Definition
Degree to which the intelligent router's volume-shifting decisions have degraded the approval rates of the processors it favors: detecting the feedback loop where routing optimization becomes self-defeating.
The Measurement Gap
The Measurement Gap
No metric tells you whether intelligent routing decisions are degrading the approval rates of the processors they favor. Intelligent payment routing works by sending each transaction to the acquirer or processor with the highest predicted approval rate for that transaction's characteristics. The optimization is sound in isolation, but it creates a feedback loop that no standard metric detects.
When the router identifies Processor A as having the highest approval rate for a given merchant category and card type, it shifts volume to Processor A. As volume increases, Processor A encounters capacity constraints, issuer-side velocity triggers, and BIN-level concentration effects that degrade its approval rate. The router observes the degradation and begins shifting volume to Processor B, which then experiences the same cycle. The system oscillates, never reaching equilibrium, with each processor's approval rate degraded by the volume the router sent to it.
Standard routing metrics track approval rate per processor and volume distribution, but they report these as independent observations rather than causally connected outcomes. Authorization approval rate looks normal in aggregate because the router is always chasing the best available option. Per-processor approval rates decline, but the decline is attributed to processor performance rather than to the volume shift that caused it. The feedback loop (routing optimization degrading the signal it optimizes against) is invisible to any metric that treats routing decisions and processor outcomes as independent variables.
Decision Supported
Is the intelligent routing system creating self-defeating feedback loops that degrade the processors it favors, and by how much?
Formula
Measurement Prerequisite
RFDI requires a baseline approval rate for each processor at a volume level not influenced by AI routing. Two methods: (1) historical pre-routing data from before intelligent routing was deployed, or (2) a randomized holdout sample where 1-5% of transactions are routed randomly regardless of the model's recommendation. Without one of these baselines, RFDI can only be tracked as a directional trend, not as an absolute degradation measure. The holdout is the stronger method because it controls for temporal changes in processor behavior.
Plain Language
For each processor, compare its current approval rate under AI-directed volume to its baseline approval rate at normal volume levels. Weight by volume share and aggregate across all processors to get the total self-inflicted degradation.
Notation
Per-processor: RFDI_p = (ApprovalRate_p_baseline − ApprovalRate_p_current) / ApprovalRate_p_baseline × 100
Aggregate: RFDI_agg = Σ(w_p × RFDI_p)
Variables
| Variable | Definition | Unit |
|---|---|---|
| ApprovalRate_p_baseline | Approval rate for processor p at baseline volume level (pre-routing historical data or randomized holdout) | % |
| ApprovalRate_p_current | Approval rate for processor p at current AI-directed volume | % |
| w_p | Volume share of processor p in the current period | Ratio (0-1) |
| RFDI_p | Per-processor feedback degradation | % |
Proposed formula, pending validation pilots.