Small and medium-sized enterprises (SMEs) run point-of-sale, inventory, finance, supplier and customer systems that do not reason jointly, so decisions in one function ignore constraints in another. We present BizTwin AI, an implemented multi-agent control tower that maintains a business digital twin, monitors it for anomalies, and coordinates nine domain agents through an Adaptive Multi-Agent Decision Intelligence Engine (AMDIE). AMDIE weights agents by a learned reliability score passed through a temperature-controlled softmax, trains that score with a quantile (pinball) loss at an owner-chosen service level, and fuses recommendations by a constrained linear program whose optimum we show to be a weighted median. We evaluate replenishment decisions in a simulated retailer across three scenarios (adequate budget, tight budget, volatile demand and supply) with 10 held-out instances each, against operations-research baselines, single agents and unweighted ensembles. AMDIE outperforms the agent mean by 3.0 percentage points (pp) of profit in the baseline and volatile scenarios (Holm-corrected p < 0.02) and learns within four weeks to place 70-80% of its weight on the most reliable agent. It does not outperform a correctly specified newsvendor policy or that best agent alone, and ablation shows that two components (median-type LP fusion and an outlier penalty) bias it toward under-ordering. We report these results, and the corrections they imply, as guidance for multi-agent decision support in SMEs.

Multi-agent systems, decision fusion, business digital twin, inventory control, explainable AI, small and medium enterprises