it is never useful because it produces biased predictions. For much better framework. Unfortunately your diagram is not correct - such diagram would be correct is one used classical frequentist statistics. But with modern frequentist methods such as conformal prediction neither large data set is required and yes it produced uncertainty quantification regardless of the underlying model, sample size and is distribution free. Not only that unlike Bayesian framework it has inbuilt mathematical guarantees of validity ( lack of bias) and unlike Bayesian framework that produces posteriors in the wrong place and overconfident prediction intervals, conformal prediction does not suffer of any of these drawbacks - if the problem is hard it will alert the user by producing flat predictive distribution rather than peaked one in the wrong place like Bayesian will do.
