How does clustering work?
Every method needs two decisions: how to measure similarity, and how to form groups from it. Similarity is usually a distance, Euclidean for numeric features, cosine for text represented as embeddings. Grouping strategies differ: k-means partitions around k centers, hierarchical clustering builds a tree of merges, and density-based methods like DBSCAN grow clusters wherever points are packed, leaving sparse points unassigned. The algorithm choice matters less than the representation: cluster bad features and you get confident nonsense.
What is clustering used for in business?
Customer segmentation is the classic: grouping buyers by behavior rather than by the categories someone guessed. Document and ticket grouping surfaces what a support queue is actually about. Image segmentation groups pixels for inspection systems. And the mirror use: a point far from every cluster is an outlier, which is the geometric heart of anomaly detection and fraud screening. Clustering answers the question you could not have asked: what groups exist here at all?
How does clustering fit modern AI systems?
Embeddings made everything clusterable. Any text, image or product can be mapped to a vector, and vectors cluster, which is why topic discovery over documents now takes an afternoon: embed, cluster, label each cluster with a model. The same geometry powers semantic search, and a vector database is, in effect, infrastructure for asking which stored items sit nearest a query. Clustering is the exploratory half of the same geometry retrieval exploits.
What goes wrong with clustering?
Three classics. Choosing k by wishful thinking: validate with silhouette scores or business sense, not round numbers. Scale effects: a feature measured in thousands dominates one measured in decimals unless you normalize. And over-interpretation: clusters are statistical artifacts until a human names what they mean and checks they are stable on fresh data. A cluster is a hypothesis, not a fact, and the useful ones survive next quarter’s data.