For "expert" systems, the analogy that I think works best is the car. It is a complex, life-threatening device that takes hours of learning to be operated safely, but that you can eventually drive for hours pretty much without thinking about it.
In my opinion, the situation depends on whether you're talking about controls or displays. A critical aspect of it is also the frequency of expected use.
For controls
A large part of what makes driving a car possible is that it relies on your muscle memory: you don't need to read the label on the switch that turns on the headlights, you just know where it is and turn it on.
The same kind of efficiency can be reached on a computer by allowing keyboard shortcuts, which in my opinion are way faster for expert UIs. Nearly all of the highly proficient experts that I've seen use the same software throughout the day rely heavily on keyboard shortcuts to do their job faster.
Icons can work to increase information density too. What's also important is that they are in reliable locations as well.
Some examples of this that you may be familiar with is Photoshop, Adobe Premiere, or any CAD software — labels are typically scarce in the UIs, which (the field being pretty competitive commercially) show that it's an advantage for expert user bases. But again past a certain point, expert users of these applications tend to switch to keyboard shortcuts.
As an example, below is the Photoshop toolbar — with labels, it would take up about 5x as much screen real estate.
All of this comes at the cost of a longer training curve simply because when you tell someone "use the tool" they may not understand right away what you are referring to. But if your tool needs significant amounts of learning anyway, it doesn't matter as much.
For displays
In the case of displaying information, we can go back to the car analogy and see that it works at least for very simple cases: people know what a red light means without it being labeled. However, you will see a lot more things labeled on the road signs than in your car.
Europe is an exception to this — being a linguistic patchwork, people have decided there to standardize on road signals, so that a "stop sign" or a "no parking" sign are supposed to look the same across the continent (to the endless puzzlement of American tourists). But for computer UX where the UI can be translated to the viewer, this makes no sense.
Back to UX, one thing to consider is that accessibility requirements will typically force you to find a way to write things out for users with visual impairments, or color blindness. Personally I would be more careful about use of icons in displays than in controls.
The role of frequency
The more your users will rely on a control or display, the more familiar it gets and therefore the more appropriate an icon treatment is. Conversely, the less your users will rely on a control, the less familiar it gets and the more important labels are.
As an example, car warning symbols are notoriously bad, i.e. many people can't tell what they mean, even when the symbol represents important, potentially life-saving advice. This is because these icons are hardly ever seen, so people's memory of what they mean (assuming they ever learned it) decays with time.
Icons and other visual treatments are appropriate for things that people need to do with high frequency. "Red light means stop" is a good example: most drivers encounter it every time they drive. It is risky to use icons for things that happen infrequently, because users may just overlook them completely out of a lack of understanding, and it may lead to bad mistakes.
Illustration: photoshop toolbar
