Henrik Olsen spins a good yarn about developing personas for a Scandinavian tin can manufacturer. As he quickly points out, he didn't exactly do it the right way. Instead of interviewing actual customers of the site, he interviewed people who spend most of their time dealing with customers of the site. This is an OK way to do it, if time and budget are limited (though experience has taught me that it takes pretty much the same amount of time to talk to customer service representatives, as it does to find and interview actual customers.) One might more accurately call the artifact of this effort a provisional persona, rather than a full-bodied persona. (See "Getting from Research to Personas".)

Now, let me just come out and say that personas are a tool, like a hammer. They're a really good tool for understanding how people behave, but they're not the perfect tool. (Nothing is.) Personas can't tell you, for instance, whether to color combination you've selected is appealing (try a focus group.) Personas can't tell you whether the check-out process on an eCommerce site is architected correctly (for that you're going to need usability testing.) To me, personas are really good at helping you choose features, the depth of their functionality, and the specific design directions that your site needs to take.

Provisional personas will work too, but you're trading accuracy for time/budget savings. That might be worthwhile in the short run (especially if you've got to get something launched by a specific, looming deadling) but it might wind up costing you a lot more in the long run. When forced to settle for provisional personas, I usually insist that usability testing occur at multiple points during the design phase.

There is one major risk of relying on provisional personas: too often, the result is a set of personas the reflect the kind of customers to whom you want to be selling your products. Or worse, they reflect only a certain type of customer (maybe just the ones who are the most vocal.) In both instances, you're working with your blinders on, perhaps missing the forest for the trees. In either case, you're getting a filtered view of your customers, which is never as reliable and informative as the real thing. Factor this into your development process, and highlight the risk appropriately.