The best implementations of a noncoherent receiver will effectively correlate the signal against sin and cos (the local oscillators don't need to be matched _exactly_ in frequency, let alone phase -- small errors are OK --) and then effectively do an energy measurement on the output. Cheaper versions use a real bandpass filter (could even be an off-the-shelf component) with some loss of performance. Also, with real bandpass filters, if FSK is being contemplated, the tones cannot be as close in frequency as can be done with IQ correlators etc. On the other hand, the cheap bandpass filter version can be cranked out very quickly (e.g. if the boss wants the design by lunchtime) whereas if you want better performance, the design is more complicated and will take more time and effort. And I don't mean the mythical man-month of IBM fame where 750 man-hours means 250 guys trying to get it done by lunchtime!