Choosing visuals
Charts are sentences: pick the form that states your finding, use AI to draft the spec, and refuse the visual sins that make executives distrust dashboards.
A chart is not decoration for data — it's a sentence with a claim. 'Gear web refunds tripled after April' is a line chart with a bend. 'The problem is one category' is a bar chart with one tall bar. If you can't say what sentence a chart speaks, the chart isn't done. That single test replaces most visualization theory.
- Change over time → line. The workhorse of 'when did it start' claims. Mark the bend (the April vertical line with a label beats a caption paragraph).
- Comparison across categories → horizontal bar, sorted by value, the interesting bar highlighted, everything else gray. Sorted-and-highlighted is half of all chart craft.
- Composition → stacked bar for a few parts; never pie beyond three slices. Your refund-reason mix over months is a stacked bar that tells the whole mechanism story in one image.
- Relationship → scatter. Rare in BA decks, devastating when apt (packaging weight vs. damage rate, if you get the audit data).
- A single number that matters → just the number, huge, with its comparison ('19.1% — was 8.0% in March'). The most under-used chart type is no chart.
AI's role: draftsman and critic, not artist
I need to show: [the claim, e.g. 'Gear web refund rate tripled starting April while store stayed flat']. Audience: COO, 10 seconds of attention. Recommend the chart form, what goes on each axis, what gets highlighted vs. grayed, the annotation text at the bend, and the title AS THE CLAIM (not 'Refund rates by channel' but the finding itself). Then tell me what the second-best chart would be and why yours beats it.
Titles-as-claims is the highest-leverage habit in this lesson: 'Gear web refunds tripled after the April packaging switch' as a title does more work than any amount of chart polish.
AI is also a ruthless chart critic: paste a screenshot or describe your draft and ask 'what will a skeptical reader misread here?' It reliably catches truncated axes implying drama the data doesn't have, dual axes implying correlation, and rainbow categories implying meaning that isn't there — the three sins that make numerate executives stop trusting your slides.
Bar charts start at zero (bars encode length; truncation lies). Line charts may zoom, but say so. Never plot two metrics on two y-axes to imply they move together. One deliberate-looking distortion, caught once, taxes every chart you show that audience forever. AI will happily generate dishonest charts if asked for 'impact' — the ethics live with the analyst.