From Statistics to Sentences: Writing Executive Summaries People Actually Read

Nobody reads the table. Executives, clients, and stakeholders read the three sentences above the table — and if those sentences are missing, they skim the charts and invent their own narrative, which is worse. Turning descriptive statistics into a decision-ready summary is a distinct skill, and it's one we've had to encode explicitly for the automated executive summary in KPI Master. Here's what a good summary says, what it must never say, and how to mechanize the difference.

The Transformation Problem

Raw statistics answer statistical questions: mean $1,850, median $1,620, std $940, n = 440, trend +8% period-over-period. A summary answers the reader's question: what is happening, and does it need my attention? The transformation has rules:

1. Lead with direction and magnitude, not the metric name. "Revenue grew 8% to $142K" beats "The revenue metric shows a positive period-over-period delta." The number is the sentence's payload; burying it in process language is the most common summary failure.
2. Give magnitude a comparison. "+8%" is noise or signal depending on baseline volatility. If the metric normally swings ±10% monthly, 8% is a shrug. If it normally moves ±2%, it's the story. A summary without a volatility baseline is numerology.
3. One insight per sentence. Compound sentences ("revenue grew while churn fell and CAC rose") force the reader to hold three threads. Three sentences, three facts, in descending order of importance.

What an Automated Summary Should Say

The honest automated summary is a ranked list of verifiable observations, each traceable to a statistic the user can check:

• The headline movement: the primary metric, its current value, its change, and its trend direction — the same Holt-forecast-informed "where is this heading" signal, stated plainly with its confidence interval when one exists.
• The distribution shape: if the median and mean diverge badly (say, mean $1,850 vs. median $980), the data is skewed by large values, and that fact changes every conclusion. Say so.
• Segments, when they exist: "values split into three clusters centered near $13K, $45K, and $90K" is an observation a clustering pass can verify — and far more useful than any average.
• The strongest relationship: the top correlation, labeled correctly as association — "moves with," never "drives."
• The data quality caveat: if 12% of a column is missing or the dataset grades a C, that caveat belongs in the summary, not in a footnote. It's the fact that most changes how much trust the rest deserves.

What It Must Never Say

The constraints matter more than the features, because language implies claims statistics can't support:

• No causal verbs. "Marketing spend boosted revenue" is a claim about causation from an r value. The correct phrasing is "moved together with." Automated summaries that use causal language manufacture false confidence at scale.
• No precision theater. "Revenue increased 8.347%" signals false exactness. "About 8%" is both honest and more readable. Round to the precision the data quality actually supports.
• No recommendations. "Consider increasing ad spend" requires context the data doesn't contain — margins, capacity, strategy. The moment a summary prescribes, it oversteps what statistics can know. Observe, don't advise.
• No hiding the caveats. A summary that reports the trend but omits that 15% of rows are duplicates isn't a summary; it's a liability.

A Before-and-After

Bad: "The analysis was performed on the dataset. Revenue metrics were computed. The mean value is 1850. Variance was detected. See charts below." This is a process log wearing a summary's clothes.

Good: "Revenue grew about 8% to $142K, continuing an upward trend — the next period is forecast near $152K (±$6K). The distribution is right-skewed: a median order of $980 versus a $1,850 mean means a small number of large orders drive the total. Note: 12% of revenue rows are missing values, so totals are understated." Three sentences, three checkable claims, one honest caveat. That's the standard — for humans writing summaries and for the software that drafts them.