Case study 8: AI tool vendor pitch
SYNTHETIC Fictional vendor materials for training only.
Scenario
A vendor demo claims their humanitarian AI platform will:
- “Predict displacement 30 days ahead with 95% accuracy”
- “Auto-approve cash assistance for low-risk cases”
- “Train on your data to improve models for the sector”
- “Comply with all privacy laws globally”
Your organization has no AI procurement framework yet.
Your task (12 minutes)
- List claims that need evidence.
- Which claims raise red flags for humanitarian use?
- What questions would you ask the vendor?
- Recommend: adopt now, pilot with limits, or decline. Who decides?
Model answer
Evidence needed: Accuracy study methodology, false positive/negative rates by subgroup, legal compliance map (which laws?), data retention and training use, human override for cash decisions.
Red flags:
- Auto-approval for assistance (accountability, appeal rights)
- Sector-wide training on your data without consent
- Vague “all laws” compliance claim
Vendor questions:
- Where is data processed? Sub-processors?
- Can we audit model decisions affecting individuals?
- What is the incident response process?
- Who is liable for wrongful denial of aid?
Recommendation: Amber at best: limited pilot only after governance framework, DPIA, and no auto-approval of benefits. Red for immediate adoption.
Decision owner: Leadership with privacy, protection, and program input.
Common errors
- Buying on accuracy marketing without subgroup testing
- Piloting on live beneficiary data without safeguards
- No exit plan if vendor changes terms or pricing
Principle taught
Vendor due diligence: Humanitarian contexts require evidence, not slogans. Governance precedes procurement.