In Stay Sharp Episode 125, Strategic Patterns for Implementing AI in Manufacturing, co-hosts Jonathan Scott and Juliann Grant sit down with Diego Tamburini, PhD, the AI Practice Director and Executive Consultant at CIMdata. A recognized leader in using AI to drive digital transformation in the industrial sector, Diego brings a wealth of experience with AI to the conversation, including pioneering the use of AI across Microsoft’s marketplace engineering organization. To shed light on the reality of the “all-AI all the time” message these days, the trio dive into real stories of how AI’s being used, what’s working, and what’s not.
Diego Tamburini, PhD
How AI is Being Used
Diego starts by clarifying that what AI means in this conversation is generative AI that’s run by large language models (LLMs). AI and machine learning aren’t new, he tells us. What’s new is having an assistant based on an LLM that’s really good with natural language—and that’s what’s changing how humans interact with software. “Instead of clicking 27 menu choices, you ask the question,” he says. “The search and query of information [is] the most common application of generative AI nowadays.” The ease of interaction is what’s made AI go mainstream.
Diego outlines five patterns of AI use and implementation within industry. The first is the kind of help we’re all used to seeing—software offering help as we do our work. This is AI embedded in a platform, typically an AI assistant driven by an LLM, but this type of implementation could also be embedded AI capabilities that we don’t even realize are AI-driven, such as generative design in CAD tools. The second type of AI use arises because customers need to span multiple platforms or disciplines, such as PLM, ERP, and supply chain management. For instance, he says, “If you have a question, let’s say, I’m going to change the design or the material of this part, what are my supply chain risks? PLM alone does not have the answer to that question.” In that case, the customer needs to do retrieval augmented generation (RAG), which involves gathering up-to-data from external, trusted sources, like other platforms and giving it to AI to process.
The third pattern of AI use in manufacturing involves the customer using and training their own machine-learning models—it’s “a whole new level of sophistication,” Diego says, that involves employing AI and machine-learning experts, as well as data governance and more. Johnathan describes it as controlling how the entire sausage gets made, and Diego notes that this is the pattern of use that’s actually been around the longest, in implementations such as predictive maintenance applications.
The last two common patterns of AI use are both about workflows. The fourth, AI augmented workflows, concerns predetermined processes for which you already know the steps or decision points. The fifth is agentic AI workflows, or AI on steroids. Diego says “They’re not just question and answer. It is more, I give it a goal and the agent can define and orchestrate the steps that it needs to perform to achieve that goal. So it is dynamic and it changes depending on the answers.”
Challenges and Advice
Diego identifies the biggest challenge for companies wanting to implement AI as sorting out what’s hype and what’s the reality. He says, “Like any technology, the biggest mistake one can make is having a technology looking for a problem.” What’s most important is identifying and prioritizing how you’re going to apply AI, how to plan for scaling it from a proof-of-concept to something larger, and evaluating if what you can buy is sufficient or if you’ll need to customize it. All of that has implications for staffing and potentially re-skilling. “It goes back to the patterns,” he says. “Depending on which ones you are actually touching, you need different skills. If you’re only using AI that is in the tooling, you need to be a power user of these capabilities and be really good at prompting and understanding how to interface with AI. On the other hand, if you’re actually developing your machine learning models, you probably need data scientists or PhDs in your company.”
Diego’s parting advice is to understand the problem you’re trying to solve, the AI technology, and your data, and then put governance and structure around how you’re going to use the new technology. “This is not only for AI, it’s for any technology around how you’re going to prioritize, manage, and resource your proof of concepts and proof of value,” he concludes.
Learn More About AI in Use
The full podcast contains more discussion and details on a variety of topics, including how surrogate models can be used to speed up design and finite element analysis, why Diego says RAG is less like retraining AI and more like cheating, how CIMdata can help your organization decide how you should be applying AI, and more.
Be sure to check out Stay Sharp Episode 125: Strategic Patterns for Implementing AI in Manufacturing and join us each week for a new podcast.
Follow the Razorleaf Podcast, Stay Sharp in Digital Engineering on:
