A high model score is not a useful customer segment.

The result that deserved a second look

An initial K-means analysis grouped 99.3% of the examined accounts into one broad segment. A high silhouette score did not make that segment useful for understanding customer jobs or prioritising a product change.

This was a diagnostic warning in a specific historical analysis, not a universal rule about clustering scores. A score needs the data distribution, features, population and product question around it.

Check the inputs before choosing the algorithm

The review inspected lifetime activity counts, highly skewed usage, account tenure, extreme users and test/demo records. Lifetime totals can blur the distinction between long tenure and high current engagement.

Alternative features included monthly rates, recency, frequency and feature breadth. The documented work considered K-means, RFM-style scoring and density-based clustering with HDBSCAN.

Evaluate the model and the decision

Clustering quality, stability and commercial usefulness are related but different questions. No single score proves that a segment corresponds to a real buyer job or should receive a different onboarding experience.

Compare like-for-like configurations where possible, inspect sensitivity to preprocessing and account exclusions, and validate the resulting interpretation against customer behavior and qualitative evidence. A lower silhouette score is not automatically better.

What this does not establish

The account describes analytical and methodological work. It is not a measured retention improvement, causal activation finding or independently audited revenue result.

It demonstrates the kind of judgement brought to offer research: check assumptions, keep the analysis reproducible and challenge an impressive-looking output before using it to make a product decision.

Why it matters for agent choice

Averaging different jobs, workloads or agent contexts can conceal the decision that matters. Research needs the right unit of analysis and meaningful segments before a single aggregate preference rate can support a recommendation.