Engineering Health Dashboard

A shipped executive report was circulated for review, and the feedback that came back was about framing, trust, and altitude rather than visuals. At the same time, a company-level pivot toward AI token cost control reframed what the report needed to say, and the CEO asked for it to be updated in that new context. The stakes were unusual: the report is both a product surface and a sales artifact, with an MVP timeline measured in days.

Primary audience: C-suite executives, VP Engineering and above, with the CTO as the exemplar reader. The format constraint drove everything: a static print-style brief, emailed standalone, with no hover states or progressive disclosure to lean on. The second cycle ran on ten objectives, condensed: reframe the whole report around one governing question ("How is agentic AI adoption impacting delivery?"), standardize AI terminology into three work models, reframe financial impact from savings to additional capacity, convert the adoption ladder into a funnel, show trends over snapshots, and cut the sections that didn't carry their weight.

Before any layout work, every metric was traced to the platform data inventory and graded: measured, estimated, partial, or unavailable. A precedence rule kept it honest: the live endpoint outranks the documentation where they disagree, which downgraded one headline metric on the spot. Three findings shaped the design. The flagship acceleration multiplier is a ratio whose denominator does not exist anywhere in the data layer. The counterfactual ("spend that would have occurred without AI") is not capturable, so a headline number had to be killed and replaced with a knowable forward figure. And no budget facts existed at all, so budget utilization could not be claimed, only spend.

The narrative-first wireframe grew from the objectives and data docs only; the existing shipped design was deliberately withheld so the new layout could not be anchored by the old one. The spine was question-as-header ("Is AI spend under control?", "How much work is produced?"), with a masthead carrying a data-basis line and a "three questions this brief answers" box. The spec then layered in full feedback coverage: 63 unique feedback tags placed 522 times across six sections, from four sources, plus the review layer, open-calls box, and scope badges.

The first persona audit ran four independent, context-free reviewers: CEO, CTO, CFO, and UX. Their best findings were the kind that save you in a boardroom. The most-quoted headline card said AI spend fell while the report's own cost bridge said it rose. A 10x denominator error appeared in three different costumes across the report. The dollar layer and the unit-economics layer described two different companies. And the one trust-establishing line, the data-basis statement, was an empty bracket.

Stakeholder feedback wasn't a punch list. It was thematic: no orientation layer; a weighting unit that read as a black box; adjacent charts covering different time periods that would not reconcile ("readers will try"); metrics that got worse undermining the story in prospect settings; and a headline figure blending hard cash recovery with modeled capacity under one banner. By process rule, stakeholder feedback does not patch the current cycle. It starts a new one, as an update project, back at step one. The follow-up working session changed the report's governing question, its vocabulary, its financial frame, and its section inventory. That is a redesign, not a revision.

The redesign collapsed six sections to four. The adoption ladder became a funnel anchored on total developer count, so percentages finally had organizational context. Snapshots became trends. The cost section was replaced by a normalized per-work-area table. "Savings" became "additional capacity" throughout, including a glossary entry stating the value is extra output for the same investment, not budget reduction.

The second audit introduced the mechanic that made the redesign defensible: a line-by-line transcript coverage audit run before the personas were spawned. Prior cycles had reviewed the wireframe in isolation; this pass walked 64 numbered transcript items with timestamp, verbatim quote, and the exact rendered element that answers each one, so nothing was silently dropped. A section scorecard then ran two independent passes per section: transcript verification, and executive landing graded by three context-free personas. Its headline tension is the most useful artifact in the set: verification scores in the 90s against landing scores averaging 4 out of 10. Proof the design was on-brief, and that the remaining gap was data and decisions, not more wireframing.

The final triage: nine findings implemented, three escalated as open calls, four parked for high fidelity, and four dismissed with a written reason naming the settled decision each one re-litigated. An arithmetic footing pass verified every within-section relationship programmatically; the one cross-section relationship that didn't foot was disclosed in the masthead and logged as an open call rather than papered over.