Method starts with processing, by a processor, audio signal to generate audio caller utterance. Processor generates an agent action ranking score associated with the audio caller utterance and determines whether the agent action ranking score is below a minimum threshold. In response to determining that the agent action ranking score is below the minimum threshold, processor generates a transcribed caller utterance using a speech-to-text processor and generates an identified task based on the transcribed caller utterance. Using the transcribed caller utterance and a task-specific agent coaching neural network associated with the identified task, processor generates an ideal response. Processor generates a feedback result and causes the feedback result to be displayed on a display device of the agent client device. Other embodiments are disclosed herein.