Recipes for AI
Risk & Underwriting

The "Risk Narrative Builder" (Prompt)

Raw model metrics are not a risk story. This turns data into a stakeholder-ready narrative.


CONTEXT

Risk teams often sit on data that leadership doesn't act on — not because it's not important, but because it's not framed accessibly. This prompt turns a table of metrics into a narrative that non-risk stakeholders can read, understand, and make decisions from. Useful for monthly risk reviews, credit committee prep, and investor reporting.


THE PROMPT (Copy & Paste)

Act as a chief risk officer preparing a risk review briefing for senior leadership.

I'll give you raw credit performance data, model output metrics, or portfolio statistics.

Your job is to:
1. Identify the top 2–3 risk signals in the data (both improving and deteriorating)
2. Write a 3-paragraph narrative explaining what's happening, why it matters, and what's driving it
3. Produce a risk register entry with: Risk Name, Likelihood (1–5), Impact (1–5), Trend (↑↓→), and Mitigation Action
4. Suggest one leading indicator we should start tracking if we're not already

Tone: Precise, not alarmist. Frame risks without minimizing them.

[PASTE PORTFOLIO OR MODEL PERFORMANCE DATA HERE]

EXAMPLES

Raw Data:

Nov vintage NCO: 14.1% (vs. 11.5% target). Dec vintage: 13.8%. Model Gini on recent cohorts: 0.38 (was 0.44 six months ago). Approval rate held flat at 31%. Self-employed segment showing 2.1x NCO vs. employed segment.

Result:

Risk Signal 1 (Deteriorating): Model discriminative power has declined materially — Gini falling from 0.44 to 0.38 over six months suggests the model's ability to separate good from bad borrowers is weakening. This is a leading indicator of future vintage performance issues.

Risk Signal 2 (Segment-Specific): Self-employed borrowers are defaulting at 2.1x the rate of employed borrowers. If this segment has grown as a share of originations, it may explain a meaningful portion of the NCO drift.

Narrative Paragraph 1: Recent vintage performance is tracking above our NCO target for the second consecutive month, with November at 14.1% and December at 13.8%. While December shows marginal improvement, both vintages are meaningfully outside the 11.5% target, suggesting this is a portfolio mix issue rather than noise...

Risk Register Entry:

FieldValue
Risk NameCredit model performance degradation
Likelihood4
Impact4
Trend↑ (worsening)
MitigationModel retraining scheduled Q1; self-employed cutoff tightened by 15 points

Suggested Leading Indicator: 30-day delinquency rate by employment type on current origination cohort — gives 60–90 days of early warning before NCO crystallizes.