Good question. I think the "data" aspect is largely a misconception. Any proprietary data might have some value in model fine tuning that would come with product setup, but almost no businesses actually have the big diverse datasets that support a nontrivial AI model. Look at the hyped up current AI language and image generation models. They're not based on proprietary data. Nor were the past generation.

So I'd argue that while proprietary data has a role in what is effectively the setup of a product, it's not the defining factor. And if anyone thinks that ML is going to come in and do something amazing on the strength of their own data, they're probably setting themselves up for disappointment. In fact, I think it's that misconception that has led companies to (imo foolishly) invest so much in in-house AI