Source: [AI-generated charts: define the comparison before the design — Skills for Business]() [← All articles](https://skillsforbusiness.ai/blog/) Workplace adoption · Guide # AI-generated charts: define the comparison before the design Before sharing an AI-generated business chart, check its population, denominator, underlying table, scale and headline. A worked example shows how. [Skills for Business](https://skillsforbusiness.ai/blog/editorial/)1 November 20267 min read On this page 1. [A correct count can answer the wrong question](https://skillsforbusiness.ai/blog/ai-generated-charts-define-the-comparison/#a-correct-count-can-answer-the-wrong-question) 2. [Separate extraction, calculation and drawing](https://skillsforbusiness.ai/blog/ai-generated-charts-define-the-comparison/#separate-extraction-calculation-and-drawing) 3. [Write the chart brief before the prompt](https://skillsforbusiness.ai/blog/ai-generated-charts-define-the-comparison/#write-the-chart-brief-before-the-prompt) 4. [Check the table before approving the graphic](https://skillsforbusiness.ai/blog/ai-generated-charts-define-the-comparison/#check-the-table-before-approving-the-graphic) 5. [Review the scale and headline together](https://skillsforbusiness.ai/blog/ai-generated-charts-define-the-comparison/#review-the-scale-and-headline-together) 6. [Make the next report reproducible](https://skillsforbusiness.ai/blog/ai-generated-charts-define-the-comparison/#make-the-next-report-reproducible) 7. [Sources](https://skillsforbusiness.ai/blog/ai-generated-charts-define-the-comparison/#sources) What to take away - Write down the business question, population and measure before asking AI to design a chart. - Inspect the chart-ready table separately from the graphic; counts and rates answer different questions. - Check scales and titles against the intended claim, and retain the source version and calculation. Before circulating an AI-generated business chart, agree on the comparison it must support and inspect the table behind it. A polished graphic can accurately draw the wrong measure. The practical safeguard is to make the population, calculation and intended claim visible before working on colors or slide layout. This matters when a manager asks an assistant to turn a service export, learning report or sales spreadsheet into an executive update. “Show which team is performing best” leaves several decisions unresolved. Best could mean handling the most work, meeting a service target most often, improving fastest or serving the hardest cases. Each requires different evidence. Our editorial recommendation is a short chart brief followed by three separate checks: the source selection, the calculated values and the visual claim. The hypothetical example below shows why those checks belong in the workflow even when the arithmetic looks simple. ## A correct count can answer the wrong question Imagine two service teams handling requests received during the same reporting week. This is an invented example, not a company case study or an AI product test. Each request has a unique identifier. Both teams use the same eligibility rules and definition of resolution within two business days. The report is prepared after every request in the cohort has had the full observation window. | Team | Eligible requests received | Resolved within two business days | Share meeting the target | | --- | --: | --: | --: | | A | 20 | 18 | 90% | | B | 100 | 90 | 90% | A count chart would show Team B resolving five times as many requests within the target: 90 compared with 18. That supports a statement about observed volume. It does not establish that Team B is five times as effective, works faster or uses staff more efficiently. Both teams meet the target for 90% of their eligible requests. Even the equal percentages do not establish equal underlying capability. The example contains no information about case difficulty, staffing, repeat contacts or longer-term variation. Twenty cases and 100 cases also provide different amounts of information. A small operational report should not quietly become a causal explanation or an individual employment ranking. Choose the output from the question. For workload planning, show the counts and make clear what was counted. For target attainment, show the percentages alongside their denominators. For this two-team comparison, the table may already be sufficient. Asking AI for a more impressive visualization does not create a better decision. ## Separate extraction, calculation and drawing The original [DEEPCHART paper](https://aclanthology.org/2026.emnlp-main.1627/), published in the October 2026 EMNLP proceedings, evaluates chart generation through distinct extraction, reasoning and visualization stages. It examines material from scientific papers, financial filings and ecosystem reports, making data fidelity an explicit evaluation target rather than relying on appearance alone. That research framing is useful here, with limits. The authors say their coverage does not exhaust chart-generation scenarios, and their value-matching metrics approximate numerical faithfulness rather than full semantic equivalence. This article does not treat the benchmark as an error-rate estimate for a particular business assistant or as validation of our proposed review process. For a business team, the transferable idea is to inspect the intermediate values. Request a chart-ready table before the graphic. It should expose the categories, numerator, denominator and calculated measure, with enough detail to trace a disputed number back to the approved source. A screenshot cannot reveal whether an assistant counted duplicate request IDs, filtered out unresolved cases or confused a precomputed percentage with a raw number. These are different defects from an incorrect label or color. Review them at the stage where they arise. If the input is a PDF, check the extracted headers and footnotes against the original page. If it is a spreadsheet, inspect filters, units and any supplied totals. A heading such as “response rate” is not enough to establish whether values are stored as fractions or whole percentages. ## Write the chart brief before the prompt The person requesting the report should own the definition of the business question. AI can help clarify the brief, but a fluent suggestion should not silently decide what the business means by performance. For the hypothetical service report, a usable brief would say: > Compare the share of eligible requests received by Teams A and B during the reporting week that were resolved within two business days. Count each