Not just old and out of date. Plain wrong, like referring to internal systems, processes or intranet URL's that are (often poorly done) duplications of enterprise standard services, processes or pages, and should be deprecated instead of incorporated into training data.
Or the unstructured data sits within structured data. Some development teams stick giant strings with their own internal, proprietary formatting into a database blob column because...I dunno, they thought it was expedient or whatever for Very Good [Undocumented] Reasons. Invariably, these fields hold all sorts of nastiness you either want to tease out, or do not want to exist at all because they expose what you wish they wouldn't in cleartext. And even though your company pays the vendors of these monstrosities (oh yes, my PTSD suppressed memories of multiple vendors committing these same terrible, no good, very bad software sins) three commas a year of "support", they adamantly refuse to share their precious proprietary format with you, much less give you the courtesy to update you when they change the byzantine format that gives the three body problem a run for the money for a deterministic solution.
If you make a Generative AI sort sense out of that, then you can practically segment the market by "how much before the customer starts to put away the blank check" and you'll make bank. Everyone I know of in this Data Loss Prevention (DLP) or adjacent spaces are deer in the headlights when the conversation turns towards these operational realities and what we might do with their solutions to put a dent in these multiverses of pain. No one has roadmaps talking of iterative pipelines, search outcome behavior analysis and automated training thereof, or really much beyond various degrees of pattern matching, whether that be regex's or LLM's.