Common uses are clustering a universe into groups that behave alike, so that a portfolio does not accidentally hold twelve versions of the same bet, and principal component analysis to strip out the market factor before looking for residual opportunities.
Regime detection with clustering or hidden Markov models produces tidy labels, but they are only known with a lag and they change when you refit. Treat them as descriptive context feeding a regime-filter, not as a signal.
The absence of labels also removes the usual accuracy check, which makes it very easy to convince yourself a clustering is meaningful. Validate by whether the structure is stable out of sample and whether it improves a downstream decision, not by whether the picture looks good.
Related: supervised-learning, market-regime, correlation, feature-engineering