Podcast: Using company-specific GPTs to unlock manufacturing data
Key Highlights
- AI adoption in manufacturing succeeds through change management, not complex technology.
- Secure, company-specific AI tools deliver more accurate results than public AI platforms.
- Manufacturers can preserve tribal knowledge by documenting expert processes with AI.
- Starting small with focused AI use cases helps manufacturers scale adoption successfully.
The concept of a “company GPT” is helping small and midsized manufacturers grasp the potential for artificial intelligence in their enterprises, and make real progress in their journeys. In this episode of Great Question: A Manufacturing Podcast, Amatrium Inc. co-founder and president Andrew Halonen discusses his recent article "AI Is a Journey for Foundries and Equipment Builders" and why he argues that manufacturers’ difficulties in implementing AI are about change management, not technology.
Below is an excerpt from the podcast:
Robert Brooks: We've spoken in the past, but recently you've alerted me to some developments in the work you're doing with metal casters and their equipment suppliers. Will you recall for our listeners what Amatrium does? What is the generative AI tool, Amatrium GPT?
Andrew Halonen: Amatrium GPT is really a collaboration tool at a higher level. So if you think about any organization, whether it's a metal casting foundry or equipment producer, they have been in business for many times decades, and they have an enormous amount of data that's sitting idle because it's hard to find. So it's really a tool to put that data to use.
RB: Again, another term you introduced, at least introduced to me in your article is a company GPT. Will you tell us what that is?
AH: The company GPT is really this hub, this hub of information. So if you were let's say a producer of a product, and it could be something like a commercial, an IR camera for detecting heat. Well, if you got your name on a product, you have an immense amount of data around that, all kinds of forms, warranties, so on and so forth, technologies, applications, applications in different market verticals. So it's really about getting that information together. And then the efficiency comes because Engineering can use that data, marketing, sales, customer service, parts supply, materials. So it's really this secure hub where a lot of people can tap in, find information, and create writing from that same data set.
RB: Going back to the first time you and I spoke, which I guess is about two years ago, the potential customers for these AI resources have needed a lot of teaching and explanation on this subject. Why do you think that's changed now?
AH: Well, AI is in the news every day of the week, whether we like it or not. And so when we spoke before, we spoke about machine learning, which is very complex. I still think it'll be years before that's a normal operation in manufacturing. However, With the newer tools, it's much easier. It's about common files, like we have PDFs and PowerPoint, Word, Excel. Those are everywhere. Those are our normal everyday business. And now the AI tools can really put that data to work.
RB: In your article, you emphasize that 80% of adoption challenges are about change management, not technology. Tell us, what is the change that has to be managed and why is it difficult?
AH: A lot of ways I look at the way our tool works. It's almost like when in the ’80s, us engineers had HP calculators and Casio calculators, and we used those for calculations. And when Excel came along, and it probably took many years, like AI, to really catch on. But when you realize you could put an equation in a cell and grab that cell and pull it across and pull it down and do all that math so fast, it was just a matter of time before that HP calculator slid into the drawer and you really never even turned it on again. So that change of habit, Do I grab the calculator or do I open up Excel?
So we're doing that now with writing, largely about what Amatrium is, we're the Excel analogy for writing, for making writing fast. And the better the data, which is your data, the better the answer. You’ve got great data plus a ingenious writer, you get tremendous, useful content out the other door. So it's a change, it's going about it differently. And when people are realizing when you start to embrace that change, and be patient with it, because change takes patience, you get tremendous upside.
RB: You also make the statement that AI can convert tribal knowledge into institutional knowledge. How can manufacturers practically identify and capture and validate what is tribal knowledge that can be converted?
AH: Let's say a foundry goes to the people that have been around for 30 years, there's a phrase, a funny phrase sliding around called the silver tsunami. That's gray-haired people retiring in the next 10 years or less. And so if you make a very intentional practice to document the best practices on how we view a new casting design, how we design tooling, source tooling, even thinking about sand and resins and the whole thing, all the way through from potential project coming in to delivering the casting. And you document the best practices around that. And you have your senior people document that, you may even interview a person, so you interview a person and you use an AI tool called Otter AI, you record that conversation. How do you think about this? How do you address this? How do you address that? Of course, you have them so they understand your goal, but really get all that stuff in written language. And then this company GPT has it all. It's tremendous. It's very searchable, and you're going to build on that in the future.
RB: You argue for company-specific secure GPT systems in contrast to or versus free public tools. Describe the risks of relying on open AI systems and tell us what the benefits are of the alternative, the secure systems.
