Most real AI use augments, it does not replace
What people actually do with AI, measured, not predicted.
Predictions are cheap. Usage data is rarer and more honest. When you look at what people actually do with capable models, the dominant mode is working alongside them, not handing the task over. That is the augmentation frontier: the large middle of work where AI makes a person faster and better rather than absent.
What the evidence shows
Across millions of real Claude conversations, roughly 57% of use was augmentative (collaborating with the user) and about 43% automative (completing a task), concentrated in software and writing.
A field study of customer-support agents found a generative assistant raised productivity about 15% on average, and around 30% for the least experienced, by spreading the know-how of the best performers.
The ILO reached the same conclusion at a global scale: augmentation, not automation, is the more likely outcome for most exposed work.
Augmentation is not automatically good for you. If a tool lets three people do the work of five, two roles still go, even though everyone left is 'augmented'. And augmentation that flattens the gap between novice and expert can erode the premium you spent years earning.
The safest position is to be the person who wields the tool well, not the person whose entire output the tool can mimic.