Once again last night, I had someone ask me what percentage of customer service calls they should survey.
The answers (as always) is "it depends."
But for most applications, the real answer is "it doesn't matter."
That's not to say that the number of surveys doesn't matter, but that the percentage of calls surveyed doesn't matter.
Surveying 1,600 customers will get you a 2% margin of error whether that's 1,600 out of 20,000 customers, or 1,600 out of two million.
The main thing that matters is making sure that the sampling method is as unbiased as possible. You could perform thousands of surveys and collect data on 50% of all calls, but if the sampling method is flawed (like an end-of-call survey), the results will be inaccurate. Worse, in fact, than surveying just a few hundred people with a proper technique.