One of the more striking examples of the power of randomization is its use in the estimation of the volume of a polytope. Given an oracle that declares (for a given point) whether it's inside or outside the polytope, a randomized algorithm can get an arbitrarily good approximation to the volume, even though the problem is #P-Complete. A deterministic algorithm, on the other hand, will fail miserably.
A related question that I was curious about is the surface area estimation problem. Given a polytope defined by the intersection of hyperplanes (we can assume that the input promises to yield a bounded polytope for now), can we estimate its surface area efficiently ? Now, one can use a volume oracle on each hyperplane to determine the area of the face it contributes to (if at all), but I was wondering if there was a more direct method (especially since the above approach is far more wasteful than necessary, especially in small dimensions).
There's even a practical reason to do fast surface area estimation, at least in 3D: this is used as a heuristic for building efficient space partitionings for ray tracing.