Mobile Navigation

Automation & Control

View Comments PDF

Generative and Agentic AI: Intelligent Manufacturing Evolves

| By Mary Page Bailey

As many chemical producers are developing and implementing comprehensive AI strategies, newer AI paradigms, such as generative and agentic platforms are set to transform the benefits AI can provide

Due to the continuous generation of data and the complexity and intricacies of the chemical processes themselves, the chemical process industries (CPI) are especially primed to benefit from the vast computing and analysis power offered by artificial intelligence (AI) technologies.

“Chemical processors have the sophistication, patience and rigor to implement these technologies at scale. The amount of benefit they can get from just a 2 or 3% improvement is dramatic, whether it’s in emissions reduction or production output, it’s a huge gain. I think that’s why we see those industries take such a leading role with these types of technologies,” said Rob McGreevy, chief product officer at AVEVA (London, U.K.; www.aveva.com) during the AVEVA World conference (April 7–10, 2025; San Francisco).

To date, most applications for industrial AI have been for predictive applications, such as for maintenance optimization or anomaly detection, or in laboratory settings to streamline experimentation and discovery. But rapid technological advances are progressing industrial AI from predictive to generative and agentic, opening up distinct and complementary capabilities, including decision-making.

“Traditional industrial AI excels in structured operational tasks — analyzing sensor data for predictive maintenance, monitoring equipment status and health and optimizing process parameters through advanced analytics. This helps plants reduce, or better avoid, unplanned downtime, extend equipment life and maintain consistent product quality. Generative AI [GenAI] brings additional capabilities by making engineering, operations and maintenance more conversational and intuitive. While industrial AI continuously monitors and optimizes operations through machine learning algorithms, GenAI enhances human decision-making by providing context-aware support and making complex information accessible. GenAI-empowered systems can act as a digital knowledge repository, allowing workers to instantly retrieve best practices, troubleshoot issues (Figure 1) or even generate reports and documentation in seconds,” explains Steffen Lamparter, head of Siemens AG’s (Munich, Germany; www.siemens.com) Industrial Artificial Intelligence Research Group.

FIGURE 1. AI-powered copilots can help simplify workflows and provide troubleshooting recommendations for engineers and field operators

Generative AI

A key differentiator between traditional and generative AI in an industrial sense is its ability to reason, even when looking at seemingly disparate data.

“Traditional AI still has a bright future, but at the end of the day, a lot of AI at the asset level is about identifying patterns from the data, making predictions and automation. When generative AI comes in, instead of being able to do prediction on structured data like tables or time series, you are able to do reasoning on unstructured data,” said Julien Debard, director of energy and utilities at Databricks (San Francisco, Calif.; www.databricks.com) during the panel “How Industrial AI is driving industry change” at AVEVA World. Examples of these unstructured data sources include everything from a photo of an asset to a video stream of a leaky valve to a maintenance log document. “GenAI allows you to resonate on a much larger amount of data and to make better decisions. Also, GenAI, because of its reasoning, is able to connect datasets that we cannot easily connect today,” he adds.

On the same panel, Amir Banifatemi, chief responsible AI officer at Cognizant Technology Solutions Corp. (Teaneck, N.J.; www.cognizant.com), reiterated the significance of the reasoning capabilities of GenAI while also not losing sight of data trustworthiness. “Right now, we’re seeing an evolution in terms of what you call reasoning. What GenAI has provided is access to a knowledge pool. The knowledge pool could be a repository of trusted information within a company, including processes, collaboration patterns or incidents. The GenAI model would be able to borrow from this pool of knowledge to reason, predict and propose scenarios. This level of prediction is the new element we did not have previously, when it was just about pattern-matching,” he explained. However, he emphasized that industry cannot look at AI as an all-encompassing shortcut, and must consider safeguards: “The operational technology and the processes themselves have to be built such that what AI brings is going to be fully assured, controlled and reliable. Trustworthiness is so important.”

