How Failures Propagate
Thesis
AI failures do not stay contained. They propagate through operational dependencies, vendor relationships, and governance gaps.
Analysis
A model hallucination is a model problem. A hallucination that triggers an operational action is an institutional problem. The architecture connecting the model to the institution determines whether a failure stays small or becomes consequential.
Failures propagate because systems are connected. An AI system that writes to a database that feeds a customer-facing application creates a propagation path. A failure in the AI system becomes a failure in the database becomes a failure in the customer experience.
Most organizations have not mapped these propagation paths. They know where their AI systems live. They do not always know where the failures would go.
Framework
Failure Propagation Map AI FAILURE ↓ DEPENDENT SYSTEM ↓ OPERATIONAL PROCESS ↓ CUSTOMER IMPACT ↓ REGULATORY EXPOSURE ↓ REPUTATIONAL DAMAGE At each propagation point: • Detection mechanism • Containment barrier • Escalation trigger • Recovery procedure The goal is not to prevent failure. The goal is to prevent propagation.
Board-Level Questions
If our primary AI system failed today, what would the cascade look like?
Where would the failure first be detected?
How far could it propagate before containment?
Have we mapped the propagation paths for every AI system?
Operating Implication
Map the failure propagation path for every AI system. Identify where a single failure can cascade into operational, financial, or regulatory consequences. Design containment at each propagation point.
Related Intelligence