Not all models do need it. And if all models were built to the same high standards, there would be less need for it still, but they are not, and we live in a world where stuff often has to be done faster than might be ideal, which means working with 'minimum standard' meshes. For example, CAD programs (from which a lot of client models originate) seem to operate in a way that is largely oblivious of quality topology (optimised as they are for design-helpful speedy iteration), or even basic poly rules - they will do literally anything that produces a viable (ie renderable) surface, and that is usually massively unhelpful for anything people might try and do later with that mesh, in a program which is different to the one it was originally made in; if they try and subdivide it, it fails - if they try and add bevels, it fails, and if they try and use half the selection and modelling tools, they will be battling uphill constantly, and fighting topology all the way because few of the tools can work predictably where edge flow is terrible or non-existent. Sometimes you get meshes that were inexpertly modelled, and that might be for 'valid' reasons, like if something was built without animation in mind, but then someone decided it had to move ! Indeed it is most usually character animation where loop placement suddenly becomes crucially important.
Likewise, if you want to use Subdivision, there are even more strict polygon rules, and not following those will produce duff, imperfect or inconsistent (across apps) SDS results. Usually this applies most to heavily curved or undulating surfaces, or to surfaces where there is a mixture of rounded curves and proximal ultra-sharp corners and planarity. Interestingly, in this last of circumstances, there are no automatic tools in existence (not even quad re-mesher or instant meshes) that will do as good a job as the experienced human modeller doing it by hand, which is why people like me are still in a job despite them ! And yet we too welcome these tools because they make our job a lot faster and eradicate a lot of the more tedious aspects of the poly-pushing.
Or, you might start a model in a very painterly / artistic way, with sculpting let's say, which allows very fast shape dev and iteration, but when you finally get your finished form the polys won't be in a particularly optimal order, having been mercilessly shifted about in the sculpt and you can almost always improve that by retopologising the base mesh after it. Also if you are baking new sculpt information, and the base mesh you started with doesn't adequately describe the final form (like if you used sculpting to pull out some horns from the head of a demon for example), so you are getting loads of stretched and uneven polys everywhere, then again, retopo is the only thing that can save the day, and return your mesh to usefulness further down the pipeline. There are also texture and UV-based reasons why topology should be decent. The short story there is that very unevenly distributed polys (even 100% quads) tend to produce sh*tty UVs prone to excessive texture distortion. But again, I could find you examples of circumstances where even that doesn't matter.
And, in the interests of fairness, I should point out the corollary, that there are occasional circumstances where quality of topology genuinely doesn't matter at all. For fully planar surfaces, for still shots or where there is no deformation animation, and no SDS involved, then your surfaces can look like a proper birds' nest of technical errors and still 'work' in that they give a flawless rendered result and have taken very little time and effort to make, which can be helpful or even necessary with tight client deadlines to be met.
Each job we undertake requires a different level of adherance to various poly modelling rules. The confusion and difficulty for beginners arises not from the rules themselves, which are surprisingly few and simple, but from knowing which ones apply in which circumstances, and why.