Why this article is worth your time ...
- Autonomous Task Execution: Agentic AI systems use contextual knowledge to autonomously access, process and visualize operational data, enabling tasks like creating dashboards and setting performance thresholds without coding.
- Enhanced Data Integration: Through self-learning, AI agents reconcile disparate industrial data, harmonizing multiple data sources to build unified digital twin models essential for modern industrial operations.
- Human-AI Synergy: While agentic AI automates complex tasks, human expertise remains crucial to refine, correct and contextualize AI-driven insights.
If you haven’t yet heard much about agentic AI yet, brace yourself because you’re about to get flooded with references to it. Why? Because it’s fast becoming an integral aspect of artificial intelligence use in industrial operations software — just like we’ve seen with AI data analytics and, more recently, with generative AI applications in industrial AI assistants and co-pilot technologies.
For an introduction or refresher on the different types of industrial AI, see the sidebar at the end of this article. And to learn more about agentic AI in particular, check out this podcast discussion with Tim Gaus from Deloitte.
To help frame agentic AI’s applicability to industrial production operations, Aveva provided a very useful demonstration at its Aveva World 2025 event in San Francisco. The demo showcased Aveva’s forthcoming industrial AI assistant running on the Microsoft Azure OpenAI Service.
In this demo, Arti Garg, chief technologist at Aveva, began by explaining that an AI agent is a system that leverages AI to perform a task by accessing system data such as operating temperature, pressure specifics or a safety metric. The key here is that, based on a prompt from a user, the AI agent can gather the data needed to create a dashboard for the user showing them the information they need without needing to hard code a new dashboard visual.
She noted that these AI agents can also work with sub-agents that know how to select and retrieve the right data for, say, an asset monitoring dashboard. Here, Garg explained the agent isn't just building the dashboard, it can also suggest the right operating thresholds for the asset it’s monitoring and create alarms.
“What's really critical to understand here is that the AI agent can do this on its own by drawing upon its contextual knowledge,” she said. “It is this autonomous nature of agentic AI that has huge potential to bring new efficiency gains to industry. It will change how we carry out our work today.”
Creating an asset monitoring agent
For the demo, Garg and Iju Vijaya Raj, executive vice president of R&D at Aveva, showed how an operator who wants to do an unplanned optimization of a condenser can use agentic AI for this task.
The initial dashboard view used in this demo came from Aveva Connect — Aveva’s platform to aggregate, curate and share information across multiple sites and data types to provide a holistic view of a business. This dashboard view displayed contextualized data from an air-cooled condenser at a power plant.
In the example, the operator, concerned about a drop in the plant's performance, wants to assess the impact of fouling on the condenser and asks the industrial AI assistant to monitor its performance and diagnose any issues it finds.
Based on this prompt, the AI Assistant retrieves a list of available agents deployed at the power plant to verify that such a monitoring agent hasn’t already been developed for this condenser. Seeing that one has not yet been set up, it asks the operator if one should be created and deployed on the unit. The operator then asks the AI assistant to create a condenser monitoring agent and train it to monitor the active power of the unit and turbine exhaust pressure based on relevant parameters that could affect its performance. Following this prompt, the AI assistant initiates the agent creation process, the details of which are visualized for transparency to the operator.
This transparency is a key issue of importance for AI applications in industry, as many experts have expressed concerns about how AI technologies arrive at the answers they provide. With this transparency, the operator can see the data sources and tags associated with this agent to verify that it is being assembled correctly.
The operator then asks for deployment of the agent and to run the model on a 30-minute interval. According to Aveva, that’s all it takes to set up the industrial AI assistant to perform a task.
Two weeks later, the operator revisits the unit and asks the AI assistant to show the results of the deployed agent. These results showed a degradation in performance of the condenser due to fouling. Visualization of the data trend lines help the operator determine if the issue needs to be corrected immediately or if it can wait until the next maintenance event.
For further insights related to the potential impacts of the condenser fouling, the operator asks the AI assistant to calculate the power loss costs and to determine the payback time associated with the cost of cleaning the condenser. The agent generates the dashboard chart for the operator to review.
Based on this, the operator asks the AI assistant to summarize the condenser issue and inspection procedures to create a maintenance work order. The AI assistant then generates the procedures and provides a link to the manual in the Connect dashboard.
Garg noted here that “this is an example of using AI to make something complicated very simple. The operator could create a new agent without developing complicated code and do it through same interface they already use every day.”
AI gets your data ready for AI
One of the biggest challenges of AI isn't the AI itself, Garg said, but getting your data AI ready — making it error free and ensuring it's complete. “And this is something that's only becoming harder as industrial settings become more complex. You have so many different types of assets and they generate more data than ever before,” she said.
This led to another demo that showed how AI agents can address typical asset data management issues, such as:
- Having an asset’s data collected by multiple applications.
- Issues with asset data stored in different places.
- Storage of data points using different naming conventions.
- Having data sets with overlapping information.
Reconciling these data issues can be extraordinarily difficult, Raj said. “Even at Aveva, we’ve wrestled with this problem for decades. That’s why this has been on top of mind for us when we we’re exploring new AI capabilities. Here, we wanted to use the self-learning process to bring separate data together and harmonize them to create a digital twin model.”
In the demo showing how AI agents can help reconcile asset data, asset maintenance information from an SAP ERP system along with stream data from an Aveva PI System were used.
The PI data showed that some asset information was encoded into stream names, while other relevant information resided in comments. It also showed that, as the company’s systems grew, inconsistent naming conventions emerged.
Garg explained that the AI agent takes this information and makes inferences. That is, it starts to figure out what the data means by looking at abbreviations and comments. From this it can learn about equipment codes and then apply that knowledge to help categorize and tag the data consistently across the facility.
For example, from the PI stream comments, it sees that HX means heat exchanger. After it learns this, Raj said, it can use the equipment code to infer equipment types, even in streams without comments.
The need for a human-in-the-loop remains
Despite all its autonomous capabilities, Garg and Raj noted that a human still needs to be involved to clarify, correct and refine the inferences the AI is making.
Here, existing naming guides can be used to teach the AI about your company’s naming standards. Once the AI understands these standards, it can proceed to make even more inferences about data in the system.
After we tell it about our naming rules, it learns from them and uses them to fill out site numbers and line numbers for us, said Garg. “Once the AI gains some basic understanding of our systems and standards, it starts linking our systems together. This demo showed the outcome of our AI driven data mapping — a combined view of our PI Data and our SAP work management system together in one place. This can then be used to set the foundation for a unified digital twin.”
Garg explained that this demo was meant to highlight how AI can help deal with one of the hardest problems in the industrial world — aligning data models to create a digital twin.
“But when I think about this example, even when using AI to help, synchronizing data across different sources requires some underlying knowledge of the data itself and the environment from which it came. It still requires some industry experience and expertise,” Garg said.
Raj concurred, adding that “what's really interesting is the relationship between data and context. To generate insights, you either need an experienced person or an established process. At Aveva, we think the next generation workforce will generate insights through AI as well as human intelligence.”
More Agentic AI and GenAI content from Automation World:
Leaders relevant to this article: