When shall we use Expectation Maximization (EM) instead of Maximum Likelihood Estimation (MLE)?

I saw in many articles that EM is an algorithm to do MLE, and we usually use it when a direct MLE is not possible.

Can someone tell me what is the meaning of "direct MLE is not possible" (and what are the reasons for not being possible)? Preferably with an example.

As an example, let us assume you have a Normal distribution that is censored above some value $$\theta$$, so that for a given sample you know how many observations are greater than $$\theta$$ but not what their values are. (You do know the values of the observations that are less than $$\theta$$.) If you want the maximum likelihood estimators of the parameters $$\mu$$ and $$\sigma^2$$ of the distribution, you'll have to resort to some iterative procedure, because there is no closed-form solution (so the "direct MLE is not possible") to the problem of finding the estimates. (In the uncensored case, you calculate the MLEs using the sample mean and sample variance with denominator equal to the sample size.) EM is one such iterative algorithm, and a widely-applicable one at that.