[I'm with Efficient Frontier, the largest ppc keyword mgmt firm in the world, so note that my viewpoints may be biased.]
Two years ago there were many more advertisers building their own keyword mgmt tools, but that has pretty much died down now that there are so many tools available. As I see it, there are many tools & services to choose from, depending on what your needs are:
Omniture Search Center
There is one HUGE industry preventing almost every bid mgmt vendor from helping its clients on AdWords (which is typicall 65% of their U.S. search efforts and 75-80% in Europe) - they can't optimize on Google AdWords.
Most bid management tools are rules-based, which means they apply one or more rules to a keyword, or to a bucket of keywords. These rules can be around position, CPC, or in some cases ROI goals. The problem is that because AdWords is an opaque environment, you cannot know what max CPC yields what bid position except to the extent you fully leverage your own historical data for the campaign. IMO, there's not a single rules-based bid mgmt tool that does, and I include AtlasOnePoint, Did-It, KeywordMax, PPCPro, Omniture, WebSideStory and the SE's own tools in that bucket. The reason for this is that none of these tools has any intelligence built into them; modelling is a foreign concept to them, yet that it *precisely* what is needed to automate the process of leveraging historical and actual cost & revenue data.
Add to that the fact that Google's Traffic Estimator is useless, and you have a situation where the rules-based tools that worked fairly good on Yahoo are now merely high-priced tracking packages whose value on AdWords is non-existent. That would be OK were not Google 65-85% of the world's major PPC markets.
Lastly, with Yahoo's Panama project [hopefully] close to production, the only remaining transparent PPC system will go opaque, meaning that only with a gigantic amount of human intervention will the rules-based systems add any value to marketers. With most good ppc advertisers running thousands to tens of thousands of keywords, I just don't see humans able to model data for more than a few dozen head-end keywords.