How to Check AI-Generated Charts Against Their Data

Before sharing a chart made by an AI agent, verify the source rows, the calculation and the visual encoding separately. A chart can reproduce every supplied number and still mislead through a changed denominator, an omitted category or a cropped axis. An attractive export is evidence of presentation quality, not evidence that the conclusion follows from the data.

Ask for the chart's underlying table and transformation steps alongside the image. Then check the labels, scales and final export against that evidence. This guide provides a practical acceptance method for business reports, with explicitly hypothetical numbers. It is not a benchmark of charting models or a claim that a particular visual style is appropriate for every dataset.

Start with the source file or query and the population it represents. Record the time period, filters, exclusions and unit of observation. One row could mean a customer, a session, an order or a support case. A chart labelled “customers” is wrong if its source counts sessions, even when the underlying aggregation is otherwise accurate.

Ask the agent to list transformations in order: filtering, joining, deduplication, grouping and calculation. A join can multiply rows, while deduplication can remove legitimate repeated observations. Neither operation should be invisible merely because it is common in reporting. Preserve the source and the transformed table so a reviewer can follow how the plotted values were produced.

For a hypothetical sales report, cancelled orders might be excluded from revenue but included in an operational workload count. Both choices can be legitimate for different questions. The acceptance rule should specify which question the chart answers. Our spreadsheet-cleanup guide explains why consistent-looking data does not necessarily retain its original meaning.

• Name the unit represented by a row.
• Keep filters and joins visible beside the plotted table.

Rates need both a numerator and a denominator. Do not average percentages unless that operation matches the question and weighting. Two regions with very different sample sizes can produce a combined rate that is far from the simple average of their individual rates. The chart should show or make available the underlying counts so the result can be recomputed.

Consider the hypothetical example below. Region A has 10 successes from 100 attempts; Region B has nine from 10. Their rates are 10% and 90%. The overall rate is 19 divided by 110, approximately 17.3%, while the unweighted average of the regional rates is 50%. Those values answer different questions, and a title that calls both “overall success” conceals the difference.

Also distinguish percentage points from relative percentage change. Moving from 20% to 25% is a five-percentage-point increase and a 25% relative increase. Preserve the intended measure in the title, axis and commentary. The companion guide on tables contradicting their formulas covers numerical consistency; a chart adds another layer where the same numbers can be framed incorrectly.

A chart type determines what comparisons are easy to see. Bars make category magnitudes easy to compare; a line emphasizes movement across an ordered sequence. A pie or stacked composition suggests parts of a whole. If categories overlap or do not exhaust the population, a whole-based display can imply a relationship the data does not support.

Ask what the reader should compare before choosing the visual form. For a hypothetical monthly report, a line may help show a trend, but it should preserve the actual time order and gaps. Sorting month names alphabetically can produce a smooth-looking line whose sequence is meaningless. The agent should not choose the chart solely because it looks more polished than a table.

Keep the simplest representation that supports the intended decision. Sometimes a short table with counts and rates is clearer than a graphic. If a chart removes the sample size, uncertainty or category definitions that the decision depends on, return those details to the presentation. Visual simplicity should remove distraction, not the evidence needed to interpret the result.

The Office for National Statistics guidance recommends zero baselines for bar and area charts because the filled length or area represents magnitude. It allows cropped axes for line and scatter charts when appropriate, with clear context. This distinction is more useful than a blanket instruction that every chart must start at zero.

For a hypothetical pair of bars showing 95 and 100, starting the axis at 90 can make the difference appear far larger than the ratio of the values. A line chart examining small movements around that range may reasonably use a narrower axis, but the scale must be legible and the title must not overstate the change. Review the visual claim as well as the printed labels.

Do not assume the plotting library's default survives generated code. Vega-Lite's scale documentation shows that zero inclusion depends on the scale and configuration. Custom domains can change behavior. Inspect the actual specification or code, then view the rendered chart; a comment saying “start at zero” is not evidence that the axis does so.

A missing observation is not necessarily a zero value. Replacing missing data with zero can invent a collapse in a time series or suggest that a category had no activity. Dropping it silently can create the opposite problem: a line may connect observations across a period with no measurement, making continuity appear stronger than the evidence supports.

Define the missing-value rule before plotting. A blank cell might mean not collected, not applicable, withheld or a failed query. Keep those meanings separate where they affect interpretation. An agent should not infer the intended replacement from the surrounding pattern merely to produce a smooth chart. If a documented statistical treatment is used, describe it and preserve the untreated source.

Review categories that disappear during filtering. A report about all regions should not quietly remove the region whose data feed failed. The missing region may be the most important operational fact in the report. A note, explicit gap or separate completeness indicator can preserve that information without pretending the underlying value is known.

• Do not convert absent observations into measured zeros by default.
• Show consequential exclusions and gaps in the chart or its note.

The title, axis labels, legend and annotation should describe the same measure. A chart of total revenue should not be titled average order value. A forecast should not be labelled as an observed result. These errors can enter when an agent writes the narrative separately from the code that computes the chart.

Use precise units and time boundaries. State whether currency values are nominal or adjusted when that distinction matters, and whether a rate is per session, customer or another unit. Avoid unsupported causal language. A line rising after a campaign began does not, by itself, establish that the campaign caused the increase. The chart can describe the timing without asserting an experiment that did not occur.

Keep the source note useful and compact. Name the dataset or query, its relevant date and any essential exclusions. The research-claim checking guide applies here: a source can be real while failing to support the exact conclusion printed above it. Review the title as a factual claim rather than as decoration.

Inspect the artifact the audience will actually receive. A correct interactive view may export with clipped labels, missing legends or a different aspect ratio. A tooltip carrying a crucial definition disappears in a static image. If that definition is necessary to understand the chart, include it in a caption or accompanying table that survives the export.

Review the chart at its intended size. Tiny text that looks acceptable on a large editing canvas may be unreadable in a report column or mobile screen. Check that colors and labels still distinguish the series, and that the visual does not depend on color alone. Preserve an accessible text explanation or data table for people who cannot use the image.

Keep the exported file connected to the exact table and code that generated it. If an operator manually edits one label after export, record that change or regenerate from source so future updates remain reproducible. A chart that cannot be traced to its inputs becomes difficult to correct when someone questions a value later.

• Check the final dimensions, labels and legend.
• Retain a table and explanation that survive static export.

A compact review record should identify the source, transformed table, calculation, chart specification and final export. Check representative values independently, including the largest, smallest and any value central to the conclusion. When the chart contains a derived rate, recompute it from the underlying counts rather than comparing two outputs produced by the same agent.

Treat the result as a sequence of linked checks. Correct source selection does not establish correct aggregation; correct aggregation does not establish honest scaling; honest scaling does not establish a readable export. If one step fails, repair that step and recheck the downstream artifacts it affects. This keeps review focused without assuming that one visual pass proves everything.

Our AI transformation service helps teams make these acceptance rules part of their reporting workflow. The agent can produce charts quickly, but the useful outcome is a chart whose claim can be followed back to the evidence. Preserve that chain and the report becomes easier to trust, update and correct.

• Recompute consequential values from source counts.
• Keep the exact data, code and export together for later review.