Write a weekly commentary from agreed metric thresholds
The Monday commentary has eight paragraphs, one per metric, and six of them say a number moved slightly. Everyone has stopped reading it, including the person who writes it.
What you'll make
a short weekly exception report naming the metrics that crossed thresholds agreed in advance, the ones that didn’t, and the ones that couldn’t be assessed, with a check table behind it.
What you'll need
a ChatGPT, Claude, Gemini or Copilot chat in your work account, a small table of this week’s and last week’s totals you’re allowed to use, the agreed threshold for each metric, and about thirty minutes. Aggregate figures can still be commercially sensitive; use fictional numbers if real ones aren’t cleared.
Example result.
Two of seven assessed metrics breached their agreed thresholds.
Return rate rose from 6.2% to 7.1%, up 0.9 percentage points.
Support tickets rose by 46 to 318, about 16.9%.
Stock-outs rose from 7 to 9 but are unassessed: no rule agreed.
The other five assessed metrics stayed within their gates.1. Write the rules before you look at the week
Do this on a different day from the commentary. Thresholds chosen after seeing the data are not thresholds; they are a story you have decided to tell.
For each metric, write the unit, the direction (rise, fall or either) and the boundary, including whether a change equal to the limit counts (> or >=). For counts, the owner may want two gates, a relative one and an absolute one, both required. That stops two complaints becoming four from reading as a 100% crisis. For rates, use percentage points: 6.2% to 7.1% is 0.9 points, which is a 14.5% relative rise, a different number. Write a metric definition your team can check covers definitions in depth.
Metric:
Unit and calculation:
Trigger: rise / fall / either / level
Threshold, with > or >=:
Two gates? AND or OR:
Missing data and zero baseline:
Owner and rule date:If a metric has no rule, it’s unassessed. Don’t let the assistant invent one.
Why this works: a rule fixed in advance lets a quiet week say nothing, and makes a breach mean something.
What you should see: one filled rule per metric, or a clear “no rule agreed”.
2. Check the input is comparable
Use full, comparable periods with the same filters and definitions. Missing isn’t zero. A previous value of zero can’t be the base for a percentage change; note it for its own rule. If last week was a holiday or a definition changed, flag the pair as not comparable before you start.
Why this works: a perfect threshold applied to two incomparable weeks gives a confident wrong answer.
What you should see: a table of this week, last week and the rule, with any non-comparable pairs marked.
3. Ask for a check table, then the commentary
Use only my data and rules. First return a check table with one row per metric: current value, previous value, absolute change, relative change where defined, rule, and status: BREACH, WITHIN AGREED GATES or NOT ASSESSED.
Apply > and >= exactly, and AND/OR exactly. A missing value or missing rule is NOT ASSESSED. Don't compute a percentage change from zero or invent a threshold. For rates, show the percentage-point change and label any relative change separately.
Then write a short commentary listing breaches and unassessed items. If nothing assessed breaches, say "No assessed metric breached its agreed threshold." Don't suggest causes, normality, significance or actions.
[data, period notes and rules][data, period notes and rules] is your table from step 2 with the rules from step 1.
Why this works: the full check table shows every metric, so a silent omission can’t pass for a quiet week.
What you should see: a table with every metric exactly once, then a few sentences.
4. Recompute every row and add your sentence
For a small table, check every change and gate with a spreadsheet or calculator: current minus previous, then that difference divided by previous, times 100. Test thresholds on unrounded values and round only for display. Confirm the count of assessed metrics leaves out the unassessed ones.
Then add what only you know. The assistant was told not to explain; you aren’t, and you’re the one who knows about the packaging change, the bank holiday or the supplier who went quiet. Label a hypothesis clearly, and say what you’ve done about it.
Why this works: fluent arithmetic is easy to believe, and your sentence is the part of the report nobody else can write.
What you should see: a check table you’ve verified line by line, and one labelled human sentence.
Check the result
- Every metric appears once, as a breach, within gates, or unassessed.
- Two-gate rules were applied as AND where the rule says AND.
- Rates are reported in percentage points.
- The commentary contains no cause the data doesn’t show.
