The ability to selectively recover from bad data input should really be part of the initial database design. It's not enough to have 'point in time" because substantial transactions may have occurred in the interim before discovery. You really can't afford to manually review multiple thousands of entries. Transactions need to be fully indentified and preserved (in an append only table, for example). This involves careful planning with the business rules.

It's expensive in development time, but if there is a potential need it is worth it.