A critical challenge during volcanic emergencies is responding to rapid changes in eruptive behaviour. Actionable advice, essential in times of rising uncertainty, demands the rapid synthesis and communication of multiple datasets with prognoses. The 2020–2021 eruption of La Soufrière volcano exemplifies these challenges: a series of explosions from 9–22 April 2021 was preceded by three months of effusive activity, which commenced with a remarkably low level of detected unrest. Here we show how the development of an evolving conceptual model, and the expression of uncertainties via both elicitation and scenarios associated with this model, were key to anticipating this transition. This not only required input from multiple monitoring datasets but contextualisation via state-of-the-art hazard assessments, and evidence-based knowledge of critical decision-making timescales and community needs. In addition, we share strategies employed as a consequence of constraints on recognising and responding to eruptive transitions in a resource-constrained setting, which may guide similarly challenged volcano observatories worldwide.