what it has learned. These models are trained on massive corpora, hundreds of billions of words or images, whose statistical regularities they absorb. The same mechanism covers LLMs for text, diffusion models for image and video, and audio models for speech and music.

According to McKinsey's State of AI 2025 report, 79% of organisations regularly use generative AI in at least one business function, up from 33% in 2023. Yet the impact is uneven: only 39% of respondents report any effect on EBIT, and most put it below 5%. Adoption is slower in Europe, at 20% of enterprises with 10 or more employees in 2025 according to Eurostat.

Because diffuse usage is not enough. The small group of “AI high performers” (about 6% of McKinsey respondents) attribute more than 5% of their EBIT to AI, and stand out by redesigning their workflows rather than layering AI on top of existing tools. Differentiation no longer plays out on the technology, but on usage discipline, case framing, and human supervision.

No. API-available models are within 5% of each other on common use cases, so the model is no longer the barrier. The real exposure is now comparable to regulatory risk: hallucination, data leakage, and supplier dependence. No serious deployment can do without an explicit guardrails policy.