Narrative insights

Narrative insights turn a table or chart into a few sentences a busy reader can act on: "Collections fell 4% week on week, driven almost entirely by two branches; the rest of the network was flat." A pipeline runs the query, applies statistical checks (period-on-period change, contribution analysis, outlier detection) and then asks a language model to write the explanation using only the computed facts. The model is a writer, not an analyst; every number in the prose is traceable to a calculation.

Good implementations are conservative. They report what the data shows, flag when a change is within normal variance, and avoid causal claims the data cannot support. They also respect the same row-level security as the underlying query, so the narrative never mentions rows the reader cannot see.

It is not a model free-associating over a screenshot of a dashboard, which is where hallucinated insights come from. It is also distinct from conversational analytics: narrative insights are pushed with a report, while conversational analytics answers questions on demand. The two are frequently combined.

• Founder / CEO: a Monday summary that says what moved and why is worth more than a dashboard you have to decode before coffee.
• CFO: variance commentary for the monthly pack can be drafted automatically from the same numbers, leaving finance to check and add judgement.
• Operations head: exceptions are surfaced in words, not buried in a heatmap, so the right person acts sooner.
• Data lead: the discipline of computing facts first and writing second is what keeps this trustworthy; insist on it in any vendor's design.

Narrative insights exist because charts do not explain themselves. Most people who receive a dashboard glance at it, fail to spot the important change, and move on; the analysts who could interpret it are too few to annotate every report. Automating the first pass of explanation gets the signal to the reader. The trade-off is the risk of confident nonsense: a model writing from raw output will invent causes. Separating calculation from writing, and limiting the prose to verified facts, is the difference between a useful summary and a liability.

• A SaaS weekly usage email to account managers that explains which customers' activity dropped and which features they stopped using.
• An NBFC portfolio report with commentary on which products and regions drove the change in overdue balances.
• A retailer's daily trading note summarising sales versus plan, promotional impact and stock-outs by category.
• A hospital operations digest explaining shifts in occupancy, average length of stay and theatre utilisation.
• A logistics network summary describing on-time performance movements and the hubs responsible.

Technique / practiceA generation technique that sits on top of computed analytics. Eazyware builds it within the natural language data querying service, with evals that check every number in the prose against the source calculation.

Eazyware service that covers it: Natural Language Data Querying. Starting prices are on the pricing page.