Consider what you are asking for: you desire an average AC model, but the device in question has 20-some pins on it. Which pins are you expecting a model through? Under what mode/condition (driver, PWM current limiter) shall it be linearized or averaged? How many combinations of pins should they model for reasonable completeness across all end users that desire this type of model? And how many people will actually make use of such models, when a reasonably complete* transient model is provided instead? (That is, even if such a model is reasonable and meaningful, it might still not be worth crafting because too few people will make use of it to justify the expense.)
And, what are you really asking for? -- A PWM cycle-averaged model of a gate/motor driver is just a buffer, give or take some phase shift due to propagation delay, and some gain factor due to modulator gain (if applicable). Besides the current limiting function, these components are trivial in such a model; were you expecting more?
Also be careful about what kind of "AC" models you are asking for. SPICE does not require models specific to this, because it performs linearization on the system automatically, using a small-signal steady-state assumption. Note that a switching circuit has approximately zero gain under this method (i.e., consider the gain of a logic gate where its input voltage is below the logic threshold), so you will be disappointed with the results from this analysis, if you use device models. (So, you are correct, with respect to your assumption that transient models will not work as-is for AC/average purposes.)
AC Analysis is only applicable to passive and analog circuits, where bias (operating point) is easily calculated from the circuit as given, and the small-signal frequency response is desired.
In contrast, as you move from simplified average models towards a practical PCB design, you must consider transient conditions within a switching cycle (peak current flows and voltage drops due to stray inductance from the layout and other component strays), cycle-to-cycle variations (e.g. ripple current, for purposes of dimensioning supply components), and quasi-steady-state conditions (e.g. heat dissipation of devices, for component selection and thermal design purposes). Few things which your average or AC model
The only duty you have at this later [implementation] stage, with respect to the earlier (average) model, is to show that they still exhibit the same averaged dynamics. Which I think you will find is a much simpler task than all the other considerations the transient models bring, and ultimately the whole design.
I'm also assuming this is in academic context, or perhaps a high-reliability design context, given the emphasis on modeling. In contrast, much practical engineering is done on assumptions and testing. If you are in fact doing the latter, I would suggest not worrying about it, and concentrate on solving implementation issues (with or without simulation).
Finally, keep in mind what you stand to learn / develop from a given model. If I understand your meaning correctly, then about all you're going to find from your average model is just the compensation components for the respective control loops. That's, at most, 9 PID parameters total across your three loops (or equivalent R and C component values if analog, or perhaps more parameters if more complex controllers are chosen). Whereas the complete design might involve, what, 50 or 100 components, or more? So the average model can only helping you with say 10% of the overall design -- a small part of the whole. And, you can assume those values will have to be chosen somehow, and just leave them as placeholders -- deferring choice of values until final modeling, or even "tune" them in testing.