If you fix this one thing.

The most-shared paragraph in every User Health Report is the one that begins with "If you do one thing this week." VPs of AI take that line, drop it into a Slack message to their CEO, and the conversation moves. We treated this as the load-bearing moment of the entire deliverable, and it took several iterations to get right.

The first version of the report was structurally similar to a senior-analyst memo: environmental headline, observations, anomalies with reasoning chains, falsifiable predictions, recommended interventions. We shipped it to early design partners and watched what they did with it. The pattern was clear: they read the headline, scanned the anomalies, and then went looking for the actionable recommendation. When the recommendation list had three to five items of roughly equal weight, they would forward the whole thing to a teammate and ask "which of these should we actually do?" When we gave them a single prioritized action with the projected behavioral impact attached, they forwarded that single line and a decision happened.

The "If You Fix This One Thing" section codifies that pattern. Every weekly memo ends with one specific action: a one-line directive ("Route any conversation containing ‘cancel’ to a human after one failed bot turn"), the qualitative expected impact ("Refund-conversation frustration buildup drops from 0.71 to below 0.50 within 14 days"), and where the evidence base supports it, the projected behavioral lift ("Trust calibration would lift ~11 points across 23% of users") and a confidence rating.

The discipline that makes this work is in what the prompt refuses to do. The named cohort is required (not "users" but "users on the cancel flow"). The named turn pattern is required (not "responses" but "the first AI turn after a refund query"). The dimension being moved must be one of the six Layer Z dimensions, and the magnitude must be derivable from the period’s averages. If any of these are missing, the prompt sets prioritized_action to null rather than fabricate. Better to surface no winner than to invent one.

The dollar impact projection lights up later — once you have business-outcome data connected (Stripe / Salesforce / HubSpot), the engine can express the projected lift in revenue terms with a confidence range. Today the field is null and the projection stays in behavioral terms. We chose to ship the percentage path now rather than wait for the dollar path because the percentage version of the projection is already enough to drive a Monday-morning conversation.

The "stranger test" is the validation gate: a non-BIE engineer reads the prioritized action and says "I would build that this week." If they would not, the prompt failed. We hold the bar at 4-of-5 prioritized actions passing the stranger test before exposing the surface to non-design-partner customers, and the founder reviews the first 5 outputs personally on every connector type before that gate opens.

The reason this specific section gets so much weight is that it is the single moment where a behavioral signal becomes a defensible decision the VP of AI can bring to their finance meeting. Everything else in the memo justifies the action. The action is the deliverable.