I don't know that there is any concrete information available about how the Larian engine renders scenes, and , of course, it's still being developed. Even with relatively dynamic lighting, it may possible to pre-calculate and store the effects of individual lights as an optimization.
I'm still not sure of the utility of RT hardware, to be honest. It isn't absolutely necessary, as you can do RT on the CPU, or using shader code on normal GPU stream processors; just not as efficiently.
There is always a question mark in a GPU if you add special purpose silicon, since you do so at the expense of general-purpose silicon. How do you balance the number of RT units against unified stream processors when you don't know what functions a game will want most? How do you design your shaders and asset pipelines efficiently when you have a different balance of RT to SP on different GPUs?
This has happened already with specialized tessellation hardware, among others, which allows you to smooth curved surfaces or add additional detail not in the original asset geometry. But it's often not necessary, requires totally different asset geometry format, and can end up being a bottleneck on the overall frame rate. As a consequence, it is often ignored making it wasted silicon.
That said, there are many frequently used geometric and optical calculations that might benefit from RT. Ray-object intersections for geometric calculations and ray-cast primary shadows are obvious candidates, in addition to reflection and refraction effects. Occlusion and radiosity for global illumination are also possibilities.
In the end I think it will be down to developer experimentation. I can see console titles benefiting the most, simply because of the high number of identical hardware units, and the need to use everything they have available.
NVidia's RT, on the other hand, seems to render a low-res image and uses their deep learning silicon on the latest GPUs to predict the correct lighting when upscaled to full output. I'm not sure what that would do for non-render calculations.
Wait and see, I guess.