The growing demand for real-time and application-specific insights in modern networks has elevated the significance of In-band Network Telemetry (INT). INT, enabled by programmable data planes and a domain-specific programming language, P4 provides fine-grained, hop-by-hop insights into network traffic. While INT enables fine-grained packet monitoring across programmable switches, its uniform deployment across all traffic flows can lead to substantial overhead and inefficient resource consumption. To address this challenge, we propose a novel AI-driven telemetry framework that combines flow intelligence with INT activation. This paper proposes an AI-assisted INT (AI-INT) framework that leverages the programmable data plane using P4 and ONOS, in combination with a Multi-Layer Perceptron (MLP) model for real-time traffic classification. Upon flow detection, the classifier predicts the traffic class, and ONOS installs appropriate telemetry and forwarding rules on the BMv2 switch, accounting for the current network state. This integrated system enables efficient, dynamic, and application-aware telemetry that minimizes overhead while maintaining observability. We present the architecture, implementation details, and evaluation results that demonstrate significant gains in overhead reduction and resource utilization.