Chapter 10 — Practice-on-your-own solutions

Why Smart Narrative is the highest-risk of the four. Three reasons compound:

1. Natural language sounds authoritative. Prose is the register the audience uses for human-authored summary. When Copilot renders a Smart Narrative sentence, the audience reads it as «the analyst summarised the data this way» rather than as «the AI produced this text from the visuals on the page.» The register carries unearned credibility.
2. Reading order privileges the text. Left-to-right, top-to-bottom reading orders put the Smart Narrative paragraph before the audience reaches the charts that would contradict it. By the time the reader gets to the EMEA chart, their frame has already been anchored by the «growth across all regions» sentence — and confirmation bias helps them ignore the EMEA bar.
3. No data trail. A chart carries its data on its face — the reader can see the bars and check the axis. A Smart Narrative paragraph carries no traceability; the reader cannot ask «which chart did that sentence come from?» without opening the Copilot dialog and re-running the query.

Why the other three AI visuals are lower-risk in comparison. Q&A visuals show the query, so the reader can see what was asked. Decomposition Trees show the split path, so the reader can inspect the branches. Key Influencers shows the ranked drivers with p-values or scores, so the reader can weigh them. Smart Narrative alone strips the traceability that lets the reader challenge the artefact.

The specific edit or removal.

1. First choice: remove the Smart Narrative entirely. Replace it with a hand-authored one-sentence Big Idea that accurately reflects the underlying data. «Q3 overall on plan; EMEA missed by 12% and APAC satisfaction dropped 3 points — the two exceptions to the on-plan headline are on the page below.» That sentence is defensible because you authored it and can point to the charts that support each claim.
2. Second choice (if the Smart Narrative feature is mandated by policy or template): pin the Smart Narrative to a subset of visuals that support the sentence, and add a manual caveat. Configure the Smart Narrative to summarise only the region-level plan-variance chart, not the whole page. Then edit the resulting sentence to accurately name the exceptions, and add a footnote: «Summary generated by Smart Narrative, edited for accuracy against the underlying data; last verified [date].»

What not to do. Do not leave the Smart Narrative in place and hope the audience reads carefully enough to notice the contradiction. They will not. The audience trusts the register of prose more than the register of a chart, especially when the prose is confidently phrased. Every hour the confidently-wrong sentence sits on the page is an hour it is training decision-makers to distrust future charts from you when a downstream decision is anchored to a false summary.

Why the two people saw different paths. The Decomposition Tree’s AI-guided splits are filter-context-dependent. When the report is opened, the tree evaluates the current filter state — which page-level filters are active, which slicer selections are set, which bookmark auto-applied — and picks the «highest impact» split against that filtered slice. Two people with two different filter states will see two different «highest impact» splits from the same underlying data.

The specific mechanism. The colleague on Friday probably had a region slicer set to a specific region (or had drilled into a specific plant earlier in the session), leaving that filter active when they clicked the Decomposition Tree. The tree then evaluated splits against that filtered slice and surfaced plant 0477. Leadership on Monday opened the report cold with no slicers active — the tree evaluated splits against the full dataset and surfaced EMEA + plant 0631, which is what «highest impact» looks like without any filter narrowing. Both trees are technically correct; they are answering different questions.

The one-sentence verification move. «Before quoting a Decomposition Tree finding, screenshot the current filter panel state (all slicers, all page filters, all cross-filter arrows) and paste it into a companion note; when quoting, quote both the finding AND the filter state that produced it — or restate the filter state in the quote itself.»

What the correct Friday quote would have looked like. «With the region slicer set to APAC and the plant-family filter active on advanced-surgical, the Decomposition Tree shows plant 0477 as the highest backorder driver within that scope.» The scoped quote is defensible; leadership can reproduce it on Monday by setting the same filters, and the finding survives the reproduction test.

Why the discipline generalises. Every AI visual that responds to filter context (Decomposition Tree, Key Influencers, Q&A over a filtered page) carries the same reproduction hazard. The discipline is: the visual’s output is only defensible in the filter context that produced it. If you quote the output outside that context, you are implicitly claiming the finding holds under all filter states — and it usually does not. Named in the Chapter 10 discipline as the filter-context inheritance rule.