Planning dashboard — before and after AI-assisted review

The dashboard on the left is what the planning team was looking at before the review pass. Thirty-seven KPIs. Four time-series charts. Two pivot tables. Everything technically present; nothing prioritised.

The prompt chain ran in three stages: first a full read of the dashboard specification, then a critique pass asking what a senior reviewer would cut, then a rebuild pass using only the ten metrics that survived. The result is on the right.

What changed: twenty-seven metrics removed, not because they were wrong but because they weren’t decision-relevant. The two time-series charts that remained were reframed around the actual planning decision (reorder point vs. current stock position) rather than around what was easy to pull from the system.

What this took: Forty minutes, one Claude session, one well-structured prompt chain. The AI didn’t understand the planning context — it was given it. The prompt included the decision the dashboard was supposed to support and the audience who would read it. That context is what did the work.

What it didn’t do: The AI had no visibility into the underlying data. It reviewed the structure and the framing, not the numbers. A human still had to verify that the ten surviving metrics were correctly calculated. Automation handles the structure; judgement handles the accuracy check.

Failure mode: The first version of the critique prompt was too aggressive — it flagged twelve metrics for removal that were actually load-bearing for a downstream process the prompt didn’t know about. The second version added a constraint: “do not suggest removing any metric that feeds into the weekly S&OP pack.” That produced a more useful result.