It is possible to include observed variables as part of the structural model. For example, to find the impact of latent variables on observed variable.
With older programs, such as LISREL you couldn't do this - you had to create a latent with a single indicator, and no error variance. Now you just do it, for e.g. in Mplus, if F is latent and x is observed, you'd write
x on f;
In Lavaan, you'd write:
x ~ F
Also, when you have a latent variable, with indicators, the loading is equivalent to the regression. So (in Mplus):
F by x;
is exactly the same as my first statement.