Prepared by the Nadali editorial team using AI.
AI can turn a chart into a confident paragraph: “performance declined”, “the new process helped”, or “this channel is doing worse”. Before a management meeting, a small-business owner needs a different check: does each sentence follow from the numbers and conditions actually recorded in the report?
The short answer: approve the link between claim, data and limits, not the elegance of the explanation. Even a correctly drawn chart does not automatically establish why something changed. The editorial workflow below is for owners and managers of shops, hotels and service teams. It does not replace checking the underlying spreadsheet or a proper statistical study.
1. Give the assistant more than a picture
Keep the chart with its source table, export date and metric definition. If only a screenshot is available, do not let AI reconstruct exact values from bar heights. Ask it to flag unreadable labels and request the underlying values from the report owner.
Create a short metric card:
- What is counted: enquiries, orders, customers or another unit.
- Which records are included and excluded; whether one customer can have several enquiries.
- The numerator and denominator of any proportion.
- The beginning and end of the period, the event date and the time zone where it affects boundaries.
- Filters, units, rounding and whether the data are final or provisional.
This card need not contain personal records. Approved aggregates or a fictional dataset are enough for an initial commentary check. Use confidential information only within an authorised environment.
2. Separate counts, proportions and changes in proportions
All the following numbers are fictional. This is a teaching example, not a test of an AI system or a real business. Imagine two complete weeks, one consistent counting rule and two enquiry channels. “On time” means a first response within the agreed deadline; that deadline has not changed. Every case is ready to be assessed, and there are no missing values in this example.
| Channel | Week A: on time / all | Rate A | Week B: on time / all | Rate B |
|---|---|---|---|---|
| Chat | 72 / 90 | 80% | 9 / 10 | 90% |
| 2 / 10 | 20% | 27 / 90 | 30% | |
| Total | 74 / 100 | 74% | 36 / 100 | 36% |
The overall rate fell from 74% to 36%, a decrease of 38 percentage points. The relative decrease is (74 − 36) ÷ 74 × 100 ≈ 51.4%. These describe the same movement in different ways; they are not interchangeable numbers. “A 38% decline” correctly expresses neither calculation here.
Singapore’s Ministry of Manpower explains the distinction between percentage changes and percentage-point changes, and why the comparison base matters. In a working caption, show the starting and ending values, not just a striking change figure. If the starting value is zero, the usual relative-change calculation divides by zero and is unsuitable. Do not replace the result with an arbitrary “100%”.
3. Check what the overall bar conceals
Chat’s on-time response rate rose by 10 percentage points; email’s did too. Yet the overall rate fell because the enquiry mix shifted: email accounted for 10 out of 100 cases in week A and 90 out of 100 in week B. In this dataset, the channel with the lower rate acquired much more weight.
“The on-time response rate fell in both channels” is therefore wrong. “The overall on-time response rate fell” is correct but incomplete without the mix change. Do not replace the overall rate with the simple average of the two percentages: (80 + 20) ÷ 2 = 50%, not the actual 74 out of 100. For a combined rate, add the relevant numerators and denominators when the categories do not overlap and use the same definition.
The practical response is to show the manager the total and channel breakdown together. That supports investigating email workload, but it does not prove that any individual employee performed better or worse. We still know nothing about enquiry complexity, staffing hours or repeat contacts.
4. Do not confuse visual scale with the size of the change
Check axis limits, units, legends and the order of periods. For bar charts, ONS recommends starting the numerical axis at zero: bar length should convey magnitude without distorting proportions. If that scale does not serve the purpose, consider a different chart type rather than exaggerating the difference.
Do not mechanically apply this rule to every line chart. For this review, the important point is that AI describes the numbers, not an impression of a “near-vertical collapse”. Compare separate charts only after checking that their scales and definitions match.
Review the period as well. Two days of the current week and seven days of the previous one are not equivalent totals. If you decide to compare matching days, label that exact comparison. Do not select the beginning of a chart simply because it makes the desired conclusion more persuasive.
5. Keep missing information visible
In a real report, a missing outcome is not zero. If some enquiries are still awaiting a response, or some data have not arrived, establish the counting rule before writing the conclusion. Do not let the assistant silently drop those records or fill a line as though the missing section were observed.
Record how many cases have unknown outcomes and explain what this means for comparability. If the treatment has not yet been agreed, label the result provisional. Inspect interactive tooltips and filters too: a screenshot may not reveal that one channel has been hidden.
6. Remove explanations the data do not establish
If a new response template was introduced in week B, that does not make it the cause of the change. Statistics Canada distinguishes association from causality: variables moving together do not, by themselves, explain why they do so.
A defensible commentary for our example is: “The overall rate of first responses within the agreed deadline fell from 74% to 36%. Rates rose in both channels, but email, which has the lower rate, accounted for a larger share of all enquiries. The reasons for the mix change and the effect of the template have not been established.”
This neither excuses nor praises the team. It marks the boundary of the evidence. A causal conclusion needs additional information and an appropriate evaluation design.
7. Approve a short commentary with a review record
A reusable AI instruction is: “Using the table and metric card, draft three sentences: an observation, an important limitation, and the next question to investigate. For each number, identify the rows and calculation. Distinguish percentages from percentage points. Do not explain causes without evidence or reconstruct missing values.”
A responsible person checks the arithmetic independently, reads the caption without the image, and confirms that the meaning has not changed. Government Analysis Function recommends providing a meaningful text alternative or accessible data table for charts. An important limitation should not be available only through a colour or the shape of a line.
Keep the table version, approved commentary, corrected claims and the responsible person’s name within the team. Repeat the check whenever a filter, period or metric definition changes. The useful output is an explanation someone can reproduce from the data, not the model’s confidence.
Read also: checking an AI-generated spreadsheet formula and measuring time in a small AI pilot.
Author: Nadali. Sources checked on 29 October 2026. The workflow, teaching dataset and example commentary are original editorial proposals, not a report of a completed test. You need access to the underlying data and someone who understands the metric definition; a chart screenshot alone may not provide enough information. Cover: a reused original Nadali AI illustration, not a real business report.