The usual failure: a two-gate rule applied as either-or. One gate is crossed, the other isn’t, and the metric is reported as a breach. Spot it by checking every two-gate row against both limits. Then:
Correct the assessment for [metric]. The checked change is [value and unit] and the agreed rule is [exact rule]. Explain which gate is met and which is not. Recalculate the breach count and the commentary without changing any thresholds or adding causes.Fill in the metric, your recomputed change and the rule as written.
Worked example: two of eight
Fictional example. A homeware wholesaler, week ending 4 October 2026. Where a rule has two gates, both must be strictly exceeded.
| Metric | This week | Last week | Rule |
|---|---|---|---|
| Orders | 2,412 | 2,498 | >5% AND >100 |
| Net revenue (£) | 281,400 | 290,100 | >5% AND >£15,000 |
| Average order value (£) | 116.67 | 116.13 | >4% AND >£2 |
| Return rate | 7.1% | 6.2% | >0.5 points |
| On-time dispatch | 94.2% | 96.0% | >2 points |
| Support tickets | 318 | 272 | >15% AND >30 |
| First response (minutes) | 142 | 118 | >20% AND >30 minutes |
| Stock-outs (SKUs) | 9 | 7 | Not agreed |
The first draft said three metrics breached, including “first response time up 20.3%”.
- The failure. First response rose 24 minutes, about 20.34%. That passes the 20% gate but not the 30-minute gate, so under the AND rule it isn’t a breach. The correction prompt fixed it, giving the example result: returns (+0.9 points) and tickets (+46, about 16.9%) breach; orders (−86, −3.4%), revenue (−£8,700, −3.0%), order value (+£0.54, +0.5%) and dispatch (−1.8 points) stay within their gates; stock-outs are unassessed.
- The human sentence. The ops lead checked returns and tickets by hand, confirmed all eight metrics were accounted for and added: “Returns and tickets both relate mostly to the Kitchen range. My working assumption is the new packaging that went live on 14 September, but I haven’t confirmed it, and I’ve asked the warehouse to photograph the next twenty returns.”
- What it doesn’t say. Nothing about causes in the assistant’s text, nothing about whether returns are at a record, and nothing claiming the within-gate changes are harmless. Select each metric in Try it, below, to see how the two gates work.
Try it
An AND rule needs both conditions
Select a metric to see the absolute and relative changes against its agreed rule.
Both strictly more than 20% and more than 30 minutes.
First response: 118 → 142; +24 min, +20.34% relative change. Does not meet the strict exception rule.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Optional training: Defending a quiet week
The first time the commentary says nothing breached, someone senior will ask what you’ve been doing. Keep the full check table, so a quiet week visibly contains eight checked metrics rather than an absence. Publish the thresholds once, with the date they were set: a threshold with a date is a decision; one without is an opinion. Review them on a fixed cycle, quarterly rather than whenever something annoying slips through. If a stakeholder wants something flagged that the gates missed, that’s a request to change a gate, not to add a paragraph. Ask which one and by how much.
And don’t let the thresholds have the last word. A major customer leaving or a system outage may not show as a percentage move at all. “No assessed metric breached” describes what you checked; “nothing material happened” is a different claim.
Optional training: A holiday week and noisy metrics
Before comparing a bank-holiday week, check the pairs:
Check these period and definition notes before comparing the metrics. Identify any non-comparable pair and the information needed to compare it. Do not calculate a change for those pairs or invent an adjustment.
[period notes]For metrics that are noisy by nature, a cheap alternative to week-on-week is to compare against the highest and lowest of the last eight weeks and flag only moves outside that band. You may find two or three metrics never needed a weekly gate at all.
Optional training: Moving the checks into your reporting tool
Spreadsheet formulas or BI rules can calculate the gates before AI writes the commentary. Tableau Pulse metric definitions include a goal or threshold and a “Value going up is” setting (favourable or unfavourable), with separate alerts; these aren’t automatically the same as a two-gate weekly rule, so check the feature against your exact rule first. For a progress update without numeric thresholds, see Use ChatGPT to draft a weekly update from your notes; for a one-off note from a dashboard, Turn a dashboard into a short decision note.