It very much depends on the auxlang. Most of them do try to make things easy, but auxlang creators tend to be a little more linguistically naïve than other artificial language creators, so they end up using themselves as the benchmark. Esperanto is a notorious example of this, where you have initial consonant clusters like 'sc' /st͡s/ and internal consonant clusters like 'tĉ' /tt͡ʃ/. Zamenhof essentially took a simplified version of the Polish consonant inventory and applied it Esperanto.
There are plenty of studies of consonant and vowel frequencies in human languages. We know, for instance, that [p] [t] [k] are found in virtually all languages, that most languages have at least one liquid, one nasal consonant (which can be realised as [n], [m] or [ŋ] depending the position of the consonant that follows it is). Most languages also have [ʔ], but you never notice it as most only allow it initially and exclude it when spelling: if you were to choose any word in English beginning with a vowel in isolation, you will almost certainly start it with a glottal stop.
As far as vowels go, the classical five-vowel system is very prevalent: some approximation of it is used in lots of languages, and there are very few languages that use simpler systems, though Modern Standard Arabic, with its three-vowel system ([a], [i], [u]) is a widely-spoken counterexample. If you wanted to include more consonant clusters, you could add schwa as a epithetic vowel: in our hypothetical language, the word _plum_ could be phonemically /plum/, but phonetically [pəlum] for speakers who have difficulty with that initial consonant cluster.
Still, knowing what sounds people can pronounce is only half the battle: a good IAL would also have well designed _phonotactics_. That's a bit of a rabbit hole though!