{"id":171135,"date":"2026-05-23T18:03:49","date_gmt":"2026-05-23T16:03:49","guid":{"rendered":"https:\/\/www.startupbusiness.it\/from-iot-to-aiot-when-data-becomes-decisions\/171135\/"},"modified":"2026-05-24T10:57:06","modified_gmt":"2026-05-24T08:57:06","slug":"from-iot-to-aiot-when-data-becomes-decisions","status":"publish","type":"post","link":"https:\/\/www.startupbusiness.it\/en\/from-iot-to-aiot-when-data-becomes-decisions\/171135\/","title":{"rendered":"From IoT to AIoT: when data becomes decisions"},"content":{"rendered":"\n

In my previous article, I described physical AI as the transition from artificial intelligence<\/a> that interprets the world to artificial intelligence that interacts with the physical world. For e-Novia, this involves integrating advanced sensory perception, automated reasoning models and real-time control, so that systems can adapt and respond to the context in which they operate. <\/p>\n\n

What is AIoT?<\/h2>\n\n

AIoT, on the other hand, arises from the integration of the Internet of Things and artificial intelligence, but it should not be reduced to the formula \u2018sensors plus algorithms\u2019. The IoT is typically characterised by fleets of sensors that send their data to a cloud platform. This is certainly useful, but it is not yet intelligence. <\/p>\n\n

The next step is to introduce algorithms that process this data in the cloud to extract new insights: by detecting anomalies, identifying correlations and predicting behaviour, the cloud is not merely a repository for data, but a space where a wide range of AI algorithms are applied, from the simplest to the most advanced, right up to generative AI algorithms.<\/p>\n\n

This is where the IoT shifts its role. It evolves from a data-collection infrastructure into a basis for inference \u2013 that is, for deducing a current state or predicting a future event. This does not always lead to automatic action. Sometimes it triggers an alert, a recommendation or a maintenance priority. In any case, the data ceases to be merely telemetry \u2013 that is, a remote measurement of a physical phenomenon \u2013 and becomes input for a decision. <\/p>\n\n

And what if the platform feeds the extracted information back to the sensors themselves, enabling them not only to perceive the context in which they operate, but also to react based on information derived from data relating to the entire fleet? This brings us to the convergence of physical AI and AIoT: smart devices that react and enable other devices in the fleet to adapt their behaviour and make better decisions based on shared information. <\/p>\n\n

Imagine machines that collect data, process it taking its context into account, and then send this data to a cloud platform. The platform, in turn, processes it using AI algorithms and sends the information back to the entire fleet. This enables, for example, a machine to anticipate a fault in a situation it has never encountered before. <\/p>\n\n

Whilst physical AI brings intelligence to individual devices, physical AI combined with AIoT enables smart devices to share experiences, thereby enriching the individual device\u2019s knowledge base.<\/p>\n\n

The IoT in 2026: from connectivity to distributed inference<\/h2>\n\n

The first phase of the IoT was driven by connectivity. The challenge was to bring objects online that had previously been unable to communicate. Industrial machinery, consumer products, vehicles and plant have become sources of data. <\/p>\n\n

In 2026, the challenge is a different one. Many companies already have sensors, platforms and data. The question is how to use that data effectively and to generate value. <\/p>\n\n

This is where distributed inference comes into play. This term refers to the ability to run AI models not only in the cloud, but also close to where the data is generated. This can take place on industrial gateways, advanced microcontrollers, edge devices or embedded systems \u2013 that is, electronic systems integrated directly into the product or machine \u2013 and it is here that physical AI converges with AIoT. <\/p>\n\n

The cloud remains essential for training models, aggregating data and managing fleets of devices. However, not all decisions can wait for the cloud. In an industrial context, certain analyses must be carried out locally for reasons of latency, operational continuity or security. <\/p>\n\n

This is one of the most significant differences between AI applied to digital data and AI applied to physical systems. In the former case, the response time can be a matter of seconds. In the latter, it must be compatible with the process. <\/p>\n\n

From the IoT platform to AI-ready architecture<\/h2>\n\n

At e-Novia, we began working on these issues years ago, long before the term \u2018AIoT\u2019 became widely used. At the time, the focus was mainly on the IoT, cloud platforms and device management \u2013 that is, the remote management of connected devices. <\/p>\n\n

The issue was already very real. An IoT platform had to do more than simply collect data from devices; it had to establish an infrastructure that would support future development. Even then, it was clear that the value would not lie solely in connecting the device. The key was to use that data to enable automation, diagnostics and new services. <\/p>\n\n

Let\u2019s consider an industrial application for predictive maintenance on machines spread across multiple plants. In this field, there are many IoT platforms, differing in architecture, functions and level of specialisation. Think.Link, the platform my team has developed, is a useful example to illustrate the role of this technological layer. It serves to collect data from devices and sensors, manage the identity of connected objects, standardise data flows and integrate that data with external applications via APIs. <\/p>\n\n

