Rounding is a choice
The rounding rule is the one part of a schedule nobody documents, which is exactly why it is worth looking at.
Present a schedule in thousands and you will meet the small problem within about a minute. Every line rounds cleanly. The total does not. The sum of the rounded lines is out by one or two against the rounded total, and the page now contains a visible contradiction that a reader will notice before they notice anything you were trying to say.
There is no clever fix. Somebody picks a line and makes it absorb the difference.
I have done this at a practice, at a software company, and at an accounting department, and I do not think I was ever asked which line I picked. It is treated as formatting. It is the last five minutes before the file goes out, and it is invisible.
Where the difference lands
The plug goes into the line nobody asks about.
Not out of dishonesty. Out of an entirely reasonable instinct, which is that you do not want a rounding artefact showing up in the figure the reader is going to interrogate. So it goes into other operating costs, or into the line that has been flat for three years, or into the one that is large enough that a thousand is invisible against it.
Every one of those is a judgment about which numbers are being watched. Which means that a schedule’s rounding decisions are a small, accurate, unintentional map of what the preparer believes nobody is checking.
The same thing happens one level up, on figures that are estimates rather than totals. An accrual is a range with a number written on it. When you round it, you round toward the end you would rather defend in three weeks. Not by much. Reliably in one direction, though, and across a year that direction is not random, because the thing generating it is a stable preference rather than a coin.
What I got wrong about it
I policed this for a while as an accuracy problem, and on accuracy grounds I was plainly wrong. The amounts are immaterial by construction. That is the point of rounding. Nobody’s decision changed because a thousand sat in one line instead of another, and I was making a fuss about a difference that could not move an outcome.
What I missed is that it was never about the amount. It was about provenance.
A rounded figure asserts something quietly: that this number is known to about this level and no further. That assertion is often correct and sometimes wildly generous, and either way it is made by a person, in the last five minutes, with no record. The rounding convention was the only element of every schedule I ever built that lived nowhere in writing.
The undocumented part of a process is where the preferences go. Not the crimes, usually. Just the preferences.
The other side of it
I want to be careful, because there is an obvious overcorrection and I have seen people make it. You can respond to all this by refusing to round, publishing to the cent, and calling it transparency. That is worse. A figure carried to the cent claims a precision that the underlying estimate does not have, and now you have replaced a small quiet choice with a large loud false one.
Rounding is not the problem. Rounding without anybody having chosen is the problem, and the two are difficult to tell apart from outside because they produce the same page.
What I do now is small and slightly unfair. When a schedule arrives, I ask who chose the unit. Thousands, millions, whole dollars. It is a harmless question and it takes four seconds to answer, and it is the fastest way I know to work out who actually built the thing and who signed it.