The exception detection system is the intelligence layer that transforms Empire AI from a monitoring dashboard into a decision-support platform. Rather than requiring executives to hunt through dashboards for problems, Empire AI's agents hunt for them automatically and deliver structured alerts with context, impact analysis, and recommended actions.
The system operates on three detection modes:
Threshold-Based Detection. The simplest but most critical layer. Each KPI in the platform has configurable thresholds — yellow (warning) and red (critical). When revenue dips below 90% of monthly target, a yellow alert fires. When it dips below 75%, red. But Empire AI goes beyond simple threshold crossing: the agents analyse velocity (how fast is the metric changing?) and trajectory (where will it be in two weeks at current velocity?). A metric that's 95% of target but declining at 3% per week is more concerning than one that's 85% but stable.
Pattern-Based Detection. Machine learning models trained on historical data identify anomalous patterns that threshold rules would miss. A sudden drop in sales conversion rate on Tuesday afternoons. An unusual spike in customer churn from a specific segment. A procurement category where costs are creeping up 2% month-over-month — invisible in any single month but clear across a quarter. The pattern detection models run continuously and surface anomalies ranked by statistical significance and business impact.
Predictive Detection. The most sophisticated layer uses forecasting models to predict problems before they materialize. The cash-flow forecasting agent models expected inflows and outflows over the next 90 days and alerts when the projected balance dips below safety thresholds. The churn prediction agent analyses customer behaviour signals and identifies accounts at risk of leaving — weeks before they actually churn.
Each alert includes a structured payload: the affected metric, the detection method, the severity level, the estimated financial impact, related metrics for context, a recommended action, and the relevant approval workflow if action requires authorization. Alerts are delivered through the command centre interface, with optional push notifications for critical items.
The blueprint includes detailed specifications for 15 detection agents spanning financial, operational, workforce, risk, and strategic domains — each with configurable sensitivity, escalation paths, and false-positive reduction strategies.