I meant workable "AGI" or "Artificial General Intelligence", which is different than the types of AI we have today. It could also be termed "strong AI", and has the ability to learn, on its own, any task a human can.
At present, what we have is "weak AI", which is generally an AI that is trained on and capable of performing one specific task (For example, image classification, driving, generating text, etc.). As a 'strong AI' is likely decades out (if not further, outside of any of our natural lifetimes), my idea is to combine multiple "weak AIs" into a whole, with a 'supervisor' AI that is tasked with figuring out the right response to a situation. It would still be unable to learn a new skill on its own and would require human aid to do so. However I believe it would, with enough single-task AIs, be able to be interacted with in a convincing enough manner for you to forget that fact.
This approach's most difficult part is still at the top, as a machine learning program to figure out the right thing to do still straddles the line between "weak" an d "strong" AI, but it has the benefits of some of the work already being done, being compartmentalized enough to be divided into sub-tasks, and also means we don't have to solve the harder problem of making software that can learn an entirely new skill.