AI chart audit: check scales and percentage claims

An AI chart audit compares a numerical claim with its source before you trust the visual. After this exercise, you can check two fictional product series, distinguish a percentage change from a percentage-point difference, and save a clear accept-or-correct record. A polished chart is not evidence that its arithmetic is right.

AI means artificial intelligence. Here, a text assistant provides another reading of a small source table, not an authoritative statistical verdict. The task is deliberately narrower than generating a dashboard: check the values, the headline and the scale of one proposed column chart. You do not need to build a new graphic.

Content checked: 16 October 2026. All business figures below are fictional teaching material. No customer outcome, live Gemini trial or measured completion time is claimed. The optional assistant route assumes a working account and a desktop browser; signing in or setting up an account may take additional time. A calculator and a plain-text editor are enough for the core exercise.

Which source should you freeze first?

Imagine a small stationery supplier comparing two complete reporting weeks. Use source version V1 and keep the original numbers unchanged. The measure is items sold, not revenue, orders or profit. A transaction containing several items would therefore not count as a single item. This distinction belongs in the source note before any percentage is calculated.

These are fictional week labels rather than dated company accounts. No price, discount, cost or campaign exposure is supplied. A statement that advertising improved profit would need information absent from V1. Write the comparison question as total items and product mix across two complete weeks, so the chart cannot quietly switch to a different business question.

Keep the four values beside the proposed graphic. Check that each product label still belongs to the right number. A correct value attached to the wrong category is still a faulty chart. The AI data-analysis guide covers a broader workflow; this exercise ends with a traceable decision on one visual claim.

Which calculations should you do yourself?

Folders rise by ten items: 60 minus 50. Relative to the starting value, the increase is 10 / 50 × 100 = 20%. Ten is the absolute increase, not the percentage increase. Enter the calculation in a calculator and write both the formula and the unit in your record. Do not rely on an assistant repeating a plausible-looking percentage.

Markers fall by ten items: 140 minus 150. Their relative change is −10 / 150 × 100, approximately −6.67%. Meanwhile, total items remain 200 in both weeks. A headline saying overall sales volume rose by 20% is therefore wrong even though the folder calculation is correct. One improving product does not establish growth in the combined total.

Product share asks a different question. Folders represent 50 / 200 = 25% initially and 60 / 200 = 30% afterwards. The share increases by five percentage points. Relative to its initial share, the increase is 5 / 25 × 100 = 20%. Both descriptions can be numerically valid, but they do not mean the same thing.

If the starting value were zero, the usual relative-change formula would divide by zero. Record the absolute change and stop the percentage claim rather than inventing an infinite or 100% increase. Missing values are not zero either. The Excel prompt collection offers other calculation tasks, but each result still needs a source-level check.

What should you inspect on the chart itself?

Consider a proposed column chart showing folders at 50 and 60, with its numerical axis beginning at 45. The visible column lengths would be five and fifteen. That can suggest a threefold relationship even though the source ratio is 60 / 50 = 1.2. Ask for a zero baseline on this column chart before accepting its presentation.

ONS guidance uses a zero baseline for bar and area charts. A line chart is different: a restricted numerical range can sometimes make a trend clearer. Do not convert that into a rule that every chart must start at zero. First identify the chart type, then judge the scale in that context. If the type is unclear, mark the scale check unresolved.

Read the axis unit and the title separately. Items, dollars and percentages are not interchangeable labels. When comparable panels show the same measure, inspect whether their ranges match. If one panel spans 0–200 and another 0–60, equal-looking columns do not establish equal values. Note the actual ranges rather than describing a vague concern about appearance.

A text description lets an assistant assess the description only. It does not prove that labels are legible in the actual image or that the rendered chart follows the stated settings. Inspect the visual yourself at a readable size. The infographic review walkthrough checks a separate, non-numerical sequence and downloaded image.

How should you ask the assistant to review the claims?

Open gemini.google.com in a browser and sign in if necessary. Start a fresh conversation. Paste the request below, replacing the bracketed source with the complete V1 table and its unit. Include all three claims and the chart description. Nothing here requires uploading business records or granting access to a spreadsheet.

Review only this fictional V1 source: [paste the complete table, item unit and complete-week labels]. Claims: A: Folder volume increased by 10%. B: Total item volume increased by 20%. C: Folder share increased by five percentage points. Proposed column chart: folders at 50 and 60, numerical axis starts at 45. For each claim, show the source values, formula, result and accept/correct/insufficient-information decision. Do not substitute revenue for items or invent a cause. State that a text description is not a visual inspection. Do not create a new chart.

Before sending, verify that the placeholder is gone and all four values remain. When a response arrives, first check the numbers it used, then its arithmetic and wording. If the assistant introduces a new week or a missing price, that is a reason to correct the answer, not additional evidence. Google documents that Gemini can present inaccurate information as factual.

If access fails or the answer is delayed, continue with your calculator. Record the assistant stage as pending rather than pretending it passed. The deliverable is your checked decision record, not a successful chat interaction. The artificial-intelligence guide explains how to define a useful first task without treating the tool as the final authority.

Which decisions should your review record contain?

Accepting claim C does not accept the entire chart. Values, units, title and scale are separate checks. Give every row an explicit status and identify who needs to correct it. An empty cell means unfinished work, not tacit approval. Preserve the original claim beside the correction so another reviewer can understand what changed.

An incomplete reporting week needs a separate decision. Do not compare three observed days with seven days and label both complete weeks. If matching three-day source data are unavailable, defer the like-for-like comparison. A daily average is not automatically a repair: opening days, counting rules and missing observations still need to be defined.

In my earlier essay on effectiveness and perfection, I described how attention to every detail can delay projects. Applied here, the practical boundary is to finish one source-linked review before redesigning an entire reporting system. That is an application of a published working principle, not evidence of a live customer experiment or a measured efficiency improvement.

What should remain after fifteen minutes?

On Windows, open Start, search for Notepad and create a blank note. Copy V1, the claims, the chart description and your completed decision table into it. Add the review date, source version, reviewer and unresolved items. If you inspected only the description, explicitly say that the actual image has not yet been visually checked.

Press Ctrl+S, select Documents, enter chart-review-V1.txt and choose Save. Close the note and reopen it from that folder. Read the four source values, the 20% folder increase, the unchanged total and the five-percentage-point share change again. The file should explain the decision without requiring someone to recover an earlier chat.

A suggested allocation is three minutes for the source, four for arithmetic, five for the optional assistant comparison and three for saving and reopening. This is a planning allowance, not a measured completion promise. If the assistant is unavailable, use that part for the manual check and retain an honest pending status for the optional stage.

When a corrected chart arrives, review the changed wording and baseline again. If the source becomes V2, do not carry V1 approval forward automatically. The translation-number check addresses a different situation: preserving figures across languages. Here, the result is a documented numerical and visual review, not publication approval or a causal explanation.