Cauchy and Fourier had a famous dispute about whether Fourier's use of trigonometric series to solve the heat equation was valid. To a large extent, the basic rigorous notions of analysis were formalized in order to resolve this dispute. Along the way, Cauchy published a famous "wrong" theorem. (It wasn't really wrong; it just used a different definition than that which became standard.)
Here's another, more applied, example. I remember going to an joint meetings talk years ago in which the speaker described a situation in which engineers had employed widely-used software to numerically compute solutions to PDE's. But they had neglected to prove that, at the level of precision they were using, their methods would converge to the correct solution. The result was disastrous: a brand new 180 million dollar oil rig collapsed so thunderously that it registered a 3 on the Richter scale.
_EDIT:_ Thanks to Kcd here's the link to more details.