Model Quantization for CPUs — INT8 Viable for Serving?

CPU serving with INT8 quantization promises cheap inference, but accuracy drops vary wildly by task. Which workloads survived INT8 in your serving stack, and which had to stay FP16?