AH: So us engineers that have done analysis, for example, finite element analysis, you'll hear the phrase all the time, garbage in, garbage out. If you don't have good data going in or, in the case of finite element analysis, it's load cases and did you properly load a bolted joint and a weld and all that stuff, you're not going to get a good, accurate answer on the other end.
So if I go to ChatGPT and I type in “what is the strongest cast aluminum?” it's going to say 356. Well, those of us in the aluminum world say we know it's 8206 or 8201, much stronger, much better. Well, why didn't they get it right? I don't know. Maybe they say 356 because it's the most common. It doesn't mean it's the strongest. So if you have your own data, and your own data points to various alloys and toughest, strongest, easiest to cast, whatever the characteristics are, I don't get that answer right. So good data, and the data that you trust, that you built your business around for all these years will produce a much better answer. So that's really the benefit of a secure in-house tool is great data will provide great answers.
RB: You offered the readers two interesting use cases involving foundries and equipment suppliers to foundries. So what priorities do you set for either of them to begin building their AI resources?
AH: I would say everybody starts small. So we have a client who's a multi-billion dollar client, they started very small. In their case, they're a manufacturer, so they'd be more on the equipment side. They started with marketing, and they started loading product data and market data. And these are PDFs, and started doing marketing, and they had a marketing goal of more marketing this year than the last two years combined, and no problem hitting that goal.
So, if you're in a foundry, you may say, we have a lot of paperwork around environmental health and safety. And so you load all your state and federal codes into the tool, and the document’s 1000 pages, that's okay. And then you load all your past history and you start to generate future reports based on your past reports.
Or in maintenance is another one. So you have all these, all this equipment, you’ve got melt equipment, that transfer ladles, you’ve got shakeout equipment, you’ve got heat-treated ovens. As you go through the boundary, it's CapEx intensive. So every piece of equipment comes with this thick manual that's hundreds of thousands of pages. Those could be loaded as PDFs, and you're quickly getting to how do you fix something, how do you do preventive maintenance, how do you do spares, all this stuff. So it's really a huge amount of information available immediately.
And when you do a search in a AmatriumGPT, it'll give you the paragraph. It'll give you the citation. You click on a citation. It's this manual, page 252. Here's where this answer came from. So it's really having all that information at your fingertips, not in the file cabinet, back in the back room.
RB: That's very, very compelling and almost irresistible to a lot of these operations. So let's look ahead. past the silver tsunami if we can. What are the new skills that operators and maintenance techs in these types of manufacturing operations will need once this new approach has been established?
AH: Again, it goes back to having the right tool for the job. So part of it will be, you know, the right AI type tools. So for example, with AmatriumGPT, we could compare to copilot. Copilot is on your computer, whether you ask for it or not. It came along with the Microsoft package. However, our customers compare and say, we can tell Amatrium is trained on engineering data. The answers are just better. Also, one of our customers sells a lot of products to the foundry business. And if you ask the question, if they were typing into their data set asking what products do we sell in aluminum foundries, if they didn't have some sort of tag on strontium, for example, it may not say it's used in aluminum applications to strengthen aluminum. So part of that would be adding these types of details to their data to make it more findable with the AI tools. Smart marketers are starting to do that too. They're testing the AI tools. Did they find my company? If not, tailor the wording so AI tools will find their company.
RB: I look forward to all of that. I hope I'm aware of it when it happens. But that seems like a good point on which to wind up. Andrew, thanks again for spending time with me today.
AH: You're welcome. Thank you.
About the Podcast
Great Question: A Manufacturing Podcast offers news and information for the people who make, store and move things and those who manage and maintain the facilities where that work gets done. Manufacturers from chemical producers to automakers to machine shops can listen for critical insights into the technologies, economic conditions and best practices that can influence how to best run facilities to reach operational excellence.
Listen to another episode and subscribe on your favorite podcast app
About the Author
Robert Brooks
Robert Brooks has been a business-to-business reporter, writer, editor, and columnist for more than 20 years, specializing in the primary metal and basic manufacturing industries. His work has covered a wide range of topics, including process technology, resource development, material selection, product design, workforce development, and industrial market strategies, among others. Currently, he specializes in subjects related to metal component and product design, development, and manufacturing — including castings, forgings, machined parts, and fabrications.
Brooks is a graduate of Kenyon College (B.A. English, Political Science) and Emory University (M.A. English.)