There are three good answers to this question, and together they have more or less answered what I wanted to know. I find it hard to choose one of them as best, but nevertheless I think this question should have an accepted answer to move it from the *unanswered list*. Hence a CW-answer summarizing the (in my eyes) most important points made.

---

* Laurent Moret-Bailly points out that $\hat{X}$ must be normal.
* Sasha then says that besides that, it can get as bad as you want. Take a normal subvariety $\hat{X} \subset \mathbb{A}^{N}$. By Noether's lemma we get a finite map $\hat{X} \to \mathbb{A}^{n} = Y$. A resolution of singularities $X \to \hat{X}$ has connected fibres. The composition $X \to Y$ is generically finite.
* Karl Schwede remarks that the pair $(\hat{X}, -\mathrm{Ram})$ is log-Gorenstein (where $\mathrm{Ram}$ is the ramification divisor). He also states the slogan *“if $\hat{X}$ has really bad singularities at some points, then $\mathrm{Ram}$ also has really bad singularities at those points too. Another way to say this is if the ramification divisor has mild singularities, then $\hat{X}$ does too.”*