Actually, trying to prove uniqueness of left inverses leads to dramatic failure! It's an interesting exercise that if $a$ is a left unit that is not a right unit, then there are infinitely many $b$'s such that $ab=1$.
So the only situations that can arise for a left unit $a$ is that it is a two-sided unit (with a unique inverse that works on either side) or else it is truly one-sided, in which case it has infinitely many distinct left inverses.
Of course, a left-unit need not be a unit. The standard example is to use a vector space of countable dimension and basis $\{b_1,b_2,\ldots\}$ and define linear transformations $f$ and $g$ by $f(b_i)=b_{i+1}$ for all $i$ and $g(b_i)=b_{i-1}$ for all $i\geq 0$, and $g(b_0)=0$. It's clear that $gf=1$ but $fg\neq 1$.
Actually, the choice of $g(b_0)$ can be chosen completely at random from the vectors in the vector space, and each such choice generates a distinct new map $g$. All of them, however, satisfy $gf=1$.