What are you measuring, what is your actual signal? Its bandwidth?
What's wrong with aliasing higher frequencies -- is there any consequence? Is it just noise, is it correlated with the signal (e.g. harmonics), is it spurious and undesirable?
How sharp is the filter?
Do you need phase-flatness (good step response, typical for time-domain applications) or frequency-sharpness (typical for RF and audio applications)?
How much quantization noise / accuracy / resolution do you need? Distortion?
In the below section, I've assumed you know what aliasing and frequency mixing effects are; if not, please read up on them. They are instrumental to the precise application of ADCs.
After you've answered all of these questions, you can design the filter's cutoff and slope, relative to the sample rate. The signal chain shall have an analog (antialias) filter, the ADC, then a digital filter. The goal is to have "100%" cutoff, at the frequency somewhat above Fs/2, which corresponds to the first aliased image of the cutoff frequency somewhat below Fs/2, that the digital signal path finally has. By "100%" cutoff, I mean, any external signal or noise is reduced to below the noise floor, which is in turn defined by analog noise or ADC quantization noise, whichever is greater (actually, the sum of both).
For example, a time-domain application, like an oscilloscope, might use an 80MSps ADC, with 20MHz (-3dB) analog bandwidth, and (1-2-2-1 weighted) FIR digital filter for similar digital bandwidth, with a Bessel filter type giving -60dB (for an 8-bit ADC) at 80 - 20 = 60MHz. Going from -3dB at 20MHz to -60dB at 60MHz requires about -120dB/dec or a 6th order filter.
Note that, for a radio application, the sample rate and analog filtering may be chosen to coincide with an RF passband, using aliasing to your advantage. In that case, the band "skirts" need to be sharp enough to prevent further aliasing. This is also known as Equivalent Time Sampling (ETS), and can be used in multiple (i.e., harmonic) bands simultaneously, which is how ETS oscilloscopes work: they're actually sampling many frequency bands, simultaneously and coherently, which correspond to the harmonics of the signal being measured. Hence, the (periodic) waveform can be reconstructed, without violating Nyquist, because we've made an assumption about the signal (its periodicity).
Tim