Data & Reporting
AI Prompt Templates

The analysis · One undefined word turned a 3% fall into a 12% rise

A metric definition states exactly how a number is calculated: numerator, denominator, grain, time window, filters and source. A metric dictionary holds the agreed definitions so every report uses the same one.

• 18,500 paused subscribers counted as active: the April migration mapped the “paused” status to “active”. The board figure of 138,000 included them.
• Up 12% or down 3%: last June there were 123,200 active subscribers. 138,000 is 12% more; 119,500, the correct figure, is 3.0% fewer.
• Three reports, three definitions: the board pack counted status, finance counted anyone billed in 60 days, and the correct definition counts billed and not paused within 35 days.

Takeaway: agree and write down every metric’s definition before reporting it, and restate past figures when the definition changes.

The analysis · The data was $410K and 2,140 customers away from the truth before analysis began

Data quality assessment checks data for completeness, duplicates, valid values and consistency. Reconciliation compares the data’s totals with a trusted source and explains every difference.

• 2,140 customers counted twice: the migration created a second record for every customer who changed plan during the cut-over. The subscriber table held 2,140 more records than there were customers.
• $410K above the ledger: the billing export for Q2 shows revenue $410K higher than the ledger, because refunds are recorded in a separate table that the export does not include.
• 3.6% with no cancellation date: these cancellations cannot be placed in a month. They are flagged, not deleted, and the churn figures state the gap.

Takeaway: reconcile every dataset to a trusted total and write down each cleaning rule before any analysis starts.

The analysis · A lookup that never failed visibly priced 1,140 subscribers on the wrong plan

A spreadsheet audit checks a workbook for hard-coded values, faulty lookups, incomplete ranges and inconsistent formulas. An exact-match lookup returns an error when it finds nothing, instead of silently returning the nearest value.

• 212 typed numbers: each month someone overwrote cells that did not calculate as expected. The workbook still looked complete, but 212 cells no longer changed when the data did.
• 1,140 subscribers on the wrong plan: VLOOKUP with approximate match returns the nearest plan code when there is no exact one. Plans created after the migration had new codes, so 1,140 rows took the next plan’s price.
• One cell that says TRUE: the rebuilt workbook compares its revenue total with the reconciled billing figure from prompt 02 and shows TRUE only when they match.

Takeaway: use exact-match lookups that fail visibly, remove every typed number from formula ranges, and add a check cell that ties to the source.

The analysis · One join added 23,760 rows and 8% to average order value

The grain of a query is what one row represents: one order, one shipment, one customer. A fan-out join joins tables of different grain and repeats rows, so any sum after it is inflated.

• 297,000 orders became 320,760 rows: 8% of orders are sent in two parcels. Joined at shipment level, each of those orders appeared twice and its value was counted twice.
• $34.56 against $32.00: the reported average order value was 8% too high. Two pricing decisions in Q2 used the inflated figure.
• A row count would have caught it: the corrected query checks that the row count equals the order count after every join, and stops if it does not.

Takeaway: state the grain of every query, check row counts after every join, and reconcile the result before anyone uses it.

The analysis · Churn rose because the mix of new subscribers changed, not because all subscribers changed

Exploratory data analysis examines distributions, outliers and segments before any conclusion is drawn. Segmentation compares measures between groups, such as acquisition channel or plan.

• 11.2% against 5.1%: subscribers who joined through a promotional offer cancel more than twice as often each month as those who joined at full price.
• From 18% to 44% of sign-ups: promotions brought in 44% of new subscribers in 2026, up from 18% in 2025, so overall churn rose even though neither group changed much.
• 9% of revenue in 0.4% of accounts: 520 office accounts order far more than households. Left in, they pull every household average up, so they are reported separately.

Takeaway: look at segments before the average, and separate the outliers before reporting any single figure.

The analysis · The loss sits between the first and third order, in the promotional cohorts

Cohort analysis groups customers by the month they started and follows each group over time. Funnel analysis measures how many customers pass from each stage of the journey to the next.

• 71% to 58%: three months after starting, 71% of the 2025 cohorts were still subscribed; for the January to March 2026 cohorts, 58% were.
• 41% for promotional subscribers: only 41% of promotional subscribers who placed a first order went on to a third, against 58% overall. The free gift arrives with order one.
• About 3,900 subscribers: 30,000 people joined from January to March 2026. A 13-point gap at month 3 is about 3,900 fewer subscribers from those cohorts alone.

Takeaway: measure retention by start cohort, and fix the stage where the loss is concentrated before buying more sign-ups.

The analysis · $1.5M was approved on a result that could easily have been chance

An A/B test compares a control with a variant on randomly split groups. Statistical significance measures how likely a difference this size would be by chance; a confidence interval gives the range the true difference probably lies in.

• p = 0.21: a difference this large would appear by chance about one time in five even if the offer had no effect. The usual threshold is one in twenty.
• −0.3 to +1.2 points: the confidence interval includes zero, so the offer may have no effect at all. It also includes a real gain, which is why the answer is to rerun it, not to drop it.
• 6,000 against 26,000: the test stopped at 6,000 customers per group. Detecting a 0.5-point difference reliably needs about 26,000, roughly five weeks at current traffic.

Takeaway: fix the sample size before a test starts, and report a result whose confidence interval includes zero as no evidence of a difference.

The analysis · 31 tiles and no definitions meant nobody trusted the one number that mattered

A dashboard requirements document states who uses a dashboard, which decisions it supports and exactly what it shows. A KPI specification gives each key performance indicator its definition, target, owner and data source.

• 24 tiles with no decision: the CEO, marketing and finance named nine decisions they make each month. 24 of the 31 tiles supported none of them.
• One definition, two places: active subscribers on the dashboard now uses the same dictionary definition as the board pack: 119,500, not 138,000.
• Published only if it reconciles: each morning’s refresh is checked against the billing total before it appears. If it does not tie, the dashboard shows the previous day with a warning.

Takeaway: build the dashboard from named decisions, and give every KPI one definition and one owner.

The analysis · One of five movements in August was real, and it was predicted

Performance commentary explains what moved in a period and why. Normal variation is the range a measure moves within from month to month without any real change; movements inside it are noise.

• 5.6% churn, outside the range: monthly churn has moved between 4.9% and 5.4% for a year. 5.6% is outside that range and needs an explanation.
• The driver was already known: the April to June 2026 cohorts, the largest promotional intake of the year, reached month 3 in July to September. The cohort analysis showed month 3 is where promotional subscribers leave.
• Four movements are noise: active subscribers −0.5%, average order value +0.4% and two other KPIs moved within their normal range. The report says so in one line each.

Takeaway: report which changes are outside normal variation and explain only those, stating plainly that the rest are noise.

The analysis · The correction comes first, and the $0.9M still unspent is the decision

An answer-first write-up states the conclusion and recommendation before the evidence. It gives a confidence level for each finding and states the limits of the data.

• The correction in the first paragraph: the June board pack reported +12%. On the agreed definition, active subscribers fell 3.0%. The write-up says this before anything else.
• $0.9M still unspent: of the $1.5M approved in June, $0.6M has been spent. Pausing the remaining $0.9M costs nothing until the test is rerun.
• High, medium and low: high confidence in the subscriber count (reconciled to billing), medium in the churn driver (cohort evidence, no cancellation reasons before May), low in the offer’s effect (no evidence either way yet).

Takeaway: open with the answer and any correction, give each finding a confidence level, and end with a decision, an owner and a date.