On the other hand in context of CRMs, helpdesks and other things that have to process received emails and then show them in web interface this gets hairy because escaping on output while correctly handling variously weird or outright broken input is non-trivial. For example every helpdesk solution I've seen gets confused by some combination of nested MIME (ie. email with attachments forwarded with additional attachments) and notionally non-text email bodies (ie. bounces from Postfix).
[Edit: and in this case storing whatever was on input without any transformations and sanitisations enables you to recover the content WHEN (not if) the breakage occurs, by either manually examining the input or, in the email case, importing it into desktop MUA. I even had once to do this with message that was unreadable in otherwise surprisingly well-behaving gmail's web interface]
However, the 'defender' in me can think of a situation where this would be good:
Input filtering mechanisms have improved to deal with new attacks/problems, but old data is not 'new input', so you'd apply the updated algorithms to existing data.
I'd definitely save input that breaks the rules when received, too - but maybe in a separate table.
If I have a comment forum with input filtering (rather than output escapong) I can't restrict it to just letters numbers and punctuation or the user's can't discuss HTML tags!
It was probably less than an hours work to add that button vs. hundreds of hours to track down all the unescaped outputs.
What happens if we view the admin panel and trigger one of these malicious scripts, because I haven't clicked this button recently enough? Oh, maybe someone should click it very frequently!
Or... Hmm, maybe rather than storing data that may be mass deleted by a script, with no review, after possibly compromising our security, we should just reject it in the first place and log a security incident for that user.