In this paper we investigate error estimates for the approximate solution of operator equations Af = u, where u needs not to be a function on the same domain as f. We use the well-established theory of generalized interpolation, also known as optimal recovery in reproducing kernel Hilbert spaces, to generate an approximation to f from finitely many samples u(x1),…, u(xN). To derive error estimates for this approximation process we will show sampling inequalities on fairly general Riemannian manifolds.