The era of Agentic IoT: Moving from monitoring to autonomous operations
The Internet of Things (IoT) started as a simple idea of connecting assets and physical devices so that their operations could be tracked, measured and optimized. Initially the IoT solutions were focused on areas like RFID based tags for asset tracking, sensors for temperature / vibration / energy consumption measurement and dashboard creation / predictive maintenance recommendation, remote monitoring to replace the manual inspection and provide real-time data. Industry focus was primarily in manufacturing, utilities, logistics and infrastructure, thereby initiating the data focused operations.
Over the years IoT has dramatically evolved with improved and cheaper connectivity, larger data storage and processing capabilities at cloud platforms, it still largely remains an observation capability coupled with recommendation (e.g. predictive maintenance) with decision making remaining in human hands. This has allowed the growth of IoT deployments in industries which would have been understandably hesitant in handing over the decision-making autonomy to the computers. This hybrid model has effectively led to the large-scale growth of the IoT deployments across various industrial sectors and contributed to significant savings and improvements in operational efficiency. Nasscom has been working with large number of enterprises to support their digital transformation journey.
With the arrival of AI, the pure dashboard creation using IoT data has started to fade in favour of AI addition for analysing the sensor data, pattern detection, generation of alerts and failure prediction. While AI provides the guidance, the final decision is still in the hands of humans. Now with the era of Agentic AI upon us, the initial steps are being taken to explore autonomous decision-making being handed to the IoT+AI combination. These connected systems go beyond the basic monitoring and analytics and move towards the autonomous operations. Advances in edge AI capability allow low latency decision making capabilities close to the physical system, thereby creating potential opportunities for autonomous decision making in tight loop environment. While the self-driving vehicles are a prominent face of such connected autonomy, there are many more areas where such systems can be deployed.
Consider a grid operation where connected devices can forecast the transformer overheating and load spikes, requiring the corrective actions like load redistribution or demand curtailment. However, with humans taking the decision, the response time will be longer thereby creating conservative thresholds and capacity underutilization. Agentic AI based operations can able to make decisions much faster, within seconds leading to higher asset utilization. Similarly in a large warehouse where large number of robots are collecting, moving and delivering assets, a real-time tracking and management of routes and actions is required. One robot’s movement decision may impact multiple other robots in terms of collision, congestion / deadlock etc. While high level policies will be created by humans, the on-ground operation requires the real-time tracking and decision making about optimal task allocation.
Looking at a chemical process plant manufacturing the advanced composite material for high stress requirements (aircraft / automotive operations), real-time monitoring of various parameters like temperature, pressure, tool vibration, current fluctuation, resin viscosity etc leads to quality prediction where variation across multiple parameters even when each individual parameter is within limit, can lead to product failure. AI can act in real-time and eliminate the failures in such situation. Adaptive spectrum sharing and interference avoidance, traffic surge handing kind of scenarios in telecom requires autonomous decision making and are relevant use cases for agentic AI deployment, as human intervention can’t handle this in real time.
While adoption of Agentic AI can be highly beneficial, there are strong challenges to be addressed. First and foremost is the issue of Trust in Agentic AI solution, as we are ceding the operational control to a computer where decisions are not explicitly coded and hence not deterministic. For example, in a telecom network an agent tasked with maximizing the spectral efficiency may aggressively reallocate the radio resource within the technical specifications but lead to the service degradation for latency sensitive scenario.
From agent’s perspective the goal is met but operator may have an unhappy customer at hand. Similarly in a manufacturing or logistics operations an agent may foresee a probabilistic failure and based upon that make decision to downgrade production rate or reroute a shipment. There may be decision made to transfer inventory from a high demand region to another region based upon weather forecast, thereby degrading the local service or creating conflict with another AI agent responsible for operation in other location. A human operator may perceive these as unnecessary decisions having adverse short-term impact, thereby creating lack of trust in the capability of the agent.
To trust AI enough to handover the decision making requires predictable behaviour, bounded autonomy and repeatable outcomes. It is crucial that there be clearly defined lines which AI can’t cross, no matter what its intelligence indicates. These can be in the form of pre-authorized corridors, service thresholds which are not to be violated. The decision reasoning by the agent must be logged so that later stage it can be analysed by human operators. For deployments, it can be gradual deployment – starting with advisor, moving to limited autonomy and increasing as system continuously behaves in trusted manner.
Agentic AI marks a major shift in IoT system design and deployments. With the growth of physical and digital infrastructure scale and complexity, the challenge is beyond analytics – going towards the ability to act responsibly and consistently without human intervention. The transition has challenges but is unavoidable and requires organizations to take Agentic AI as engineering and process discipline to realize its benefits while keeping safety, trust and control.