Let's talk math: In some very formal, reductionist sense, what we're doing when we do math is to start with a set of axioms, operate on them using a fairly small set of logical rules, and arrive at theorems that seem interesting.
If we do this very carefully -- much more carefully than a human would ever normally do, showing 100% of our work -- it's trivial to check such a proof automatically. You just look at each step, and verify that that step follows from the axioms + logical rules you're using + previous steps. This checking is clearly (handwave) polynomial: at worst, at each step N, you have to examine every axiom, every logical rule, and each prior step to make sure that step N+1 is legal.
So proof checking is in P.
Let's say I have a set of axioms, and a logic, and I want to know whether a theorem T is true given that system. One strategy I could use is to decide to only consider proofs of length less than N, and generate every legal theorem that I can reach in fewer than N logical steps. When I'm done, I check my list of proven theorems and see if any of them match T. Since I can always my answer in polynomial time, I know this is in NP.
The try-everything-possible algorithm has the downside of being a bit exponential (since we expect N to be very (very!) large for any interesting theorem) and so also a bit useless.
But what if a P=NP proof can provide guidance on how we should combine our axioms in order to reach theorem T? We suddenly have not a proof-checker, but a proof-generator! Anything that we can state in a formal language, we can simply ask the prover to prove -- it will either give a proof, or say that a proof of size < N doesn't exist. (That wouldn't be a disproof, and it's a Goedel/halting problem kind of argument that we may never really know how large N needs to be.)
Finding new proofs is a large part of what's considered creative in mathematics. It's hard to over-state how math might change with a tool like that available. The field has seen any number of revolutions before -- e.g., the effort of hundreds who worked on the algebraic roots of polynomials pretty much only survives in homework exercises nowadays -- but automating proofs would be Big.
(Finding interesting conjectures is valuable, too, but I doubt Erdos would be nearly as famous if he'd only posed his questions without his resume of theorems behind them.)