Educators are increasingly confronted with technology-driven learning scenarios. Even before the push from the current pandemic, digital learning apps became an integrated didactic tool. Advanced computing can thereby support the digital content creation for educational courses offered on mobile platforms. Computed media content such as natural voices or generated images of real looking but non-existent persons has become easy to access and feasible to use. Particularly in exploratory learning approaches, natural environments and lifelike characters could be utilised to create compelling instructional scenarios or reality-oriented training assignments. However, there is still a limited understanding of the effects such generated lifelike persons used as Pedagogical Agents (PAs) may have on cognition and behaviour. This study presents the findings of an international field test on the Google Play Store that investigated the effects of an artificially generated instructor's facial expression on risk-taking in a decision-making task. In the field study, an established measure of risk propensity, the Balloon Analogue Risk Task (BART), was extended to include instructions by an artificially generated lifelike persona with different facial expressions. The resulting research game for mobile phones was internationally distributed on the app store for a two-month between-subjects field experiment. The participants were instructed either by a smiling female pedagogical agent, the same agent with a neutral expression or no agent while deciding to risk further pumping up a balloon for more profit or safely collect balloons for realising a current, more modest profit. The results (n = 379) indicate that instructions presented by a smiling female agent reduce risk propensity in decision-making compared to instructions presented with a neutral facial expression. Instruction design considerations and experiences from the distribution of a research game on Google Play are summarised in the concluding implications.