I'd have to agree with you. I have found Pair programming useful at certain points when the code being developed was particularly complicated (although a bit more time spent on design and review may in some cases have been a better solution)...I recall a post on slashdot, that pair programming makes a lot of sense when its "natural"...as you describe--but forcing full-time pair programming can be quite expensive--there's always the 80% of boilerplate/braindead code that may not really benefit as much from 2 people looking at it...whereas the more risky/complicated code might benefit from it. I'd propose that a well managed project should be able to identify those risky areas and use pair programming in those situations--I've yet to try this myself though :)