Sometimes I do sales, and in that case I use the word "accuracy". Other times I do engineering, and then I use the word, "error". While the two words are opposites, they both describe exactly the same thing.
For em analysis, error can range from 5% down to below 0.01%. In a very very few cases, em analysis can provide answers that are exact to within numerical precision, but this requires specialized code. For example, this is possible for Zo of lossless infinitely thin stripline.
If you know the correct answer in advance, you can always "tune" and em analysis to give a very accurate answer (for example, by precisely adjusting the aspect ratio of the subsection size), but this is useful only for sales to engineers who are not very smart (our customers, I work for Sonnet, are smart, so we do not do this).
More typically, accuracy is determined by doing a convergence analysis. Keep cutting the cell size in half (on a simple circuit) and see how it converges. If the difference between each successive analysis is cut by about half, you have a good idea what the error is. This is the typical (with occasional exceptions) case for Sonnet. A nice smooth convergence is always best.
If the answer oscillates about some final correct answer as you keep cutting the cell size in half, it is much harder to determine the error.
If no convergence analysis is done, i.e., if only one em analysis is done of a given circuit, then "accuracy" becomes a philosophical/political "yes it is"/"no it isn't" argument. In this case, it is not an engineering discussion. I think this description fits much of the information provided in the above posts.
For RFIC, where cost of a mistake is very high, it is best to analyze the circuit with at least two different tools from two different vendors. Now, if there are significant differences, you MUST understand why those differences are there before fabrication.
If I may make a promotional sales statement for our own tool, which works well in conjuction with ADS, MWO, Eagleware, and Cadence, (whichever one you think is best) we have some new features that will have significant impact in RFIC design. The response from our half dozen beta testers is strong. Please feel free to contact us if interested in using our "accurate" em software for RFIC.