Design correction dialogs so they follow the conversational convention of allowing one-step correction
For example, avoid:
Instead, use grammars that allow callers to say “No” and reasonable variants of “No” before the corrected amount, as well as allowing them to just say the corrected amount. The improved dialog would be something like:
Look at real data to see where these types of corrections occur before building them all over an application
Note that these are more complicated grammars to write, and can get wildly complicated depending on how many different ways the caller could correct. There are therefore tradeoffs to be made. Sometimes it's best to wait and see exactly where people are actually behaving like this on a regular basis. It will be much easier to sell a team on the effort to put it in a couple of key places where the data shows it's needed than to do it in many places before pilot.