Earlier quoted context omitted.
Funny. This is the one to which I had the most negative response. Zipcodes are terrible for inferring anything geographically interesting. A single zipcode is just a collection of USPS postal routes; nothing more, nothing less. They don't have to be contained within a single city, or even a single state, nor do they have to be contiguous, and they can overlap with one another. New ones are also added all the time, an…
ZIPs are almost always fine at a state granularity, and pretty good for cities. If the form is street/city/state/zip, you can autofill State correctly from ZIP 99.9% of the time.
Do you have data supporting the 99.9%? I suspect that's high; there are buildings within five blocks of where I'm sitting that have a zipcode that crosses a state line, and I don't think it's the only state-crossing zip in my area (Washington, DC, for what it's worth). In any event, even if it's accurate, it doesn't take that much traffic for a .1% error rate to start to add up to non-trivial number of users, especially because that .1% is going to have a noticeably-jarring experience, as opposed to the alternative (collecting all fields) which everyone is already completely accustomed to.