Mobile-Money Billing Reconciliation at an African MNO — From 14 Days to T+0 on 280M Daily Transactions

MindMap deployed Bank Reconciliation (Br) as the multi-rail probabilistic-matching engine, Anomaly Detector (Ad) as the exception-pattern-detection layer, Multi-Agent Orchestrator (Mo) for the cross-system orchestration, and Financial Close (Fc) for the daily-reconciliation cycle workflow.

Phase one was the matching-engine rebuild. The previous reconciliation engine was rule-based with sequential rule application; the new engine uses a probabilistic-matching model that evaluates all candidate matches across all match dimensions simultaneously and emits the most-likely match with confidence scoring. The model is trained on the operator's historical-match-outcome data with the labelled match outcomes as supervision signal.

Phase two was the cross-rail integration. The mobile-money transaction lifecycle spans multiple rails — the in-network mobile-money platform, the partner-bank settlement rails, the agent-network commission rails, the third-party-merchant settlement rails — and the previous reconciliation engine had handled each rail in isolation. The new engine handles the cross-rail transaction-lifecycle holistically, matching across rails where the transaction's lifecycle spans rails.

Phase three was the exception-investigation workflow. For the transactions that the matching engine cannot auto-match with high confidence, the investigation queue is prioritised by customer-impact (subscribers with open disputes), commission-impact (agents waiting for settlement) and revenue-impact (high-value transactions affecting revenue recognition). The investigator's workspace surfaces the candidate matches the model considered, the reasons each was rejected, and the recommended next investigation step.

Phase four was the daily-reconciliation cycle integration. Financial Close drives the daily cycle workflow — the timing of the reconciliation runs, the cross-team handoffs between operations, finance and the partner-relationship teams, and the close-cycle reporting that the operator's finance team uses for revenue-recognition.

The platform runs on the operator's private cloud inside its primary data centre, with high-availability failover to the secondary site. The transaction data — substantial PII and financial-data exposure — stays inside the operator's perimeter.

The matching engine processes approximately 280 million daily transactions on a streaming basis, with end-to-end matching latency typically under 5 minutes from transaction occurrence to match-confirmed status. The engine's throughput design has 4x peak-volume headroom to accommodate continued mobile-money growth.

Bank Reconciliation's matching model is a gradient-boosted-tree ensemble with features spanning the standard match dimensions (reference identifiers, amount, counterparty, timestamp) plus mobile-money-specific features (agent-network correlation, third-party-merchant correlation, partner-bank-settlement correlation). The cross-rail transaction-lifecycle tracking is implemented as a transaction-state graph that follows each transaction from initiation through to final settlement across all involved rails.

Anomaly Detector's exception-pattern layer identifies systemic issues in the transaction flow — recurring mismatch patterns from specific agents or merchants suggesting integration issues, anomalous volume patterns suggesting partner-side data-feed issues, fraud patterns suggesting coordinated attack on the mobile-money rails. The pattern-level insights drive the partner-relationship and engineering-improvement work.

Financial Close orchestrates the daily reconciliation cycle workflow with the operator's finance team, including the cycle-completion reporting that the finance team uses for revenue recognition. The reconciliation outputs flow into the operator's ERP via the standard integration.

The full reconciliation audit trail — every transaction, every match decision, every exception investigation, every settlement event — is persisted with the local-jurisdiction-required retention.