We determine criteria for modelling plan-based “but” in task-oriented dialogue (TOD), following work by Lagerwerf [5] and focusing on cases in which it signals denial of expectation (DofE) and concession, to which end we propose a novel treatment of concession in TOD. We present initial considerations for an algorithm to address plan-based “but” in an Information State (IS) model of dialogue that predicts which interpretations (DofE and/or concessive) to generate. We motivate this work by showing how it updates beliefs in the PTT [6] model of dialogue and can be used to facilitate recognition of planning mismatches and more generally, discourse understanding and natural language generation (NLG).