Yes, even AMD said Mantle isn't for everyone and it's quite right. You don't need roads that F1 cars can race on when all you have is light traffic.
The issue may be due tu multiple bottlenecks. In ArmA you can test yourself with the provided 2D/3D editor without any AI. Performance tanks once you increase object complexity and draw distance and it's more severe in cities and towns. I'd guess 32bit limitations, probably some API here and there, plus poor occlusion and multitrheading, all take part in that. And yes, you have poor performance while the CPU and GPU don't do much. Since it's the same engine, the same happens on Day Z, ArmA 2, Take on Helicopters, etc.
Not sure how's in Total War, but you can play test quite easy different scenarios in ArmA and perhaps even analyse the draw calls if you have the
know how.
Beyond the API matter, let's not forget about GPU physics and AI, already demonstrated by
nVIDIA and
AMD. None of those really took off, although they hold a far greater potential even compared to DX12 or Vulkan. If an old HD4850 can run 3000 AIs with great complexity and advanced graphics at HD resolution, what could modern day GPUs do with physics and AI plus a great API?
The technology is there and doable, too bad Crysis didn't sell that well. Gaming would be a whole lot different than the same copy -> paste games year after year.