The estimator must be the maximum likelihood estimator.

There must exist a sufficient statistic $T$ for $\theta$.

The parameter space must be finite.

The sampling distribution must be normal.