Industrial AI adoption has happened in a series of fits and starts for many companies who may have been early adopters for a very specific domain — for instance, pump predictive maintenance — but then have found it difficult to expand their AI programs due to vendor-specific data siloes. But these challenges can be avoided for companies who have used open-format cloud storage for their data, explained Databricks’ Debard. “Looking at the innovation feed that we have today, you cannot put you all your eggs in one basket with one vendor. My recommendation is to get your data estate in check, so that you don’t copy your data all over the place, and where you have clear access and governance to your data. Then anyone can come and help you with your AI.”

AVEVA’s chief technologist Arti Garg underlined the criticality of data-readiness in one of AVEVA World’s keynote addresses (Figure 2): “One of the biggest challenges isn’t the AI itself, but it’s getting your data AI-ready, making it error-free, and ensuring it is complete. 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, so making those data usable is a significant barrier to unlocking any insights.”

FIGURE 2. At AVEVA World, Arti Garg discussed agentic AI and its role in the next generation of industrial intelligence

Agentic AI

Agentic AI — purpose-built and prompt-oriented — helps with some of these barriers in that it can be tasked specifically with decision-making. Creating situation-specific dashboards on the fly to help guide troubleshooting decisions and visualize relevant data is an area where agentic AI excels for industrial users. “Instead of a hard-coded dashboard layout, agentic AI can make something more tailored to an organization and a specific use case,” said Garg. The agent creates a sequence of tasks designed to create the desired dashboard, and as it executes the tasks, it may even collaborate with subagents to help retrieve data. “Maybe it will work with a subagent that knows how to pick the right data for the dashboard asset you’re building, and maybe it will start suggesting the right thresholds or creating alarms. But what’s really critical is that it does this all on its own. It draws upon AI, it draws upon contextual knowledge and creates something unique without a lot of help,” she continued.

Beyond dashboards, other examples of agentic AI showcased at AVEVA World included an automated pump-troubleshooting interface and a plant-design agent that helped to optimize pipe routing. “Agentic and generative AI are not just about creating dashboards. AI actually makes the process of designing a new asset much, much faster. That’s a very useful step for anyone who has to design complex assets and probably is having trouble finding the talent to work on engineering design,” said Caspar Herzberg, chief executive officer of AVEVA. Furthermore, he stated that AI “means you can get people up to speed more quickly, and it means switching between technologies will become easier.”

Siemens’ Lamparter underlined the potential for AI agents and copilots to improve workforce concerns: “Industrial copilots are emerging as a particularly effective solution to address labor shortages by enabling workers to accomplish complex tasks more efficiently. For instance, these AI assistants can help engineers create panel visualizations in just 30 seconds and generate code that requires minimal adaptation, significantly reducing the expertise barrier for new workers.”

Taking a step further into agentic AI in operational environments, he sees the agentic AI as the “future of industrial automation, where AI systems could potentially take limited autonomous actions within carefully defined operational parameters. This might manifest as systems that not only detect issues but also initiate corrective actions — adjusting process variables, rerouting production flows or scheduling maintenance interventions without human prompting.”

 

How is industrial AI being used?

Many specialists may tout the incredible capabilities of industrial AI, but how are these types of technologies actually creating practical value for users? “There were periods of experimentation in 2023 and 2024, especially with GenAI and large language models [LLMs], but the focus now is on real business value. I think predictive maintenance is one area that we’ve been seeing value from for a number of years, but LLMs are also adding a lot of new use categories,” said Erik Brynjolfsson, director of the Digital Economy Lab at Stanford University (digitaleconomy.stanford.edu), in one of AVEVA World’s keynote panels (Figure 3).

FIGURE 3. Erik Brynjolfsson (left) and Caspar Herzberg (right) join Schneider Electric CEO Olivier Blum (center) at a keynote panel at AVEVA World

A number of these real-world examples were highlighted at AVEVA World. In his keynote address, AVEVA CEO Herzberg detailed AI projects with global mining firm Boliden AB (Stockholm, Sweden; www.boliden.com) and specialty chemicals manufacturer Albemarle Corp. (Charlotte, N.C.; www.albemarle.com).

