Automatic recommendation reads data types, not analytical intent

What it is

BI tools' "smart charts" and Show-Me panels recommend from the data schema: one categorical plus one numeric variable → bar chart; two numeric variables → scatterplot. Useful, but with a ceiling: analytical intent is not in the schema. For the same "categorical + numeric," one analyst wants magnitude comparison (bars optimal), another wants composition shares (stacked or pie), a third wants a single category's trend over time (line) — the intent lives in the analyst's head, not in the field types. What the engine can deliver is a legality floor (a chart that commits no encoding error), not the question's optimum (lowest reading cost for the actual question).

Why it happens

The recommender's inputs are field types and cardinality (dimension/measure, discrete/continuous, cardinality high or low), mapped to that type's default safe chart. Missing intent produces two systematic biases: convergence on generic charts (bars, lines, and pies dominate recommendation rates because they are never quite wrong under any intent) and convergence on showing everything (recommendations tend to use every dragged-in field rather than serve one question). The rules themselves are sound — Mackinlay-style automated design ranks encodings by expressiveness, exactly these systems' principle — the fault is handing over "expressively viable" as if it were "task-optimal."

Where it stops holding

Recommendation genuinely earns its place at: the first glance at unfamiliar data, non-experts' starting point, and guarding against elementary encoding errors (categorical palettes on continuous fields and the like). It cannot replace: question definition, reading-cost trade-offs within the viable set (previous leaf), or splitting compound questions. The sound division of labour is recommendation proposing candidates and the human fixing intent — treating the recommender as a draft generator, not a decision maker. The reverse automation (question templates generating charts) requires intent to be input explicitly, which still means a human defining the question.

Applying it

• Integrate recommendation as a "candidate panel with one-click switching," not auto-rendering; keep intent prompts beside the default chart ("compare sizes? see composition? track a trend?").
• Team practice: every auto-recommended chart passes the question check before delivery (does it directly encode the comparison being asked?).
• Verification: take the last ten recommended charts and annotate their actual intents; if more than about three in ten are "legal chart, mismatched intent," the recommender has been promoted to decision maker.

Related

• Same group: U2.17.1 A mismatched chart forces readers to compute the answer mentally · U2.17.2 One question, several viable charts — they differ in reading cost · U2.17.4 Choosing the chart first manufactures questions to fit it · U2.17.5 As data magnitude shifts, a once-fitting chart stops working
• Nearby: U2.01.2 Define the question before choosing the chart · U1.02.1 Automated design ranks channels by expressiveness
• Search terms: chart recommendation · Show Me · automated visualization · expressiveness