How does your company handle data science and AI portfolio responsibility / P&L impact and ROI

I've been in data science for about a decade and I'm in the process of forming some views of how we best organise data science and related disciplines in companies.

The standard organisational model that has emerged over the past few years seems to be a "Hub and Spoke" model where you have the central hub providing feature stores, MLOps standards and capabilities, line management, technical community, and so on, and the spokes which is where the data scientists (et al.) are embedded in the business units. The primary alternatives to this are fully centralised or decentralised organisational models, which I think are comparatively rare these days.

One thing that I am less clear about is how portfolio responsibility tends to play out. By that I mean who's ultimately responsible for the P&L impact of data science work and whether those resources get used in an intelligent way?

There are two primary ways to set this up, as far as I can gather:

1. Portfolio responsibility in the business units. In this model, data science is essentially treated as a utility/capability that is delivered by the DS/ML/AI department and the business units are ultimately responsible for whether the data scientists are delivering an appropriate ROI. Portfolio development/management in one business unit can be completely different to that in another.
2. Portfolio responsibility in the data science dept. The Hub or some other body ultimately decides where the data science resources are deployed, ensuring maximum ROI across business areas. Data science products/services are treated more like ventures or bets with uncertain payoffs and portfolio management is handled as a dedicated function.

And then I guess there are many half-way houses in between.

So my question is how does this work in your company?