We examine the problem of determining the fluid flow at the core surface given a model of the temporal variation of the magnetic field, seeking to fit the field over a finite time interval. The flow is assumed steady over the time interval, and we adopt the frozen-flux approximation. We produce a range of solutions of varying complexity and subject to different constraints, estimate the errors on the solutions, and assess their consistency with the underlying assumptions. We find both direct evidence for ageostrophic flow in the form of flow across the geographical equator, and indirect evidence in the form of poorer fits to the field when the flow is contained to be geostrophic. On account of this evidence we do not adopt the geostrophic approximation. However, we find that unconstrained flows are poorly determined, and consider the various alternatives for producing better determined solutions. We choose to restrict our attention to determining the large-scale toroidal part of the flow, and find a simple pattern of flow which is nearly symmetric about the equator. From a sequence of maps of the flow over 5-year time intervals back to 1915 we are still able to identify many of the basic aspects of this simple pattern, but changes in the flow pattern are discernible. Although our models of the flow explain a large proportion of the signal (typically 95 per cent of the variance), the fits which we obtain are not as good as might be expected. Failure of the underlying assumptions, together with our decision to map only the large-scale toroidal part of the flow, explains the poorer than expected fit.