Why does the group position stop agreeing once you look per entity?
Because a total hides disagreement and a breakdown does not. Group cash is a sum over external balances, and a sum is forgiving: two entities can hold opposite views of the same internal movement and the group figure still comes out right.
The moment somebody asks for the position by entity, by country, or for whichever scope a lender looks at, those opposite views have to be reconciled to each other. That is a different and harder question than the one the group total answered, and many systems only ever answered the easy one.
It is also the question that arrives at the worst time. Nobody asks for the entity breakdown on a quiet Tuesday. They ask for it when a bank, a lender or an auditor has asked them first.
What breaks first between two entities?
The pair of entries. One movement inside a group is two bookings, in two ledgers, made by two people who never spoke, and there are four ways they diverge before anybody has made a mistake.
They diverge on date, because the sending entity books when it instructs and the receiving one books when it sees the money. They diverge on amount, because a cross-border transfer arrives net of a charge only one side saw. They diverge on currency, because each entity books in the currency it actually books in and the pair only agrees at a rate somebody has to name. They diverge on identity, because the counterparty is a code on one side and a typed name on the other.
Each of those is small. Together they mean the two legs of a movement will not find each other by amount and date, which is exactly how a system that matches on amount and date reports a clean intercompany position that is not one.
Why is netting the hard part rather than the easy part?
Because you cannot net balances you have not agreed. Netting is arithmetic performed on a set of positions both sides accept, and getting to that acceptance is the whole job.
In practice it means every pair of legs matched to each other, every currency stated with the rate and the moment it was taken, and every disagreement named rather than averaged. A position with one leg missing is not a smaller position, it is a different one, and settling on it moves real money on the strength of a figure that was never true.
It also has to survive being redone. If the same period nets to a different number next week because somebody re-ran it, nobody will trust the first number again, which is why the matching decisions have to be kept rather than recomputed from whatever the ledgers look like today.
What has to be true before a group figure is worth quoting?
Three things, and they are checkable in an afternoon. Both legs of every internal movement have been matched to each other, and the ones that have not are visible with their count rather than folded into the total. Every currency in the figure carries the rate it was converted at and the moment that rate was taken. And every name appearing twice has been resolved to one counterparty, with the resolution recorded where a reviewer can disagree with it.
If any of the three is missing, the figure is still useful internally and it is not a figure to hand to a lender. That distinction, between an operating number and a number somebody else will rely on, is worth making out loud, because the same total is usually used for both.