A lot of AI features in Figma so far seem to be generative, which IMO is mostly fruitless for the core auidence as current models are not capable of new problem solving, only generating variations of the combined knowledge of everything that has been observed before, without understanding the context of decisions made to implement design choices as well. 

Where I can see see value add is in other aras. QA for example - perhaps pointing a dev URL at the tool to see how front-end have developed the designs and comparing to what was in the original figma (visually, not comapring the code generated by Figma, which is often not all that helpful in established systems) - could help catch mistakes and also documwent intentional changes in the design to development process.