Most machine-learning teaching starts with supervised learning: inputs, labels, a model that maps one to the other. Most real data has no labels. The interesting question is what remains possible when they are absent.
What clustering is for
Clustering groups observations by similarity without being told what the groups mean. The output is not an answer — it is a hypothesis about structure. A clustering of customer records, sensor readings, or plant-waste samples says "these look alike" and leaves the meaning to whoever knows the domain. Treating the groups as facts rather than hypotheses is the first way unsupervised work goes wrong.
The second way is expecting one true grouping. k-means with a different seed, a different k, or a different distance measure returns a different partition, and all of them can be locally sensible. Stability across reruns is evidence; a single run is not.
Evaluation without ground truth
With labels, accuracy is a number. Without them, evaluation is indirect: within-cluster cohesion against between-cluster separation (silhouette scores), stability under resampling, and — the one that matters most — whether the groups survive contact with someone who knows the domain. A cluster that a domain expert recognizes as meaningful is worth more than any internal metric.
When not to cluster
If a decision needs to be defensible — pricing, eligibility, triage — an unverifiable grouping is a weak foundation. Unsupervised structure is best used for exploration, segmentation for human review, and features for a later supervised step. The technique earns trust by producing leads for verification, not by replacing it.