With a field moving as quickly as AI right now, you should probably take any predictions over what it will be able to do and not do with a big grain of salt, especially if they are made for any timespan greater than the very near term future (months? weeks? That term will itself get shorter and shorter as time goes on).
Very few people, including the author of this study most likely, predicted a few years ago where AI would be today. That should tell you something.
The model of how AI would influence work, including splitting work into tasks, classifying tasks by what type of AI could complete it, the complementarity of prediction with other aspects in an organization, etc., all have been studied for decades. The people working on this, including the author of this study, have not changed their conclusion dramatically recently.
What has changed is the apparent capacity of LLMs: they got convincing earlier than other specialists thought possible. However, that timing isn’t the most relevant aspect of industrial economics and transaction cost theory. Many people make that mistake, but Oliver Williamson himself has corrected that misunderstanding in front of me several times.
What matters is that, when the technology is available, people will reorganize functions.
Very few people, including the author of this study most likely, predicted a few years ago where AI would be today. That should tell you something.