As of July 2026, MMG Dugald River Mine (DRM) will have operated its closed-loop digital twin for flotation optimisation for 18 months in production. This paper reviews the outcomes of one of the first operational deployments of a fully autonomous, AI-assisted process control (AIPC) systems for flotation in the Australian minerals processing sector. The authors outline both the measured metallurgical benefits and the key technical learnings associated with operating deep neural networkbased technology for optimising Net Smelter Return (NSR) in a live production environment. To quantify the system’s value, DRM conducted a structured, randomised on/off trial over a representative operating window, enabling a direct comparison of twin-driven and operator-driven control outcomes. The trial concluded a positive impact of +0.53 per cent Zn recovery improvement (𝑝𝑝 < 0.001) by the twin, on a high baseline rougher-scavenger recovery performance exceeding 97 per cent. The trial also concluded no statistical difference in rougher-scavenger concentrate or feed grades, while the twin did achieve a significant reduction in reagent consumption through an improved multivariate control action. Time utilised in full autonomous control over the first 12 months of twin operation was approximately 75 per cent. Several post-commissioning enhancements supported these outcomes which the authors have detailed in the paper. These can be summarised as: improvements to flexible adjustment of twin set point optimisation bounds directly from supervisory control and data acquisition (SCADA), improved generalisation of the twin’s internal rougher and scavenger control models; major enhancements to system infrastructure allow automated retraining of rougher and scavenger models for efficient maintenance and redeployments; and improvements in the usage of model errors in real-time to aid in short-interval self-correction of the NSR objective function calculations. Overall, the advancements achieved by the DRM team have demonstrated a new novel form of advanced process automation which contrasts significantly with traditional fuzzy logic and linear matrix model predictive control systems. Most notably, the self-learning and continuous learning aspects of the underlying AIPC systems utilised in the rougher scavenger deployment will allow greater benefits to plant metallurgists to directly model, maintain and deploy neural network-based control systems more flexibly from the plant as part of their day-to-day role. The results and findings outlined in this paper will inform both the incremental scale of the DRM rougher scavenger system to the downstream cleaner block, as well as the future application of AIPC technologies across other MMG flow sheets.