It's typically automatically exported/imported. People aren't manually editing cells with millions of rows.
The wrong entry is only wrong if it came out of the source data, and generally it isn't a case of "ruining the whole thing" any more than it is already incorrect inside the line of business app.
extracted from a database and then formulas/pivots are done on tables (though people often don't use the actual table feature so in practice a "table" is just a hardcode rectangular range of cells). However, it is still a very error-prone design.
So the ERP system can't do formulas and pivot tables like excel? I never used one.
Assuming someone imports the data correctly, it is still very easy to corrupt. Many years ago I was given a junior analyst to "fix" and one of their assignments was to analyze some data (de facto we used Excel). I couldn't figure out what they did to it. Best guess is they tried to sort the table but only included some of the columns in the sort so at that point the data looked like it went through a blender. Plus somehow the column totals were still off from the original in ways I couldn't figure out either...
Putting that aside, formulas typically have to be repeated for every row using a filldown logic. So instead of something like NewVar := OldVar1 + OldVar2 in real data analysis systems, one writes "=A1 + B1" in cell C1 (row 1, column 3) and then "fills down" (the 3rd column) letting Excel update the formulas as it goes along to "=A2 + B2", "=A3 + B3",.... Don't forget to get your cell anchor references right in more complex situations. And yes your formula for C25313 could accidently get corrupted to "=A25131+B25313" and Excel will happily chug away (it will flag such cell with a little mark though if you happen to notice while scrolling through it).
There are best practices to avoid the most common mistakes like the above and various new features designed to make it work better with "big data" but I go back to my thesis: the spreadsheet model is very error-prone design.
is there a way to make the source data/cells "immutable" so you can only analyse them but not alter them by mistake?