Industrial operators are increasingly looking to AI, and specifically, AI agent technology, to help make more efficient use of available data and to automate industrial workflows. Learn how.
Industrial operators are besieged with data these days. Putting that data to use can be a challenge. Increasingly, the industry is looking to AI, and specifically, AI agent technology, to help make more efficient use of available data and to automate industrial workflows.
RTInsights recently sat down with Jason Schern, Field CTO at Cognite, to talk about the growing importance of AI agents in industrial operations, what issues they help overcome, how to get started, and what benefits an organization might reap from using them.
Here is a lightly edited transcript of our conversation.
RTInsights: What is an AI agent, and why are they a compelling technology for industrial operators today?
Schern: An industrial agent is simply an intelligent entity that can be used to perform a very specific task like finding data, answering a query, creating an analysis, or even writing code. What makes it different from a chatbot is that it doesn’t need to be programmed specifically for every possible scenario or task; rather, it uses data and reasoning in its response.
Industrial agents require a few key components: They need a 1) language model with reasoning, 2) Instructions for solving the task, 3) Specialized tools for industry, and 4) access to relevant data in context.
This combination makes industrial agents very powerful at performing certain tasks, and they can even get better at them over time. The human SMEs or engineers that use these industrial agents get an immediate productivity boost, and they can even be used to automate certain parts of a workflow. I predict that the best-in-class operators of the future will use a portfolio of industrial agents discerningly across their operations as this technology further matures.
RTInsights: What are some examples of these types of agents already in the field?
Schern: Industrial agents can be used for many cases, such as improving root cause analysis, issue summarization, creating charts, populating data models, etc. They are especially useful when the business problem requires a lot of human-hours to execute and complete.
For example, we have a large energy customer who currently gets lots of technical documentation from their suppliers. All this information comes through in the format that the suppliers have always given – but this information has to be normalized for the customer to make sense of it. A human expert could go through and normalize it all, but it would take years.
This is a perfect use case for an industrial agent. It would be impossible otherwise to write a program that takes into account every variation, so instead, the agent can use reasoning to
go through the documentation, find the relevant information, and put it correctly into their data model for easy reference and retrieval.
RTInsights: Gen AI is in its second “wave” of maturity – what is different this time around that makes it more suited for Agent-based operations?
Schern: Reliability is key for industrial operations in energy, chemicals, etc, so users have to be able to trust the outcomes from LLMs before they can adopt it into a critical workflow.
Back in 2023, when generative AI’s results initially started to be taken seriously, tech thought that you could solve almost any search-based use case with a large language model that “knew everything” because it was trained on all data possible. But that type of solution led to LLMs that would create hallucinations – where answers were made up and not always based on reality. As you can see pretty quickly, this questionable accuracy doesn’t meet the standards for critical operations.
Fast forward to today, the approach that is working much more effectively is to use the most appropriate language model for a given task. Additionally, rather than relying only upon the information a model has access to through training, we have frameworks that allow us to provide access to proprietary and up-to-date operational data as the source and context for the model to use in formulating an answer. That way, you can take advantage of the reasoning capabilities provided by language models and use them only with data that you know is relevant to the problem you are trying to solve. When you’re using the best language model for the problem with the right set of data, then you can get to outcomes that are trustworthy enough to be used in daily decision-making.
RTInsights: What technical components and techniques do you need to bring these agents to life? What’s the secret to being able to trust them?
Schern: There are a few key things here. First, you need to be able to identify and use the right language model for the problem at hand. At Cognite, we’re constantly evaluating and benchmarking the models we use for certain problems because the performance can change over time. This is important to help prevent hallucinations and get you the most accurate results.
Secondly, the language models need to be provided with access to relevant content that they should utilize as the source for responding to a question or prompt. This works when you have all your proprietary data in context- you can identify the landscape of data that is available to solve the problem, gate irrelevant data, and then instruct the model to respond using only that set of data. Because the data sits in a contextualized data model, you can enrich the solution space over time without sacrificing the quality of output.
Lastly, you need to give the agent very specific instructions. Think about how if you had a new intern, you’d need to teach them how to solve a general set of problems with clear guidance. The better you understand your business problem, the better you can “teach” the agent with clear instructions to get what you want.
RTInsights: What should organizations focus on first in order to succeed in their roadmap?
Schern: Anyone who wants to have success with industrial AI agents has to get their data in order first. Just like any AI application, if it’s garbage in, it’s garbage out – and the technology will never get adopted and improve operational workflows.
What we’ve seen work best so far is when structured and unstructured data can come together in an industrial knowledge graph. This means that the data is contextualized together so that the industrial agent knows exactly where to look for the information that’s relevant to the business problem.
The beauty of this approach is that if you get your data in order, you can get a whole lot more value from it, even if you’re not ready to start using industrial agents. Think easier dashboarding, faster time to answer, smarter troubleshooting, etc. You’ll increase the productivity of your workers and make smarter decisions on a day-to-day basis. The only question you have to ask yourself is how quickly you want to get to that future.