In a scenario such as this, the platform is not the end goal. It is the layer that makes the next step possible. If data arrives haphazardly, in different formats, or with incorrect sampling, even the most advanced AI model will not deliver the expected results. If, on the other hand, the architecture is designed correctly, AI can be integrated into the process with greater reliability. <\/p>\n\n

From a technical perspective, an IoT platform must therefore manage communication security, software updates, data quality, integration with existing systems and, above all, be designed to support the execution of any type of AI algorithm. Without these elements, AI cannot unleash its full potential. It may be highly sophisticated, but it operates on fragile or incomplete data, or simply cannot find an architecture that allows it to run correctly. <\/p>\n\n

Over the last ten years at e-Novia, having led dozens of IoT projects, I have often seen the same pattern emerge: the AI model has rarely been the main constraint. Far more often, the problem to be solved lies upstream \u2013 in the quality of the data, in compatibility with existing systems, or in the difficulty of integrating a prototype into a real-world process. <\/p>\n\n

Data, interoperability and legacy systems<\/h2>\n\n

In industrial settings, AIoT immediately reveals its architectural nature. The challenge lies not merely in capturing a signal from the physical world, but in integrating it into a data chain that is compatible with automation systems, management platforms and processes that have often been in place for years. <\/p>\n\n

A factory is not an abstract concept. It consists of machinery installed at different times, PLCs, SCADA systems, MES, ERP, local databases, operating procedures and the tacit knowledge of the operators. The PLC, or Programmable Logic Controller, is the controller that governs many industrial automation systems. SCADA refers to supervisory control and data acquisition systems. MES and ERP manage production and business processes respectively. <\/p>\n\n

These systems are rarely designed to interact seamlessly with AI models. Interoperability therefore becomes a design requirement, not an afterthought. <\/p>\n\n

Integrating AIoT means building a robust data pipeline. Data must be collected, cleaned, contextualised, filtered, stored and made available to algorithms or operators. Each step can introduce errors. A predictive model trained on unrepresentative data will produce unreliable results even if the algorithm is correct: AI algorithms are almost always \u2018garbage-in, garbage-out\u2019 systems. <\/p>\n\n

This is why AIoT cannot be treated as an application layer to be added at the end of a project. We see this frequently in the projects I oversee at e-Novia. If data, software, hardware and operational constraints are not considered together from the outset, even the best AI model will fail to deliver value. <\/p>\n\n

AIoT in industrial products and processes<\/h2>\n\n

In our innovation projects, whether within the venture studio or in collaboration with other companies, we have consistently found that the greatest value is created when physical data is linked to a specific operational need. It is not just about measuring; it is about making that measurement usable for predictive maintenance, alerts and new services. <\/p>\n\n

One of the projects we have worked on that can be publicly disclosed, and which fits neatly within an AIoT framework, is Enyring, a Yamaha venture focused on electric urban mobility and the management of battery fleets for battery-swapping services. In this context, the challenge was not simply to collect data from batteries and vehicles. It was to build a cloud-to-vehicle system capable of monitoring availability, operational status and usage cycles. <\/p>\n\n

This reasoning also applies to applications that are less visible but very common in industry. These include quality control using computer vision, energy monitoring, diagnostics on distributed machines, and maintenance optimisation. AIoT is not a single product category. It is an approach to designing connected systems that are designed to generate operational intelligence. <\/p>\n\n

What changes for start-ups and businesses<\/h2>\n\n

For a start-up, AIoT is a promising field, but one that is far more complex than pure software. It is not enough to have a good algorithm. You need expertise in electronics, firmware, cloud computing, AI engineering and industrial integration. The prototype must work in a real-world environment, with data noise, physical constraints and connectivity that is not always perfect. <\/p>\n\n

For a business, however, AIoT offers a practical way to improve existing products and processes. It often starts with a simple question: how can I reduce machine downtime? How can I gain a better understanding of how a product is actually used? How can I turn an asset into a service? <\/p>\n\n

The answer is never purely technological. It is architectural. We need to understand what data is available, what is missing, where it is best to process it, which systems need to be integrated, and which decision we want to improve. <\/p>\n\n

In this sense, AIoT does not replace the concept of physical AI, but makes it more tangible in many industrial applications. Physical AI refers to the development of systems capable of perceiving and interacting with the physical world. AIoT describes one of the layers that make this development possible: the layer in which connected objects, sensors and algorithms begin to generate useful intelligence for products and processes. <\/p>\n\n

It is a transition that we at e-Novia have seen emerge in numerous projects over recent years, even before the market had given it a name. Today, that name exists. The challenge remains to design systems that do not merely collect data, but ensure that it is reliable, interoperable and usable to inform better decision-making. <\/p>\n\n

References<\/em><\/strong><\/p>\n\n