Write a weekly report that changes a decision

A report can be accurate sentence by sentence and still mislead

“Enquiries rose 50%, response time improved and qualified leads increased.” It sounds like a good week. But the enquiry total includes a new spam source, response time counts automatic acknowledgements, and the team changed what qualifies as a lead on Wednesday.

AI is good at turning a table into fluent prose. Before asking it to explain the business, decide what the figures mean and which comparisons are valid. A report should help someone choose what to do, not simply make every metric sound like a story.

This guide uses a fictional service-business report. The numbers are teaching examples. The task is to decide whether to expand a campaign or repair the enquiry process first, using a small amount of evidence that the team can actually inspect.

The denominator changes the conclusion

All enquiries increased by 50%. Genuine enquiries did not increase. Confirmed qualified enquiries rose from 12 to 14, but six current cases remain unresolved. The confirmed qualification share is 12/40, or 30%, versus 14/40, or 35%, with unresolved cases separately disclosed.

Do not exclude the six unresolved cases from the denominator simply to make the current week look stronger. If you also report qualification among completed reviews, label that measure and its smaller denominator explicitly. Neither ratio proves the campaign caused a change.

The useful question becomes: are the unresolved cases waiting because the team lacks capacity, information or a clear qualification rule?

Keep a small metric dictionary beside the report

For each important measure, name the source, period, unit, inclusion rule and owner. “Enquiries” might mean unique submitted forms, conversation threads or people. Those are different measures. A person using LINE and the website should not become two customers unless the report deliberately counts channel contacts.

Define response time from the customer's relevant request to the first substantive response during the service hours you promise. If you report automatic acknowledgement time, label it separately. An immediate bot greeting does not establish that the customer's question was answered.

Use a fixed reporting timezone and disclose incomplete periods. A Monday report assembled before all weekend records arrive should not compare its partial current week with a fully reconciled previous one. Identify late-arriving records and decide how the next report will restate them.

Keep definitions stable across weeks. When a rule changes, annotate the break and, where possible, recalculate the comparison under one rule.

Ask the assistant to challenge the input first

Before requesting the summary, ask for a source check: missing periods, duplicate records, changed definitions, unknown statuses and totals that do not reconcile. Keep the output small enough that a human can resolve the material gaps.

Suppose the assistant finds that the advertising dashboard attributes 18 enquiries while the CRM has only 14 linked records. Do not make them agree by copying one number into the other. Different attribution windows, unmatched identifiers or failed submissions could explain the difference. The report should state which measure each system supplies and which reconciliation remains open.

For important claims, retain a source reference or query result. A citation to the whole dashboard is less useful than the specific saved view, filter and extraction time. Keep private customer data in the controlled source system rather than embedding it in a widely shared management memo.

Replace the celebration with a decision memo

The first AI draft says: “The new campaign is working, with enquiries up 50% and stronger qualification. Increase the budget.” That confuses volume with quality and correlation with cause.

A corrected memo reads: “Incoming enquiries rose from 40 to 60, but the additional 20 were classified as spam. Genuine enquiry volume remained 40. Confirmed qualified enquiries were 14 versus 12 under the same definition; six cases are still awaiting review. We do not yet have evidence to expand spending. The operations owner will resolve those six cases and inspect the spam source before the next campaign decision.”

This is not a claim that the campaign failed. It is a decision to wait for a specific piece of evidence. The proposed work has an owner and a purpose. The report does not need a longer narrative to become more useful.

Keep fact, explanation and action in separate sentences

A fact is an observed number or event with a source. An explanation is a proposed reason. An action is what someone will do next. Put them next to one another without pretending they have equal certainty.

For example: “Six cases remain unreviewed” is a fact if the queue confirms it. “The new intake form may be missing a required detail” is a hypothesis. “Inspect those six records and compare their missing fields” is a useful action. The inspection may show staffing, not the form, caused the delay.

Include evidence that would change the decision. If the six cases are strong opportunities and the backlog is cleared, the campaign may deserve another controlled test. If most are outside the service area, the targeting or page explanation may need revision.

Small weekly changes are noisy. Avoid inventing statistical certainty from two short periods or attributing every movement to the team's latest edit.

Use a one-page report brief

Decision this report must support:
Period, timezone and data cutoff:
Metric definitions and source references:
Facts we can verify:
Comparisons that are not valid, and why:
Missing or unresolved records:
Possible explanation, clearly labelled:
Recommended action and owner:
Evidence that would change the decision:
Next review condition:
Do not invent causation or treat unknown values as zero.

When using an assistant, give it the metric dictionary and checked totals before the previous week's prose. Otherwise it may preserve the old narrative while replacing the numbers. Ask it to identify the sentence with the weakest evidence, then inspect that sentence yourself.

Microsoft's data-cleaning guidance is a useful reference for preparing tabular inputs. The decision-memo structure here is our editorial recommendation. For your next weekly report, remove any metric that does not inform a current decision and resolve one denominator before asking AI to write the summary.

Sources and dates

Sources checked on 20 October 2026. Worked examples are teaching examples unless explicitly attributed. Provider documentation describes product behaviour; the operating recommendations are DLVX editorial judgement.

• Microsoft: top ten ways to clean your dataSpreadsheet cleaning reference; the worked dataset and reconciliation rules here are fictional teaching material. Checked 2026-10-20.