One option is to start building from scratch, but often it might be better to prove out the values first. That way you have momentum before making a huge investment for very high-value resources. Coming up with some smaller proof of concept opportunities to actually see how machine learning can deliver for particular use cases is probably a more successful pattern compared to building an entire organization from the ground up.
Something that is also a common mistake is hiring a bunch of machine learning experts who are very expensive before getting the data infrastructure and the data accessibility in their organizations in order. That’s something that is often a big challenge and then what ends up happening is a low or delayed return on investment, or a hacked together system which leads to bad engineering for years.