Andrew Ng wrote insightful article in HBR a while ago about what an AI can and cannot do. He highlights that if there’s a mental task that a human being can do in few seconds, you can probably automate it.

In my opinion, these types of problems have a following structure,

1. Single person decisions: these decisions are often replacing decision making of single human being.
2. Scale factor: You need AI to help you in tons of such decisions in order to make meanigful impact.
3. Correlation enough: you do not need to prove causation, insights based on correlation is enough.

However, real work often gets done with decisions of the following structures,

1. Collaborative: imagine a CEO, finance team and sales team deciding whether to offer discount to close a sale.
2. Scale factor: Few dozens of such decisions made correctly and you are home. Company doesn’t need to make million such decisions correctly to succeed.
3. Causation important: if I offer discount, will the sale close and will the client stick around longer?

In current Avatar, most AI (to be honest, machine learning algorithms) are shining at the first bucket. Real impact of AI will come when it can start assisting human beings with second set of problems.