Distributed Intelligence in the Artificial Intelligence of Things: A Comprehensive Review of Architectures, Applications, and Challenges

Artificial Intelligence of Things (AIoT) applications increasingly exceed the limits of centralized cloud processing because they require low latency, privacy preservation, scalability, and operational resilience. This review synthesizes distributed intelligence across the edge–fog–cloud continuum through a structured integrative methodology comprising multi-stage literature search, two-stage filtering, and thematic synthesis of more than 100 sources. The analysis covers four representative domains—industrial IoT, smart cities, connected healthcare, and smart agriculture—to identify recurring architectural patterns and shared deployment challenges. The review organizes these challenges around power and computational constraints, data management, security and privacy, interoperability, and model lifecycle management. Building on this synthesis, the paper formalizes an Edge–Fog–Cloud distributed intelligence model and develops a workload-placement taxonomy based on latency, privacy, power, and model complexity. Comparative analysis shows that on-device TinyML is best suited to ultra-low-latency and privacy-sensitive inference, edge and fog layers provide an effective compromise for localized near-real-time intelligence, and cloud infrastructures remain essential for large-scale analytics and model training. Across domains, the evidence supports hybrid multi-layer architectures as the most robust strategy for advanced AIoT deployments. The review also identifies key future directions, including human-in-the-loop AIoT, multimodal sensor fusion, energy-harvesting devices, federated learning, and the Tactile Internet.