Case 1 — Wildfire evacuation SMS draft

A communications officer wants to paste beneficiary phone numbers into a public AI chatbot to personalise evacuation SMS messages.

Answer: Red — prohibited input (R1). Use an approved SMS platform; no PII in AI prompts.

Case 2 — Translation with harmful omission

AI translation from Dari to English omits a sentence about threats from an armed group. The officer submits without re-reading the original.

Answer: Amber — protection review (A2); bilingual human review required.

Case 3 — Donor update with invented statistics

AI drafts a donor report with "14,200 households" and "37% reduction" — numbers not verified against MEAL data.

Answer: Amber before send; Red if published with fabricated statistics. Use the source-check ledger.

Case 4 — Do not use AI

A caseworker wants AI to decide cash-assistance eligibility from case notes.

Answer: Red — high-stakes eligibility decision (R3). Use structured human assessment with appeal rights.