For a Generative AI company, an example of performance metric may be factual correctness. The threshold maybe 100% and in this case if the Gen AI produces any factually incorrect text, the loss will be covered. A relevant metric for speech transcription can be Words Error Rate (WER). For a predictive analytics AI, the performance can be accuracy or recall of the predictions, whichever is most crucial in the context the AI is being used.
Generally, an AI company chooses a performance metric that is most relevant to their users/customers.