Simple tables
Single source, mostly normalized, low temporal pressure.
Rich mixed records
Text, numeric, categorical, joins, and normalization choices interact.
Temporal / graph / high-dimensional
Embeddings, linked entities, sequences, drift, or local neighborhoods matter.
Legal / financial / safety cost
A wrong group creates loss, audit exposure, or safety risk.
Audit before use
High-consequence clustering needs a defensible failure policy even when the table looks simple.
Defend every assumption
Similarity design, sampling, and review evidence become part of the deliverable.
Defect finding
Specialist territory
Mistakes are expensive, non-obvious, and usually worth an independent review before production.
Fraudulent transactionsLegal clausesHistorical legal driftRobotic IoT eventsPsychographic data
Operational cost
A wrong group sends money, work, or staff attention the wrong way.
Measure before routing
Even simple groupings need error checks when they drive queues, spend, or escalation.
User complaints
Second opinion pays
Mixed features, entity joins, and text normalization can flip assignments silently.
InvoicesProduct parts
Approval problem
Time, graphs, embeddings, and drift make errors harder to see and harder to unwind.
Outlier behaviorsUser behavior
Low cost
A wrong group mostly slows exploration.
Baseline can be enough
Use defaults as a quick read, then check stability before anyone acts.
Prototype carefully
Normalization and joins can manufacture groups. Validate before interpreting them.
Do not trust the projection
A neat two-dimensional view is not evidence that neighbors are meaningful.
Cost of the wrong cluster
Data shape, richness, joins, graphs, normalization, and time