This answer probably won't be coming from the perspective that you want, since it'll use even more stable homotopy theory than Adams did.  But I think it's pretty clear in its own way.

First let me make a small correction to Neil's comment.  Actually, from the J-homomorphism construction you only get a map of spectra from the *connective cover* of $\Sigma^{-1}ko$ to $gl_1(S^0)$.  So you have to kill that bottom Z in ko.  But actually the better thing to do is to "find" the same Z below $gl_1(S^0)$ instead, by changing the target to $pic(S^0)$, the classifying spectrum for invertible spectra.  So the J-homomorphism should be thought of as a map of spectra $ko\rightarrow pic(S^0)$.  In these terms it has an easy intuitive description too: $ko$ classifies vector bundles, $pic(S^0)$ classifies stable spheres, and the J-homomorphism sends a vector bundle to its one-point compactification.  Now you see why it's the multiplicative structure on $S^0$ that's relevant, because this operation sends direct sum of vector bundles to smash product of spheres.

This isn't just me being pedantic, by the way.  It's good to understand exactly what structure the J-homomorphism carries, and it's also very handy to have that extra Z around.  One measure of this is the following:

Claim:  Let $J: ko \rightarrow pic(S^0)$ be *any* map of spectra which sends the unit class in $\pi_0 ko$ to the class of the 1-sphere in $\pi_0 pic(S^0)$.  Then the image of J on homotopy groups satisfies the ``lower bound'' estimates established by Adams for the J-homomorphism.

Unless I'm misreading your question it's these lower bound estimates that you're interested in.  So what I'm saying is that for those purposes, all you need to know about the J-homomorphism, besides the natural structure it carries, is that it ``sends 1 to 1'' on that bottom Z we attached (or rather, never bothered to forget). 

The way we'll prove the above claim is to give a more modern version of Adams' e-invariant argument.  This will be based on Rezk's study of logarithmic operations coming from the Bousfield-Kuhn functor.

The motivation is that, as Neil says, the spectrum $pic(S^0)$ (or $gl_1(S^0)$) is kind of mysterious.  That's too bad for us, since apparently we need to understand it to prove the above claim.  However, the Bousfield-Kuhn functor tells you that for any prime p, the "part of a spectrum which mod p K-theory understands", meaning the K/p-localization of the spectrum, only depends on the n^{th} space of the spectrum, for any n you like.  This is some amazing consequence of periodicity.  The upshot is that the K/p-localization doesn't care that $pic(S^0)$ has this exotic infinite loop structure, and you get a natural map
$$pic(S^0)\rightarrow L_{K/p}S^{1}$$
which exhibits the latter as the K/p-localization of the former.

Now, a priori this map is somewhat mysterious on the level of homotopy groups.  But actually Rezk made a fantastic study of it in his paper on logarithmic cohomology operations (which the above is essentially an instance of.)  One consequence of Rezk's work is a calculation of the effect of the above map on $\pi_0$.  It is as follows.  First, recall that $\pi_{-1} L_{K/p}S$ identifies with $Hom(Z_p^\times,Z_p)$ (noncanonically, this is just $Z_p$, the $p$-adic integers.)  Then, the above map $log:pic(S^0)\rightarrow L_{K/p}S^{1}$ sends the class of the $1$-sphere to the homomorphism $Z_p^\times\rightarrow Z_p$ given by
$$x\mapsto \frac{1}{2p}log(x^{p-1}).$$
(I might have a sign wrong, but that doesn't matter for our purposes.) Here $log$ stands for the $p$-adic logarithm.  It's an interesting thing that the above expression makes sense and is primitive.  That is, the set of all $(p-1)^{st}$ powers of $p$-adic units is exactly the domain of convergence of $log$, and, also, $2p$ is the GCD of all the values of $log$.

We deduce from this that the composition $log\circ J:ko\rightarrow L_{K/p}S^{1}$ sends the unit class in $\pi_0ko$ to that same homomorphism.  Now the point is that $K/p$-local homotopy is computationally friendly.  Also, $ko$ is nearly $K/p$-local: its $K/p$-localiztion is the $p$-adic $KO$, via the connective cover map $ko\rightarrow KO$.  Thus it's not hard to calculate that a homotopy class of maps $ko\rightarrow L_{K/p}S^{1}$ is completely determined by its effect on $\pi_0$.  So actually the above minimal information tells us exactly what $log\circ J$ is.  Then if we look on higher homotopy groups, we see that the image of $log\circ J$ has size given by Adams' upper bound.  This implies the claim.

By the way, I think the cleanest way to perform the above calculations is to fix the generator $g = exp(\frac{2p}{p-1})$ for the group $Z_p^\times/\pm$ of p-adic Adams operations acting on p-adic $KO$.  Then you get to write down the fiber sequence
$$L_{K/p}S \rightarrow KO\overset{1-\psi^g}{\rightarrow} KO,$$
and the result is that $log\circ J$ identifies with $ko\rightarrow KO\rightarrow L_{K/p}S^{1}$, where the last map is the boundary map in the above sequence.  (Again my signs might be off, sorry for that.)