Do you store also predicted intent with the conversation texts? if yes, you can do two things:
Use predicted intents in the evaluation
Use unrecognized texts for generating new training data, but remember, that directly putting everything in the training data is a bad idea, it’s better to use this data to generate little bit different examples so that model correctly recognizes similar examples too
Unfortunately I don’t know how you can automatically evaluate this saved data, so at some point you still have manual work and overall I believe to improve bot, manual work is important, at least in the beginning.