The article is partially correct.
AI has to be continually trained for different markets, property types, appraisal types, and so on. "Training" requires preparing enormous numbers of data-input + data-output pairs, e.g., pictures of rooms with labels (Kind-of-room, condition-of-room, quality-of-room, range-type, counter-type, oven-type, ...). The Neural Net uses these thousands of instances to learn how to create useful property description labels for some new, unknown property that comes along. It takes people to do this work of labeling photos (a relatively menial job). But also, AI neural network models (which are basically matrices of weights) are not always right and can, in fact, be very poor. A trained person has to monitor their use and be able to fix problems or otherwise make improvements.
What we will see is fast development and correspondingly more complex software systems. Indiduals who can succeed will be those fascinated by complexity, with good memories and a great ability to understand large complex systems.