On Friday afternoon, an operator sends AI a dashboard screenshot, a spreadsheet, and a few chat messages: “Write this week’s review.” Minutes later comes a fluent report: reach grew, conversion improved, and strong content should be scaled. The word “grew” may combine paid and organic traffic; “conversion improved” may come from a tiny sample in one channel. The smoother the writing, the easier it is to skip the first question: where did these numbers come from?
AI is excellent at organising confirmed data into a weekly report. It is not qualified to choose metric definitions or invent reasons. Start with a one-page metric brief that freezes the number, time window, source, and comparison baseline. Then keep confirmed facts, testable hypotheses, and open questions separate.
Define the observation window
“This week” is not always Monday through Sunday. Ad platforms use account time zones, ecommerce orders can have refund delays, and content data keeps accumulating after publication. State the start and end time, time zone, export time, whether today is included, and whether comparison is the previous calendar week or the previous seven days. Keep only five to eight metrics tied to the goal.
Give every number an identity card
Use a row for each metric: name, current value, comparison value, change rate, source link or report name, filters, owner, and data status. Use statuses such as checked, pending refresh, estimated, or anomaly under review. “Registrations +18%” should also show 354 versus 300, channel scope, time window, and deduplication rule. Absolute numbers and definitions help readers judge a percentage change.
Put cause, evidence, and speculation in different columns
Separate confirmed drivers, testable hypotheses, and external context. A landing-page release with matching timing may be a confirmed driver. A topic that may lead to more saves is a hypothesis needing a breakdown. A holiday or industry event is background, not causal proof. AI can write these clearly; it must not turn “may” into “therefore.”
A prompt for weekly reporting
You are an operations-analysis assistant. Write a weekly report from the metric brief and fact record below. Preserve all numbers, time windows, metric definitions, and data statuses. Do not calculate metrics that are not provided. Use sections for results, key changes, confirmed drivers, hypotheses to test, risks and data limits, and next-week experiments. Keep confirmed drivers separate from hypotheses. Every action needs an owner, validation signal, and deadline. Avoid unsupported words such as significant, major, or strong.
Use ChatGPT or Claude only after removing customer emails, order IDs, account credentials, and non-public revenue data. The model can structure and rewrite; sensitive data still follows your company policy.
Replace “keep optimising” with an experiment
A useful action answers four questions: what changes, for whom, how it will be compared, and what result counts as evidence. For example, change the first registration-screen section from a feature list to a use-case explanation, keep traffic from the same channel, then compare completion rate and form exits. If the sample is insufficient, do not conclude. AI can propose experiments, but it should not promise uplift or confuse correlation with causation.
Run a number audit before sending
Compare this weekly report with the metric brief. Do not polish it. List numbers that differ from the brief, comparisons without sources, hypotheses written as conclusions, claims without a time window or sample size, and actions missing an owner or validation signal. Return the original sentence and the information needed to fix it.
Have a data owner sample-check two or three core values at the source. The valuable part of a report is not the visual polish; it is that anyone can trace a conclusion back through a number, source, and definition.
Release checklist
- Does every core metric have a value, baseline, time window, and source?
- Are absolute numbers shown with percentage changes, and are small samples labelled?
- Are facts, hypotheses, and external context clearly separated?
- Does each next action have an owner, validation method, and deadline?
- Has sensitive data been redacted and access to source links limited appropriately?
Once the metric brief becomes fixed input, AI can genuinely shorten reporting time: it organises evidence into a clear narrative while the team confirms definitions, sets priorities, and tests hypotheses. If exports are messy, start with the AI spreadsheet cleanup workflow.