Although AI shows growing promise for mental-health care, current evaluation approaches remain fragmented and poorly aligned with clinical practice, social context, and first-hand user experience. We introduce an interdisciplinary framework combining clinical soundness, social context, and equity. An analysis of 135 recent computational-linguistics publications identifies recurring limitations, including reliance on generic metrics, limited participation by mental-health professionals, and insufficient attention to safety and equity. We propose a taxonomy of assessment-, intervention-, and information-synthesis-oriented AI support, each with distinct risks and evaluation requirements, and demonstrate the framework through case studies.