We should be eternally grateful to work in a field where every single last thing can be modeled, emulated or solved, without having to touch a single tool.
There are no guesses here. There can only be a lack of knowledge.
Anyone who insists that it is a required design step, to perform physical testing, and that luck is involved: they are admitting their ignorance.
Now, it often is the case that, for a variety of reasons, an engineer is unable to attain full knowledge of a system, and therefore must make assumptions, guesses, and perform experiments. One must understand that this is not a strictly necessary route, and these steps are more as exceptions that we make for various reasons (insufficient time or budget to study or model the system; insufficient working knowledge; insufficient time/budget to pursue the education to obtain that working knowledge; etc.).
It's a practical problem, not a theoretical one.
A very good engineer might go her entire career without having undertaken such a project -- that is, one with knowledge so comprehensive as to be practically complete -- but this is only statistical evidence, and not proof!
Conversely, a less skilled engineer might go his whole career assuming that such processes are necessary, and never taking up the education (or making the realization) that there are other ways to do it.
Mind, it's not that one or the other approach is necessarily superior. Perfectly merchantable products can be made either way. It's perhaps interesting to consider the ways in which the two approaches might fail. For example, the low-level method might be fragile under variation in parts value (something of a liability for long-term products, as parts age in use, or as substitutes parts are put into production over the years), whereas the high level method might get all the basics right, but have a structural quirk that makes the design unexpectedly difficult to evolve over time, as the customer's requirements or production needs change (e.g., using just the perfect part, and then the requirement changes just outside its capability; or it's from Maxim and they discontinue it, amirite?). Either way, if the absolute best results are demanded, one should allow for multiple cycles of refinement.
So, in regards to EMC, we can -- at least in principle -- model everything perfectly, make just a few component changes, and have high confidence in passing tests the first time. We don't need to rely on luck in the physical test, but we might have no reasonable alternative due to the amount of effort required to create such a simulation (namely, ca. $6-figures for the EM simulator, plus some months setting up the model).
(Note that the argument applies recursively, so that a more talented engineer might change just a few values in the simulation and be done, whereas the other may make hundreds of relatively undirected changes. Simulations are just virtual testing procedures, after all. Or in the meta, so that ones' level of knowledge might be approached in a comprehensive way, or through scattered experience.)
(Note also that this is not an endorsement of "CAD jockey" design. If one can reach a level of practical experience from their armchair, by all means do, but do not leave it unchecked against reality. Many will need lots of hours in the lab to reach a functional level; most prople, I suppose, will never reach it at all. (That is, a lot of people don't care for engineering at all, or tried and didn't turn out to be any good at it.))
Tl;dr: disagree.
Tim