Specify a high-latency / response-time alert (e.g., p95 or p99 response time above an SLA threshold) rather than alerting on average latency, and justify the percentile choice in terms of client-experienced performance.