# Prompt 11 — Direct an AI-Generated Artifact Through the Four Responsible-AI Moves > Adapted from Chapter 14 (Anand worked example, headcount-allocation dashboard for the CFO's Q4 planning meeting). The full audit pattern that runs after the CSAR loop, before any high-stakes AI-generated artifact ships. > > **Verified against**: Power BI Copilot, June 2026 surface (any Copilot-generated artifact bound for a recommendation or decision role) ## Crystallize An eight-line brief pasted into Copilot for Power BI: > "Build a one-page headcount-allocation dashboard for the CFO's Q4 planning meeting. Audience: CFO and finance office. Decision: which functions get the ten net-new heads budgeted for Q4. Measure: full-time-equivalent (FTE) count by function; overlay attrition rate and open requisitions. Comparison: headcount versus each team's approved scope, not raw count alone. Time: as-of the previous Friday's HR snapshot (do not run against the live semantic model). Grain: function-level; do not aggregate below function. Success criterion: the page has to be defensible if a regulator subpoenas it next year." Eight lines that name all six BRD fields plus the defensibility bar. The subpoena question is the tell: any brief whose success criterion invokes external audit is a brief that will need all four moves. ## Scope Copilot restates the audience, decision, measure, comparison, time, and grain. The first restatement often misses a constraint that reads as obvious to the analyst but is not surfaced in the brief — in the Anand case, Copilot read the ten net-new heads as a *redistribution* of the existing pool rather than as *additive*. Anand re-prompts to correct: > "To confirm: the ten heads are additive to the current FTE base, not a redistribution. The dashboard needs to make that additivity readable." Copilot restates with the correction. **Both restatements go into the provenance bundle** — the first-pass version documents the constraint that needed to be surfaced, and the corrected version documents the resolution. ## Assemble A one-page dashboard returns in about fifteen seconds: FTE by function (sorted bar), attrition rate overlay, open-requisition overlay, small treemap of function-to-team distribution, KPI card for total headcount. Looks defensible on first glance. ## Refine (Part I rubric corrections) - **Ch 4 (chart-family fit)**: The treemap answers a composition question the BRD does not ask. Remove. - **Ch 3 (Big Idea) + Ch 12 (hierarchy)**: The primary KPI needs to be the *recommended net-new allocation*, not the total headcount card. Make it visually dominant. - **Ch 12 (5-Visual Rule)**: With the treemap removed, the page is at five visuals. Hold the line. ## The four responsible-AI moves (Chapter 14) - **Move 1 — Provenance.** Assemble the bundle: the eight-line brief, both Scope restatements (pre- and post-correction), the HR semantic-model snapshot version (previous Friday), and the three Refine corrections against the Part I rubric. File to the shared folder with the governance-policy filename pattern. - **Move 2 — Fairness.** Three concerns in sequence. - *Representational*: add a secondary panel comparing headcount to each team's approved scope, so small-but-strategic functions (legal, security, DEI) are not buried by the ranking. - *Distributional*: click into the 8% aggregate attrition rate; the underlying data reveals early-tenure and mid-tenure women engineers at 14%. Add a callout for the subpopulation gap — the aggregate number would have driven a decision that ignored the more urgent retention problem. - *Procedural*: confirm the dashboard will be shared read-only with the function heads before the CFO meeting. - **Move 3 — Decision boundaries.** The artifact is *recommendation* (the CFO will make the call), which requires standard rubric + evidence-on-every-claim + provenance footer. Pre-write the anti-escalation clause into the artifact footer: > "This dashboard was built as a Q4-planning recommendation. If it is being read as the Q4 headcount plan of record, request the fairness audit and archived prompt history before quoting." - **Move 4 — Disclosure.** Add the standard AI-directed marking (*"AI-directed — built with Copilot, reviewed by the analyst against the rubric"*) plus one specific line naming the post-generation fairness callout: > "Fairness callout added post-generation; see notes for the subpopulation attrition context that the aggregate number understates." The four moves compound: Fairness (Move 2) changes what the decision-boundary escalation clause (Move 3) has to name; the escalation clause changes what the disclosure line (Move 4) has to disclose. Running them separately loses the compounding, and the issue that only the skipped move would have caught ships silently. ## Ship decision Ship. Deciding evidence: the four-move pass surfaced two substantive issues (the subpopulation fairness callout and the role-escalation clause) that would have shipped invisibly without it. The bundle is complete, the disclosure is honest, and the decision-boundary escalation clause names the risk explicitly on the artifact itself. See the [companion to Chapter 14](/ch/14/) for the full four-moves treatment. --- **When to use this prompt**: Any AI-generated artifact bound for a recommendation or decision role. Skip only for exploratory work the audience will not act on. The four moves are asymmetric — the cost of running them is a few minutes per chart when the artifacts are fresh; the cost of reconstructing them months later when a regulator, auditor, or executive asks is much larger. **Dataset**: [anand-headcount-fourmoves](/datasets/anand-headcount-fourmoves/) — download CSV + JSON to reproduce the four-move pass end-to-end.