Generative AI is the family of models that produce content: text, images, code, audio — new artifacts shaped by patterns learned from training data, steered by a prompt. The term covers the large language models behind every chat assistant as well as image and audio generators, and for most organisations it arrived as a writing and coding aid: a person prompts, the model drafts, the person decides what to do with the draft. That last clause — *the person decides* — turns out to be the load-bearing part of the whole arrangement.
Three properties define the technology operationally. It is probabilistic: the same prompt can yield different outputs, and quality varies in ways no test fully pins down. It is fluent regardless of correctness: wrong answers arrive in the same confident prose as right ones, which is why hallucination is a workflow problem, not just a model flaw. And it is steerable by whatever it reads: instructions and content blend in the input, a property that becomes a security issue the moment outputs stop being reviewed. Techniques like retrieval-augmented generation ground outputs in real documents, and prompt engineering shapes behaviour — both manage these properties; neither abolishes them.
The line that matters: content versus action
The consequential distinction in 2026 is not between generative models — it is between systems where a human stands downstream of the output and systems where software does. Generative AI used as a content tool is governed like one: review the output, set usage policy, train the users. But the same models now sit inside agentic systems that execute their output as actions — and every property above changes meaning. Probabilistic becomes *unpredictable behaviour on real systems*; fluent-but-wrong becomes *confidently wrong actions*; steerable-by-input becomes *prompt injection with consequences*. The full comparison draws this line carefully, because it is the line your governance has to follow: same model, different blast radius, different controls.
What readiness looks like at the generative layer
Even before agents, generative deployments carry real obligations: knowing where model output enters decisions and records, keeping sensitive data out of prompts and traces (or accounting for where it lands — see AI security risks), and writing usage policy that names what may and may not be produced and shipped. The governance practices apply here in their content-era form. The reason to do them well now is partly their own value and partly preparation: every habit — inventory, ownership, evidence — transfers directly to the agent era, where it stops being good practice and becomes survival. The organisations that governed their generative tools are the ones whose agent transition goes quietly.