One interesting case for MLEs arises for the MLE of the parameters of the Laplace distribution.
The MLE for the location parameter ($\theta$, say) is the median and for the scale ($\tau$, say) is the mean deviation from the median, but I the minimal sufficient statistic is the complete set of order statistics; the minimal sufficient statistic doesn't reduce to only some function of the MLE for $(\theta,\tau)$.
However, the thing stated in the question would still seem to hold (since the MLE of $(\theta,\tau)$ is some function of the MSS), but it seems you could write some functions of the MSS that those two statistics don't capture; say the difference-in-the-extreme-gaps $X_{(n)}-X_{(n-1)}-(X_{(2)}-X_{(1)})$, could not be recovered from them.