The problem is that Devs making estimates tends not to be accurate IME. Some stuff is easy to estimate, but adding in new & significant functionality is notoriously usually hideously mis-estimated (either high or low). So much so that the numbers basically b come meaningless and can't be relied upon because every major new thing was a lot easier/a lot harder than expected.
The reaction to this is to split the task up, but this often requires that you do a lot of pre-work-work to architect and design the work so you know how it can be broken up, and that design work needs to be estimated so you are back at square one with wildly inaccurate estimates.