This depends on the radiation, the phase types (carbides, nitrides, oxides, intermetallic copiunds), the grain size of the matrix. Investigations of Fe with Cu you can reach depth up to 5µm (it is an exponential function) but Cu is commonly not the first choice because it generates Fe fluorescence radiation so that an enormous amount of signal is simply Fe radiation (high background). Alternatively, for stress measurements Cr is preferred. Then the information depth is in the scale of only 1 µm. Perfect for residual strain measurements but too low for your problem of phase analysis. Therefore, one uses Co radiation. Comparable to Fe and no fluorescence radiation. W radiation goes even deeper but because of the small wavelength Bragg angles become very small and therefore the overlapping of peaks. Good in case of residual austenite determination since the peaks of austenite and ferrite/martensite are still separated. In case of other phases it is perhaps not the best choice since it becomes very inaccurate. Also because of the small Bragg angles the effective penetration depth is not that big as the linear absorption coefficient will tell you. Moreover, steel has perhaps a too big grain size, i.e. you will get no statistics and the pattern is perhaps only useful for a phase identification. You also need to know the composition (solid solution) of each phase as accurate as possible since the composition determines the scaling factor  and in case of ordered structures also the relative intensities of all reflections (structure factor).

This depends on the radiation, the phase types (carbides, nitrides, oxides, intermetallic copiunds), the grain size of the matrix. Investigations of Fe with Cu you can reach depth up to 5µm (it is an exponential function) but Cu is commonly not the first choice because it generates Fe fluorescence radiation so that an enormous amount of signal is simply Fe radiation (high background). Alternatively, for stress measurements Cr is preferred. Then the information depth is in the scale of only 1 µm. Perfect for residual strain measurements but too low for your problem of phase analysis. Therefore, one uses Co radiation. Comparable to Fe and no fluorescence radiation. W radiation goes even deeper but because of the small wavelength Bragg angles become very small and therefore the overlapping of peaks. Good in case of residual austenite determination since the peaks of austenite and ferrite/martensite are still separated. In case of other phases it is perhaps not the best choice since it becomes very inaccurate. Also because of the small Bragg angles the effective penetration depth is not that big as the linear absorption coefficient will tell you. Moreover, steel has perhaps a too big grain size, i.e. you will get no statistics and the pattern is perhaps only useful for a phase identification. You also need to know the composition (solid solution) of each phase as accurate as possible since the composition determines the scaling factor  and in case of ordered structures also the relative intensities of all reflections (structure factor).

The deeper the X-ray enters the sample, the more it interacts with the sample’s atoms. The penetration depth depends on the wavelength (i.e., the type of detected atom), the intensity of the X-rays any many other factors that Dr. Nolze mentioned previously.

In general, heavy elements in the matrix decrease the saturation depth considerably. In overall, I think with good sample preparation and using the conventional X-ray  method penetration depth in steel could be 100-250µm.

the penetration depth has nothing to do with sample preparation (as long as you do not change the materials density). It has the same interaction depth for amorphous and for crystalline materials.

I=I0 exp(-µd)   (µ...attenuation coefficient (linear absorption coefficient/density), d...depth)

i.e. you are interested in I/I0 and you need to say which ratio you are asking for. µ is listed or can be calculated for different radiations. 

Please note that the penetration depth is commonly the value where the intensity is attenuated to a certain ratio. Of course, radiation reaches also higher depth (see equation) but their contribution to the signal is respectively low. This also points out, that any gradient in your material will overestimate the layer close to the surface. 

The equation also shows tha each depth estimation should also give the attenuation level assumed. 100-250µm would assume a very short wavelength, or a very low attenuation level. Practically, it is for my taste too big.  Typical values are related to I/I0=0.5. But please consider, that the diffracted radiation has to leave the sample again, i.e., d is twice that long... Also consider that d is in that case not deff since there the Bragg angle is still not yet considered. This means that the effective depth deff always depends on the theta (diffraction angle).