Have you ever had an LLM silently pass its own bad assumptions down your pipeline?

I ran into a messy pipeline issue recently where an LLM was supposed to clean up incomplete user inputs before a database lookup.

The problem was that downstream functions started treating the model's cleaned-up version as the actual ground truth instead of just a search hint. It ended up matching fake data with real database entries because it was basically judging its own guesses instead of what the user originally gave us.

Now I keep raw user data completely locked away from any AI modifications during the final check.

How do you keep your AI outputs isolated so they don't corrupt the actual logic?