How do I review a misleading dashboard? Fix one with four questions

Review a dashboard with four questions and fix the chart that would send Lucía's calls to the wrong site.

Free course. Your Maestro, an AI, .

You critique a dashboard the same way you design one: starting from the decision. Before you weigh in on colors or fonts, ask these four questions, in order. They're the checklist for reviewing any dashboard: yours, a coworker's, or the one your AI builds for you:

Who is it for, and what do they decide?
If the dashboard doesn't make clear who uses it and for what decision, everything else is opinion. Look for the purpose sentence, even if it's only implied.
Does every number have a definition and a comparison?
Do you know what was counted and what it was divided by? Does every number have its target or its previous period next to it?
Does every chart answer its question?
Write down the question each chart claims to answer and check whether it really does: with the right measure (counts or rates) and in a form that's quick to read.
Does the layout follow the decision?
The answer on top, then the where and the since when, the detail last. Also check the period and that each category keeps its color.

The dashboard you inherited

Lucía inherited this dashboard from the person who had her job before. It has two charts: on the left, a pie chart titled Which site has the problem this week?; on the right, the clinic's weekly rate with the target. Run it through the four questions.

You can guess the decision: which site gets the calls. The line on the right is fine: it answers "how is the clinic doing against the target?" with the right measure and a reference point. The problem is the third question: the pie claims to answer "which site has the problem?", but it slices up the count of no-shows. Centro gets the biggest slice because it's the biggest site. If Lucía trusts the pie, she sends the calls to Centro, and Cholula, the site losing one appointment in six, goes another week with nothing done.

Check yourself: Why does the pie chart on this dashboard lead to a bad decision?

Try it

The editor holds the dashboard as : the chart's description in , which the page draws on the right. The first item in the hconcat list is the pie. Fix it with three changes:

Change "mark": "arc" to "mark": "bar".
Change theta to x, using the tasa (rate) field instead of inasistencias (no-shows), with "axis": {"format": "%"} so it reads as a percentage.
Change color to y, with the sede (site) field and "sort": "-x" so the site with the highest rate ends up on top.

The data already has the tasa column; you don't need to calculate anything. You're done when you see three horizontal bars on the left and Cholula's is the longest. If the editor flags an error and you can't find it, paste it to your Maestro.

Your turn

Turn the pie on the left into horizontal bars of the no-show rate by site, sorted from highest to lowest. Don't change the chart on the right. You're done when Cholula is the longest bar and sits on top.

Chart spec (Vega-Lite JSON)

{
  "hconcat": [
    {
      "data": {
        "values": [
          {"inasistencias": 17, "sede": "Centro", "tasa": 0.101},
          {"inasistencias": 12, "sede": "Angelópolis", "tasa": 0.098},
          {"inasistencias": 14, "sede": "Cholula", "tasa": 0.171}
        ]
      },
      "encoding": {
        "color": {"field": "sede", "title": "Site", "type": "nominal"},
        "theta": {"field": "inasistencias", "type": "quantitative"}
      },
      "mark": "arc",
      "title": "Which site has the problem this week?"
    },
    {
      "data": {
        "values": [
          {"semana": "2026-06-29", "tasa": 0.099},
          {"semana": "2026-07-06", "tasa": 0.101},
          {"semana": "2026-07-13", "tasa": 0.096},
          {"semana": "2026-07-20", "tasa": 0.086},
          {"semana": "2026-07-27", "tasa": 0.097},
          {"semana": "2026-08-03", "tasa": 0.099},
          {"semana": "2026-08-10", "tasa": 0.1},
          {"semana": "2026-08-17", "tasa": 0.101},
          {"semana": "2026-08-24", "tasa": 0.098},
          {"semana": "2026-08-31", "tasa": 0.115},
          {"semana": "2026-09-07", "tasa": 0.101},
          {"semana": "2026-09-14", "tasa": 0.112},
          {"semana": "2026-09-21", "tasa": 0.115}
        ]
      },
      "layer": [
        {
          "encoding": {
            "x": {
              "axis": {
                "format": "%d/%m",
                "values": [
                  {"date": 29, "month": 6, "utc": true, "year": 2026},
                  {"date": 13, "month": 7, "utc": true, "year": 2026},
                  {"date": 27, "month": 7, "utc": true, "year": 2026},
                  {"date": 10, "month": 8, "utc": true, "year": 2026},
                  {"date": 24, "month": 8, "utc": true, "year": 2026},
                  {"date": 7, "month": 9, "utc": true, "year": 2026},
                  {"date": 21, "month": 9, "utc": true, "year": 2026}
                ]
              },
              "field": "semana",
              "timeUnit": "utcyearmonthdate",
              "title": "Week",
              "type": "temporal"
            },
            "y": {
              "axis": {"format": "%"},
              "field": "tasa",
              "title": "No-show rate",
              "type": "quantitative"
            }
          },
          "mark": {"point": true, "type": "line"}
        },
        {
          "encoding": {
            "y": {"datum": 0.1}
          },
          "mark": {
            "strokeDash": [4, 4],
            "type": "rule"
          }
        }
      ],
      "title": "How is the clinic doing against the target?"
    }
  ]
}

Edit the JSON; the preview updates when you pause.

How this maps to Altair

Altair builds this same JSON: chart.to_json() prints it. "mark": "point" is alt.Chart(df).mark_point(). Each encoding channel is an argument to .encode(), such as x=alt.X("field:Q", scale=alt.Scale(type="log")) or color="field:N". Each transform entry is a call such as .transform_filter("datum.year == 2007").

Run it yourself

With bars, the dashboard tells a different story: Cholula on top, with almost twice the rate of the other two. The data didn't change; what changed is the measure the chart puts in front of you.

Notice what you did not have to touch: the line on the right. Critiquing a dashboard isn't rebuilding it; it's finding the piece that leads to a bad decision and fixing that piece. At work, that difference matters: a dashboard people already know gets fixed one piece at a time, and each change gets a one-line explanation ("we switched the site chart from counts to rates, because counts made the small sites look better").

If you have a dashboard at work, run it through the four questions with your Maestro.