This is slightly an edge case but I think it can add some value to the other answers.
If you add more details to the graph (say an annotation or a line) you'll soon discover that it is relevant when you call legend on the axis: if you call it at the bottom of the script it will capture different handles for the legend elements, messing everything.
For instance the following script:
df = pd.DataFrame({'A':26, 'B':20}, index=['N'])
ax = df.plot(kind='bar')
ax.hlines(23, -.5,.5, linestyles='dashed')
ax.annotate('average',(-0.4,23.5))
ax.legend(["AAA", "BBB"]); #quickfix: move this at the third line
Will give you this figure, which is wrong:
While this a toy example which can be easily fixed by changing the order of the commands, sometimes you'll need to modify the legend after several operations and hence the next method will give you more flexibility. Here for instance I've also changed the fontsize and position of the legend:
df = pd.DataFrame({'A':26, 'B':20}, index=['N'])
ax = df.plot(kind='bar')
ax.hlines(23, -.5,.5, linestyles='dashed')
ax.annotate('average',(-0.4,23.5))
ax.legend(["AAA", "BBB"]);
# do potentially more stuff here
h,l = ax.get_legend_handles_labels()
ax.legend(h[:2],["AAA", "BBB"], loc=3, fontsize=12)
This is what you'll get:
