The antiproton flux measurements from AMS-02 offer valuable information about the nature of dark matter, but their interpretation is complicated by large uncertainties in the modeling of cosmic ray propagation. In this work we present a novel framework to efficiently marginalise over propagation uncertainties in order to obtain robust AMS-02 likelihoods for arbitrary dark matter models. The three central ingredients of this framework are: the neural emulator , which provides highly flexible predictions of the antiproton flux; the likelihood calculator , which performs the marginalisation, taking into account the effects of solar modulation and correlations in AMS-02 data; and the global fitting framework , which allows for the combination of the resulting likelihood with a wide range of dark matter observables. We illustrate our approach by providing updated constraints on the annihilation cross section of WIMP dark matter into bottom quarks and by performing a state-of-the-art global fit of the scalar singlet dark matter model, including also recent results from direct detection and the LHC.