AWS AutoScaling CPUUtilization metric not accurate?

I do heavy computation on incoming data traffic, similar to web server requests but not exactly. The computation uses mainly cpu. Memory or disk read/write is hardly used at all. I deployed this application to an autoscaling group. I also have some customised measurements on the performance of the system, beside the default AWS AutoScaling CPUUtilization metrics.

The strange thing I found out is, sometimes the default AWS AutoScaling CPUUtilization metrics can be as high as 95%, but my customised measurements show that the system works just fine, as also justified by my visual user check.

I'm quite confident on my customised measurements and I believe that it really shows the true performance. But should I consider the high CPUUtilization in this case as abnormal, and it's just simply not accurate?