Given worker access to generative LLMs, plus training and motivation to use them, LLMs are effective for certain workflows. Those workflows tend to be personal, one-offs, or summarization in nature: write a bash script for this headache I have every day; tell me what colleague X is trying to say in his 1200-word email, since his writing is garbage and he can't get to the point; "what's the Excel formula syntax for this other thing that I keep forgetting?"; etc.
So the time and mental-energy savings inures to the workers, mostly from coordination tasks that don't directly create core value. And then those savings aren't "reinvested" into value-producing activities whose benefits would inure to the firm because the workers have no incentive to do so; don't know how to create core value; don't have the skills to create core value; or aren't permitted to do those activities by higher-ups.
Bottom line: LLMs are eating busywork coordination activities — hence no impact on most firms' bottom lines.