What does an honest, buffer-less estimate mean? Is it the number of person-hours to make you 50 % sure the task is completed? Or 90 %? Or something else?
For planning on film sets I would use the "this is how long it would normally take to do X if you do it at a normal tempo and no unforseen blockers appear".
For these you add the buffer. Sets that I scheduled were most of the time slightly faster than the schedule, until a blocker appeared which ate the buffer, but was usually resolved without eating it away completely. So at the end of a day we were always either on time or a bit ahead.
In film this is the ideal case, if tou depend on the weather or on certain locations being open, missing your slot can mean that you have to try for a whole week to get it again (or have everybody work on something else and rush them over once it works)
PM guru Eliyahu Goldratt advocated for a 50% likelihood of completion with no buffers for individual tasks, but one shared global project buffer.
I don’t think there’s a single right answer for that percentage, but the key is that it should be determined by the person or group defining the global buffer and should be done with consideration for the distribution of possible durations the project tasks might take.
The expectation when removing per task buffers is that some estimation errors for individual tasks will be high and others low so ultimately to an extent they'll cancel each other out.
In Goldratt's system, the risk comes from the longest (in terms of duration) chain of dependent tasks. Delays early in the chain can only be canceled out by speed ups in later dependent tasks, that's the primary motivation for that global buffer.
The technique was adapted from factory production line optimization. The case can be made that it can work well for projects doing repeatable and thoroughly understood things like constructing buildings. There are reasons it's not a slam dunk for software, but IMO there's still a lot to learn from his work.
