Let’s be honest about Industrial IoT: we are incredibly good at generating alerts. We have sensors for everything, rules for every threshold, and dashboards that light up the moment a metric slips. But there is a massive gap between detecting an anomaly and actually solving the problem.
Operations, Support or Field Service teams today often find themselves managing “something is wrong” notifications that lack the specific context needed for immediate action. Instead, they need to do several manual look-ups, check historical records in another system and/or call upon help from another expert. This leads to high Mean Time To Resolution (MTTR) and a frustrating sense that analytics is just adding noise rather than improving quality of life.
On the other hand, we are living in the age of Large Language Models (LLMs) and Agentic AI—technologies that promise to move us from “data processing” to “reasoning.” In our daily lives, these tools act as co-pilots, summarizing complexity and suggesting actions.
At Cumulocity, we believe it is time we bring that same “brain” to the data stream. Streaming Analytics should evolve beyond (advanced) analytics; it needs to combine real-time processing with human-like reasoning.
The Two-Expert Approach: The Watchdog and the Investigator
Today, we are thrilled to introduce two new preview blocks designed to shift the paradigm from raw alerts to ready-to-use answers. By bringing Agentic AI directly into Streaming Analytics, Cumulocity is pushing the next evolution of IoT. We are moving from a platform that just “watches” to one that “thinks.”
This requires two distinct experts working in tandem: one for robust monitoring and anomaly detection at scale, and another for context-aware investigation. If you were to draw this architecture on a whiteboard, it’s a beautifully simple two-stage filter. Rather than building complex custom integrations, you simply configure a high-level pattern that bridges math and meaning.
As always, Streaming Analytics is your “fast expert.” It is brilliant at math and excels at crunching thousands of data points a second with incredible precision. However, raw data by itself often needs a layer of situational context to become truly actionable. Conversely, Agentic AI acts as your “thoughtful brain.” It is excellent at reasoning and synthesis, but you wouldn’t want it trying to process raw vibration telemetry in real-time.
By splitting these responsibilities through easy-to-use blocks, we create a powerful new pattern:
- The Watchdog (Fast Math): Sitting in the high-speed data stream, it constantly asks, “Is this normal?” This layer handles everything from simple rules to complex patterns powered by ML, catching the subtle drifts that traditional logic might miss.
- The Investigator (Deep Context): When the Watchdog barks, the Investigator wakes up. It looks at the anomaly context and asks, “What is actually happening, and what should the engineer do?”
This ensures you aren’t wasting expensive LLM resources on healthy data, while moving beyond the limitations of static thresholds that often fail to capture the nuance of complex machine failures.
Bringing it to Life: The ONNX and AI Agent Blocks for Analytics Builder
To make this Watchdog/Investigator pattern a reality, we’ve added two blocks into Analytics Builder that handle the heavy lifting of data orchestration and model execution behind the scenes. Crucially, because these are integrated into our Analytics Builder, they are accessible to any kind of business user—not just developers. If you can drag and drop a block and fill in a configuration field, you can now deploy sophisticated AI. By removing the need for custom coding or complex “plumbing,” we allow the people who understand the machines best to be the ones who define the logic.
The ONNX Block (Giving Your Watchdog a Bigger Brain)
Analytics Builder has always provided the foundational blocks you need to create reliable watchdog logic—often, a simple threshold rule is all it takes to spot an obvious issue. But complex failure patterns, like subtle sensor drift or multi-variable anomalies, require more than basic logic. That’s where the ONNX block comes in. It essentially gives your watchdog a bigger brain, letting you take machine learning models your data scientists have already built and run them natively in the stream. By handling the model execution “plumbing,” Cumulocity allows business users to simply plug in a model and start seeing results without writing a single line of deployment code.
The AI Agent Block (Your Investigator)
Once an anomaly is flagged, the AI Agent block securely passes the context to an LLM. Using a simple configuration-driven prompt template, you can define exactly how the agent should reason. This bridges the gap between a mathematical “blip” and a practical maintenance taskIt empowers domain experts to tune the AI’s “advice” to the specific needs of their use-case, ensuring the output is always practical and actionable.
Stop Alerting, Start Solving
By combining the precision of traditional ML with the reasoning of the Agentic Age, Cumulocity is changing how operations teams work. Because we provide the ready-made interfaces for these advanced technologies, your team can spend less time on engineering “the how” and more time perfecting “the what.” Instead of red lights, your engineers get prioritized incident dossiers that cut investigation time to nearly zero.
Interested to try it out yourself? Activate the preview feature and you will see these blocks appear in the list of “Calculation” blocks when working on the Analytics Builder Canvas.
- Detailed information on the ONNX block can be found here in the documentation.
- Want to learn more about why Cumulocity uses ONNX as its standard for ML inferencing? Read more in this article.
- Detailed information on the block can be found here in the documentation.
- Want to learn more about Cumulocity’s AI Agent Manager? Read more in this article.
Coming Soon: In our next post, we’ll move from theory to practice with a deep-dive into a predictive maintenance use-case, showing you exactly how to wire these blocks for real-world results.