Cascade Amplification in Administrative AI Pipelines: When Feedback Loops Turn Additive Errors into System Failures

Administrative artificial intelligence tools in primary care — scheduling, documentation, coding, billing, referral management — are routinely classified as lower-risk than clinical decision-support systems, on the basis of evaluating each tool in isolation. We simulate a five-node administrative AI pipeline on a Synthea primary care cohort (n = 1,000 patients, 100 Monte Carlo iterations, seed 42), calibrated with parameters from published empirical evidence including a conservative +2% wRVU inflation from independent longitudinal data, an 8 percentage point E/M upshift from national claims analysis, and a 28% triage error rate from external validation. We introduce the Cascade Amplification Factor (CAF), a metric that distinguishes additive (CAF = 1.0) from superlinear (CAF > 1.0) error accumulation. Two adjacent AI nodes (documentation + coding) yield CAF = 0.503 [0.458, 0.540], indicating partial error cancellation. Five AI nodes without feedback yield CAF = 1.009 [0.925, 1.099], indicating near-additive accumulation. Five AI nodes with a referral-to-scheduling feedback loop yield CAF = 2.245 [2.032, 2.486], exceeding the pre-specified clinical relevance threshold (CAF = 1.2) by a factor of 1.87. Care-related harm dominates the four-dimension harm taxonomy under full pipeline deployment (share ≈ 61%, driven by intervention deficits). We frame the transition through Perrow's Normal Accidents Theory: the feedback loop converts a linearly interactive system into a complexly interactive one. The regulatory implication is that governing individual tools is necessary but insufficient; the feedback architecture of the deployed pipeline is the decisive governance target. Govern the loop, not the tool.