AI and IoT Synergies in Enhancing Business Operations: A Simulation-Based Evaluation of Predictive Maintenance, Edge–Cloud Placement and Business Value

Abstract

The Internet of Things (IoT) gives businesses continuous, granular data about their physical operations; artificial intelligence (AI) turns those data into forecasts and decisions. Their combination is widely credited with lower downtime and cost, but most published figures are vendor claims or case studies whose data cannot be reproduced. This article develops a reference architecture and a synergy framework for AI–IoT integration across five operational domains, and tests the central claim, that AI and IoT together deliver more than either alone, in a transparent simulation of predictive maintenance. A synthetic fleet of 96 machines over three years has two latent degradation modes, load- and temperature-dependent wear, shocks and machine-specific failure thresholds; five sensor streams carry machine-specific baselines, load confounding, noise, drift and dropout. Gradient-boosted trees on lagged rolling features forecast failure within 72 hours with a test AUC of 0.91, against 0.83 for the best single-sensor rule and 0.75 for a model without sensor data. Six maintenance policies were simulated on identical degradation paths in five replications. The AI+IoT policy cut failures from 1.93 to 0.20 per machine-year and cost per machine-year by 56% against reactive maintenance, 37% against time-based maintenance, 18 to 21% against tuned condition-based rules and 39% against an AI-only policy, while still costing 63% more than a perfect-foresight oracle. The additive interaction between AI and IoT was negative in every replication because the separate contributions already approached the attainable ceiling; the combination captured 77% of the attainable saving, the share expected from two complementary, independent contributions. A five-year value model gives a net present value that is large against reactive maintenance, modest against an existing condition-based programme and negative for small or reliable fleets, with break-even failure intensities for each baseline; a placement model shows cloud inference cheaper and edge inference better on latency, bandwidth and autonomy. All results describe the simulated fleet; the contributions are a reproducible testbed, a decision-focused evaluation protocol, an implementation roadmap and a critical reading of circulating percentage claims.