Check out this paper from Salesforce AI Research showing that even leading LLM agents achieve only 58% success rate.
And it is not just in CRM agent tasks that LLMs fail often.
Even coding tasks will give you results that look great but are terribly misleading.
EXAMPLE: I was coding a LWC and just for testing gave that same task to the LLM.
It coded a call to a function like this
closeTab( { tabId: aTab.id } );
THE PROBLEM: the actual function parameter is not "tabId" as a property enclosed in an object, but the simple value of the tabId itself. And that function has to be called with async/await.
Running that LLM-generated code would not work and, on top of that, would fail in a way that does not obviously indicate what the issue is.