AI/ML accelerators in these things are mainly for inference purposes - training most NN's takes considerable time and power (watts and TOPS), whereas inference using a fully trained NN can be extremely efficient even on these relatively small ASIC's.
As for actual use cases, ML can be used for various vision/imaging and audio based tasks (feature/facial recognition, image stabilisation, speech synthesis, voice recognition, handwriting recognition?) - basically anything with identifiable patterns will benefit from it, if the NN can be compressed efficiently enough to run on a small footprint.
One of the Daala/AV1 developers (Xiphmont) recently developed a new audio codec that uses ML for ultra low bandwidth (1.6 kbps), presumably for narrow band only (voice). Link here.