Nobody reads dashboards
Most dashboards are opened weekly, glanced at and closed. The interpretation people want takes time nobody has, so the data sits unused.
Narrative summaries generated from the same data are read, because they say what changed rather than showing that something did.
What a good generated summary contains
- The two or three things that changed materially, with numbers
- Anything outside its normal range, flagged as such
- Where a change is explained by a known factor, saying so
- What has not changed but was expected to
- No padding — a short summary is a used summary
The fourth is the one humans miss most. A metric that should have moved and did not is frequently more informative than one that did.
Anomaly detection needs the boring statistics
Use proper statistical methods to find what is unusual, then use a model to explain it in plain language. Asking a language model to spot anomalies in numbers is the wrong tool.
That division — maths for detection, language for explanation — is the pattern that works.
Answering questions over reports
- Grounded in the actual data, not in prose about it
- Showing the query or the figures used
- Refusing when the data cannot answer it
- Respecting who is allowed to see what
Where to be careful
Never let a generated narrative assert causation from correlation. “Sales rose after the campaign” is a fact; “the campaign drove sales” is a claim requiring evidence.
That distinction matters most when the summary goes to a board, which is exactly where it is most tempting to blur.