As Hurkyl said "the Central limit theorem".
Whenever we apply mathematics to a "real life" problem, we have to establish a "mathematical model". Often, since "real life" problems involve measurement which is never exact (like mathematics is), there are several different models we could use and we have to decide which to use. To solve statistics problems, we have to decide on an underlying probability distribution. Sometimes there are clear reasons for using a binomial distribution or Poisson's distribution.
However, for the vast majority of statistics problems, the normal distribution will give the best fit precisely because of the "Central Limit Theorem"- one of very few mathematical theorems that say "this is a good model for certain types of applications".
The Central Limit Theorem says: "If you take a n samples from a given probability distribution (any distribution as long as it has finite mean,x\overline, and standard deviation, \sigma), the sum of those samples will be approximately normally distributed with mean n x\overline and standard deviation \sqrt{n}\sigma while the mean of the samples will be approximately normally distributed with mean x\overline and standard deviation \frac{\sigma}{\sqrt{n}}.
The larger n is, the better the approximation.