At the Aitik open-pit copper mine in Northern Sweden — one of Europe’s largest, producing some 40 million tons/yr of ore — Boliden has implemented prescriptive AI analytics supported by AVEVA to integrate its conveyor condition-monitoring data with vibration data from its bearing manufacturer SKF AB (Gothenburg, Sweden; www.skf.com). “The two companies are integrating SCADA [supervisory control and data acquisition] and condition monitoring in context, sharing continuously with remote experts and leveraging advanced analytics for real-time, predictive and prescriptive insights to reduce downtime across multiple applications and multiple vendors, centralizing information in the cloud. This is smart maintenance. This is a new way of working at a distance, transversally, across organizations as a connected ecosystem,” said Herzberg.

At Albemarle, Herzberg explained that AI is being used to capture and manage vast amounts of data across facilities on four continents “so that decision-makers can spend less time searching for information and more time turning insights into improved efficiency.” According to Albemarle, these programs have saved the company over $150 million.

Jonathan Alexander, Albemarle’s global manufacturing AI and advanced analytics manager, further discussed the expansion of the company’s global “Albemarle Intelligence” AI strategy: “It started with providing context around process signals, and sampling and curating them to solve certain business problems at scale. Then, with data curation, data governance and data organization, we had a standardized data infrastructure, and with all this extra context, we could add standardized analytics and algorithms that could scale.” He notes that the platform uses two main algorithms: one in the form of unsupervised machine learning using principle component analysis for outlier anomaly detection to help with predictive maintenance and identification of process variations; and a supervised machine-learning algorithm that is akin to advanced linear regression for value prediction.

Another real-world AI use case demonstrated at AVEVA World was the use of agentic AI for asset management and support across Methanex Corp.’s (Vancouver, B.C., Canada; www.methanex.com) global methanol production network. “It is a software program where you don’t need to write the business rules into the platform. You don’t program them. You allow the AI to make the decision about how to complete the task requested by the user,” explained Julio Figueroa, IT business relationship manager at Methanex, noting that the platform is enabled to access file data, as well as a document library containing procedures, forms and manuals, and to send emails. The agent can be integrated with various communications platforms, such as Microsoft Teams or Telegram, to respond to queries about assets. The recommendations proposed by the AI agent can be used to help operators prioritize their actions to focus on plant reliability or efficiency in a given scenario.

Franco Branca, lead project engineer at Methanex in New Zealand, detailed another AI operator-assistance project underway: “We have a process in our plant that is manual and non-linear, which means if an operator makes a change, there’s not a direct, clear correlation to what that change is going to cause. That means these changes are normally time-consuming, and they’re left for when the night shift comes on or a weekend. What we realized is it would be helpful to train an AI. This is a perfect problem that an AI can solve by looking at history. So we used the traditional method to train an AI, taking process data and manual data that we’ve collected over time.”

After training, the AI was able to provide a visual output to the field operator with recommendations on the best next steps to optimize the process. “The operator then makes those changes and feeds in the output of what is done to the AI model. It then updates and it recommends another change, so it’s helping the operator to quickly optimize the plant without having too many iterations. It’s helped us bring together raw data that we’ve collected over a long time, making the data real for our operators, making their lives easier and making our plants more efficient,” said Branca.

 

AI for decarbonization

As companies move forward with ambitious decarbonization strategies, AI technologies can help improve efficiency and expedite the adoption of newer, climate-friendly technologies.

“I think AI is going to end up being able to reduce waste that occurs from the use of water, air, steam, gas or electricity in a meaningful way. The other area where AI will be impactful for decarbonization, I think, is in the sophistication of some of these direct-air-capture mechanisms. There are a lot of physics-based, first-principle-based methodologies that are being applied, optimized and tuned. And AI is just perfectly suited for that,” said AVEVA’s Rob McGreevy.

Other decarbonization avenues being supported by AI include plastics recycling, for example via AI-facilitated sorting processes, and “green” hydrogen production via new approaches to optimization. “In hydrogen and ammonia production, AI can optimize electrolyzer efficiency, predict renewable-energy availability and improve process stability,” noted Siemens’ Lamparter. Siemens is currently developing an “intelligent chatbot” dedicated to creating GenAI-supported plant designs for hydrogen production. ■