request once, retain unresolved requests in the denominator and wait until the full observation window has elapsed. Show the numerator and denominator. Describe target attainment and volume without inferring productivity or causes. Add the approved source filename or export identifier and its retrieval time. Specify the timezone, business calendar, excluded request types and handling of transfers where those could change the result. The exact rules depend on the service; writing them down makes disagreement visible. Be particularly careful about the reporting window. “Requests closed this month” and “requests received this month that met the target” select different populations. The former can include old backlog; the latter needs follow-up time. Dividing one by a convenient monthly total can produce a plausible percentage with no coherent interpretation. If timestamps or status records are missing, report that limitation and its effect on the eligible population. Do not ask the assistant to fill gaps merely to produce a complete-looking chart. Resolve ambiguity with the data owner or narrow the comparison. ## Check the table before approving the graphic Recalculate the central result independently of the assistant’s explanation. In the example, that means checking 18 divided by 20 and 90 divided by 100, as well as reconciling the counts to the selected request records. A second paragraph from the same assistant repeating the percentages is not a separate arithmetic check. For a larger export, use an inspected query, spreadsheet calculation or script that another person can rerun against the same source. Retain the transformation alongside the output. Our guide to [AI spreadsheet formulas and row-level tests](https://skillsforbusiness.ai/blog/ai-spreadsheet-formulas-row-level-tests/) covers checking the row logic; the chart review adds whether the chosen aggregate answers the reader’s question. Check a boundary case that could change the comparison: a transferred request, a duplicate identifier or a request received near the reporting cutoff. This is an editorial recommendation for exposing definition errors, not a statistically sufficient validation sample. Where the decision is consequential, reconcile the full population and involve someone who understands the data. Do not accept a revised graphic until its table has also been revised. Otherwise the assistant can repair a caption while leaving the wrong values in the rendering code, or repair a calculation while retaining an outdated image. ## Review the scale and headline together Once the values are accepted, inspect what the picture asks the reader to compare. In its [practitioner guidance on bar and column baselines](https://www.datawrapper.de/academy/why-our-column-and-bar-charts-start-at-zero), Datawrapper explains why cutting off the numerical axis distorts comparisons based on bar length. It also discusses alternatives, including dot and line charts, that do not carry the same requirement. This is a chart vendor’s explanatory guidance, not independent testing of AI chart tools. For the hypothetical count comparison, bars should express 18 and 90 from a zero baseline. For target-attainment bars, show the percentage scale clearly; a 0–100% range is a straightforward choice here. Avoid turning a narrowly cropped axis into a visual claim about a dramatic difference. Equally, do not impose a universal zero-baseline rule on every line chart regardless of purpose. Text needs the same scrutiny. The [Office for National Statistics chart-text guidance](https://service-manual.ons.gov.uk/data-visualisation/guidance/chart-text) distinguishes the main message from supporting details such as the statistical measure, coverage and period, and calls for a source. These are useful reporting practices; ONS house-style requirements are not universal rules for business presentations. For this example, an evidence-matched headline is: “Both teams met the two-day target for 90% of eligible requests.” Supporting text should identify the receipt window, the request counts and the source. “Team B outperforms Team A” would exceed the evidence even if every plotted value were correct. Apply the same check to the slide narration and alternative text. They should not add a cause, forecast or performance judgment absent from the data. Supply an accessible table or equivalent text so the comparison does not depend entirely on recognizing colors or reading an image. ## Make the next report reproducible Keep the approved brief, source version, transformation, chart-ready table and final export together. Record who checked the data and what they actually checked. If nobody independently reconciled the source records, say so; naming a reviewer without describing the scope can overstate assurance. When the export changes, rerun the calculation and revisit the population rules. A stable prompt does not guarantee a stable interpretation when a column is renamed or a status category is added. If the graphic, table and narrative disagree, stop distribution until the discrepancy is resolved. Start with one recurring report. Ask its owner to state the comparison in one sentence, reproduce its main number and identify a plausible claim the evidence cannot support. Those three answers provide a concrete basis for deciding whether the AI-generated chart is ready to share. ## Sources & further reading Sources checked for this article. Research findings and vendor claims are distinguished in the text. 1. [DEEPCHART: How Far are LLMs from Faithful Data-Science Chart Generation?](https://aclanthology.org/2026.emnlp-main.1627/)Tang et al.; ACL Anthology · Accessed 1 November 2026 2. [Why our column and bar charts start at zero (or below)](https://www.datawrapper.de/academy/why-our-column-and-bar-charts-start-at-zero)Datawrapper Academy · Accessed 1 November 2026 3. [Chart details: Chart text](https://service-manual.ons.gov.uk/data-visualisation/guidance/chart-text)Office for National Statistics · Accessed 1 November 2026 Produced with AI-assisted research and writing. Examples and recommendations are our editorial synthesis unless attributed. This article does not claim independent product testing or human expert review. [Read our editorial policy.](https://skillsforbusiness.ai/blog/editorial/) ## Keep reading [AI-generated customer interviews: separate hypotheses from findings](https://skillsforbusiness.ai/blog/ai-generated-customer-interviews-hypotheses-findings/)[Document AI: verify extracted fields before drafting the brief](https://skillsforbusiness.ai/blog/document-ai-extracted-fields-source-checks/)