InSpin had 2,000 answers from its workers and an Excel file nobody wanted to open.
End-to-end: research, information architecture, high-fidelity UI, accessibility.
Overview First, Then Zoom, Then Details
14: Decisions traced to evidence
Scope, Constraints & Reality
The 2,000 rows are generated to match InSpin's question types, so the patterns are plausible, not proven. Wherever a decision depends on how real analysts behave, the document flags it as something to test.
Published work on how people read charts, NN/g, Google and Microsoft guidance on dashboards and AI, the WCAG 2.2 spec, and the public help docs of five tools that already do this job.
They work in ops, workforce or research. They're comfortable with numbers, and they've probably opened the Excel file once and closed it again. The tool has to be readable without training.
Whether workers were promised anonymity isn't in the brief, so setup asks. Anonymous is the default: groups under 5 stay hidden everywhere, and answers open only for people who said yes to a follow-up. An original version with no privacy setting is kept in case InSpin never promised it.
Design Decisions
Overview First, Then Zoom, Then Details
Shneiderman's 1996 rule is the whole information architecture: an Overview, segments to zoom and filter, then the respondent record on demand. The Overview shows four numbers and two lists, and everything else is one click away. Categorical questions are sorted horizontal bars, because people read length more accurately than angle, so the donut chart was dropped.
AI That Can Be Checked, Not Just Cited
Theme cards show a short AI summary labelled as one, then real quotes word for word. Each theme list ends with a coverage line, like “96% of answers fit a theme. Read the other 4%.” NN/g found participants never clicked AI citations, so every Ask AI answer ends with a plain “based on” line and an Open this view button that lands on the same charts you'd trust anyway.
Segments Apply Instantly And Stay Visible
Someone exploring survey results shouldn't press Apply. Picking a segment updates every chart at once, the applied filters stay on screen as chips, and a live count reads “Showing 248 of 2,000.” Below about 30 people the screen warns that percentages are unstable, and an empty result tells you which segment to remove.
Impact
Each row of the evidence-to-decision table pairs a published source with the specific thing it changed in the design.
Dovetail, Qualtrics Text iQ, SurveyMonkey, Typeform Smart Insights and Marvin, each with what was borrowed and what was deliberately done differently.
Writing up the contrast checks found that segment chips and secondary buttons used a 1.75:1 border. They moved to the darker control border. A user-test plan for the next round is in the document.