It's that (already) old joke: we give the LLM 5 bullet points to write a memo and the recipient uses an LLM to turn it back to 5 bullet points.
Some plausible (to me) possibilities:
1. Bifurcation: Maybe a subset of knowledge workers continue to write and read and therefore drive the decisions of the business. The remainder just do what the LLM says and eventually get automated away.
2. Augmentation: Thinking is primarily done by humans, but augmented by AI. E.g., I write my thoughts down (maybe in 5 bullet points or maybe in paragraphs) and I give it to the LLM to critique. The LLM helps by poking holes and providing better arguments. The result can be distributed to everyone else by LLMs in customized form (some people get bullet points, some get slide decks, some get the full document).
3. Transformation: Maybe the AI does the thinking. Would that be so bad? The board of directors sets goals and approves the basic strategy. The executive team is far smaller and just oversees the AI. The AI decides how to allocate resources, align incentives, and communicate plans. Just as programmers let the compiler write the machine code, why bother with the minutiae of resource allocation? That sounds like something an algorithm could do. And since nobody reads anyway, the AI can direct people individually, but in a coordinated fashion. Indeed, the AI can be far more coordinated than an executive team.