Ask AI to explain a difficult chart without inventing its data
Someone forwards a dashboard screenshot with two axes, no data labels and the question "what's going on here?". Ask an AI tool and it will often answer to one decimal place, axis or no axis.
What you'll make
a short explanation of a chart image that separates what it can read, what it’s estimating and what it would need the source data for.
What you'll need
a screenshot of the chart, the question your reader is asking and, if you can get it, the table behind the chart. ChatGPT, Claude, Gemini and Microsoft 365 Copilot Chat (with a work account) all accept a pasted or uploaded image. A text-only chat can’t see a chart it was never given, so attach the image itself rather than describing it. Allow about 10 minutes.
Example result.
Readable: bars are monthly orders on the left axis (0 to 4,000). The line is average delivery days on the right axis, which starts at 2.5, not zero. Estimated from gridlines: orders went from about 2,000 in January to about 3,000 in June. Delivery time went from about 2.7 to about 3.5 days, a rise of roughly 30%. Not shown: exact values, and any reason for the change. To be exact, I need: the monthly table behind the chart.
1. Ask for an inventory before the story
Paste the chart into a new chat and ask for what’s on it before any interpretation:
Before explaining this chart, list what you can read on it: title, each
axis with its unit and where it starts, which series uses which axis,
categories, data labels, footnotes and filters. Mark anything you can't
read, or that isn't shown, as "not shown". Don't interpret anything yet.Why this works: most chart mistakes start with a misread axis. An inventory makes the model commit to what it saw, and it gives you a short list to check against the image in a few seconds.
What you should see: a list you can tick off against the chart, including where each axis starts and a plain “not shown” for anything missing, such as data labels.
2. Ask for the explanation your reader needs
Once the inventory matches the chart, ask the real question:
Now explain this chart for [reader] who wants to know [question].
Describe the relationships you can see. Give any number you read from
the chart as an estimate, labelled "read from the chart", and say which
axis it comes from. Don't give a cause unless the chart states one. List
the source data you would need to give exact figures.[reader] is who the explanation is for, such as “the operations manager”. [question] is what they asked, such as “whether delivery is getting worse as orders grow”.
Why this works: the labels keep a rough reading from turning into a precise-sounding fact once someone pastes it into a report.
What you should see: estimates with “read from the chart”, each tied to an axis, plus a short request for the data behind it.
3. Swap in the source table when it arrives
If the chart’s owner sends the underlying data, give it to the same chat:
Here is the table behind the chart. Check that its period, categories,
units and filters match the chart. Replace your estimates with the table's
values and show any calculation. If the table and the chart disagree,
list each disagreement instead of picking one. Still don't add a cause.Why this works: a related spreadsheet isn’t always the chart’s source. Checking the period and filters first stops you replacing a rough estimate with an exact number for the wrong thing.
What you should see: exact values with their working, and a list of any mismatches between the table and the chart.
Check the result
- Every number in the answer is either labelled as an estimate or comes from the table.
- Each series is read against its own axis, and the answer says where each axis starts.
- No unit, date range or category appears that isn’t on the chart.
- Any reason for a change is phrased as a question to investigate, not a finding.
The usual failure: the answer reads a line against the wrong axis, or treats a truncated axis as if it started at zero, then adds a confident cause. “Delivery time doubled because of staff shortages” is the classic version. Spot it by asking where each axis starts and checking the numbers against the gridlines yourself.
Check your explanation against the chart. For each number, name the axis
you read it from and where that axis starts. Correct any change that
assumed an axis starts at zero. Remove or rephrase any cause the chart
doesn't show, and keep the relationships you can see.Worked example: orders and delivery time on two axes
Fictional example. An operations lead pastes this dashboard tile into a chat and asks “Is delivery getting worse as orders grow?”
Title: Orders and average delivery time, January to June
Bars: orders per month, left axis 0 to 4,000, gridlines every 1,000
Line: average delivery days, right axis 2.5 to 4.0, gridlines every 0.5
No data labels
Footnote: Excludes marketplace ordersThe first answer said:
Weak: “Orders rose by 50% and delivery time doubled, from 2.7 to 5.4 days. The warehouse clearly couldn’t keep up with demand.”
The repair prompt brought back the version in the example result: delivery went from about 2.7 to about 3.5 days on the right axis, roughly 30%, and no cause is shown. When the source table arrived, step 3 gave delivery of 2.7 and 3.5 days and orders of 2,040 in January and 2,980 in June, up 46%. The table also had a June total of 3,410 that includes marketplace orders. Because the chart’s footnote excludes them, the model used 2,980 and listed the difference rather than choosing silently.
- The line looked like it doubled. On an axis starting at 2.5, it climbs from near the bottom to the middle. Asking where each axis starts catches this.
- 5.4 days isn’t on the chart. The right axis stops at 4.0. A number outside the axis range is a quick sign the model has invented a value.
- The table didn’t quite match. Picking the table’s 3,410 would have quietly changed what the chart measures. Listing the mismatch lets the chart’s owner say which version is right.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Check percentage denominators
A percentage label needs its population and period. If either is missing, ask for the definition before comparing the chart with another one. “Up 12%” of a quarter’s orders and “up 12%” of one region’s orders are different claims. The overlapping segments guide covers the case where the groups share members.
Ask for a readable replacement chart
Once you have the source table, a clearer chart often answers the question better than an explanation of a muddled one. Two separate charts with shared months usually beat one dual-axis chart. The accessible chart guide shows how to add direct labels and a text summary so the meaning doesn’t depend on colour or a legend.
Turn a hypothesis into a question for the owner
A possible reason can still be useful, as long as it’s labelled. Ask the model to list up to three hypotheses, each with the evidence that would confirm or rule it out, and send that list to whoever owns the data. “Did marketplace orders move to the main warehouse in April?” is a question someone can answer. “The warehouse couldn’t keep up” is a guess that sounds like a finding.