Critiquing an Executive Dashboard With AI Before It Ships

By the time a dashboard is ready for an executive review, you have looked at it so many times that you cannot see it anymore. You know where every number lives, so you stop noticing that the most important one is buried in the bottom-right quadrant. You wrote the axis labels, so the abbreviation that means nothing to a CFO reads as perfectly clear to you. This is the exact moment a vision model earns its place in the workflow: not as a designer, but as a stand-in for a viewer who has never seen the thing before.

I run a layout critique with Claude (the Opus tier, for its screenshot reasoning) or Gemini on every executive-facing dashboard before it ships. It takes about ten minutes and it consistently surfaces two or three problems I had gone blind to. What follows is how I do it, and — just as important — what I have learned not to trust it for.

What a vision model is actually good at here

Feed a model a full-resolution screenshot and it reliably catches a specific class of problem: things that are visible on the surface and do not require knowing your business.

• Visual hierarchy. Does the eye land on the headline metric first, or on a chart border, a logo, or a big block of filter chips? Models are good at reporting what draws attention and in what order.
• Alignment and spacing. Tiles that are two pixels off, inconsistent gutters, a legend that floats without a clear anchor. This is tedious for a human reviewer and easy for a model.
• Label and abbreviation clarity. Ask it to read every label aloud and flag anything ambiguous. "WoW" and "QTD" and "blended CAC" all get caught as jargon that an executive audience may not parse instantly.
• Redundancy and chartjunk. Duplicate KPIs shown two ways, gridlines fighting the data, a pie chart doing a bar chart's job.
• Cognitive load. A rough count of distinct visual elements and whether the view can be read in the five to ten seconds an executive will actually give it.

The prompt structure that works

A vague "what do you think of this dashboard?" gets you flattery. The model wants to be agreeable, and it will find nice things to say. You have to force it into a critical posture with a rubric and a persona.

Here is the skeleton I use:

You are a skeptical executive who has 8 seconds to read this dashboard
before a board meeting. You do not know our internal acronyms.

Walk through it in this order and be specific about location:
1. What is the single number or message you see first? Is that the
   most important thing on the page?
2. Read every text label. List any that are ambiguous, jargon, or
   unreadable at this size.
3. Trace your eye path. Where does attention go 1st, 2nd, 3rd?
4. List every element that could be removed without losing meaning.
5. Name the 3 most serious layout problems, ranked, with the fix.

Do not compliment the design. Assume it has problems and find them.

Two things matter in that prompt. The eight-second constraint forces it to prioritize like a real executive rather than a patient analyst. The "assume it has problems" line counters the sycophancy that otherwise produces a list of strengths you did not ask for.

I run the same screenshot through twice with slightly different personas — once as the skeptical executive, once as "a colorblind analyst reviewing this on a laptop in a bright conference room." The second persona catches contrast and color-reliance issues the first one glosses over.

A concrete example

On a revenue dashboard last quarter, the model told me the first thing it saw was a large green sparkline in the top banner — which happened to be a vanity metric (total sessions), not the number the CFO cared about (net revenue retention, which sat mid-page in a modest tile). That is a hierarchy failure I had built and could no longer perceive. I swapped their sizes and positions. It also flagged that three of my tiles used "MRR" while one used "recurring rev/mo," an inconsistency I had genuinely never noticed across four revisions.

Neither of those is a brilliant design insight. They are the obvious-in-hindsight mistakes that survive precisely because the person who made them cannot see them. That is the whole value proposition.

What it cannot do, and where it will hurt you

The failure mode to fear is a confident, well-written critique that is simply wrong. Treat everything the model says as a hypothesis to verify, not a verdict.

• It cannot judge whether your data is correct. It has no idea if your churn number is plausible, whether the aggregation is right, or if two charts contradict each other on the underlying facts. It reasons about pixels, not about your business.
• It hallucinates precision. It will say "the third tile is misaligned by roughly 6 pixels" with total confidence. The direction is often right; the number is invented. Use it as a pointer, then check by hand.
• It is weak on domain-appropriate defaults. It does not know that in your finance org, red always means "below plan" regardless of whether the metric going down is good. It may flag your correct color convention as a problem.
• It over-indexes on generic best practice. Ask it enough times and it will tell every dashboard to "reduce clutter" and "add whitespace." Weigh its specific, located observations far more than its general advice.

Fitting it into a real review cycle

I treat the AI pass as the step before human review, not a replacement for it. The sequence I use:

1. Build the dashboard to what I think is ship-ready.
2. Export a full-resolution screenshot at the actual delivery size (a Power BI screenshot scaled down lies about label legibility — capture at the real resolution the executive will see).
3. Run the two-persona critique. Read the output as a list of suspects.
4. Verify each flagged item myself. Discard the hallucinations, fix the real ones.
5. Take the revised version to a human — ideally someone from the actual audience.

The AI pass makes the human review more valuable, because you have already cleared out the embarrassing surface-level stuff. Your reviewer spends their attention on the questions only a human with context can answer: is this the right metric, does this tell the story leadership needs, is the framing honest.

A note on what you send

A dashboard screenshot is data. If it contains real revenue figures, customer names, or anything under NDA, run it through a tool your organization has actually approved for that class of data — an enterprise deployment with the right data-handling terms, not a personal account. When in doubt, blur or dummy the numbers before the critique; the layout feedback does not depend on the values being real. The model is judging where things sit and how they read, and a redacted screenshot supports that just as well.

Used this way, an AI critique is not a shortcut around design skill. It is a cheap, fast way to recover the outside perspective you lose the moment you start iterating — the difference between shipping the dashboard you built and shipping the dashboard an executive can actually read.