By Derek Moeller, president, CognitionWorks
Talk with the team at Fabrik Plastics, based in McHenry, Illinois, about its generative AI initiative, and one topic keeps coming up: speed.
“We can sit in a meeting now, and a question comes up, and we have the ability to ask that question using AI and get the results right in that meeting without reports, without opening IQMS, without going through folders,” said Ken Bennick, vice president of engineering and tooling at Fabrik.
Fabrik connected its ERP system, IQMS, to SprocketAI – an AI agent with a built-in data warehouse.
Generative AI famously understands regular human conversation. But it also understands computer code. And it’s an excellent translator between the two. That means it can reason over large quantities of data by translating an executive’s request into a series of programmed queries.
Traditional chatbots struggle with large quantities of data because uploaded data fills their “context,” or working memory; not only does this limit how much analysis they can process, but it also tends to suffer in intelligence as their context fills.
By giving AI systems the ability to run their own data analysis code against a data warehouse, the AI can reason across millions of rows of data, combining everything from bills of material, sensor data, production schedules and scrap records to customer invoices and supplier data.
Fabrik’s President Don Hardin provided an example: “Recently we were in a meeting, doing a deep dive on our scrap on a part. We were looking at how it has changed month to month over the last six years we’ve been molding it. And the AI’s graphical visualization of the data is incredible compared to just having a report out of Crystal Reports. We were sitting there in the meeting getting the scrap percentage charted out back to us over years of information, and being able to have the data in front of us that we can tell why and when the changes occurred, and we were able to pinpoint those dates in relation to those changes.”
Nate Harvey, operations manager at i2-tech located in West Des Moines, Iowa, agreed. “We’re able to give everyone a look into what’s happening behind the scenes,” he said, referring to the AI’s ability to query data from its installation of IQMS. Analyzing its enterprise data with AI allows a broader group of people to make data-driven decisions.
Fabrik and i2-tech’s success in deploying generative AI points to an essential element of AI strategy: integrating it with a company’s systems.
Opening the Box
Conventional chatbots have limited integration with manufacturers’ data systems. The models are intelligent, but they’re not connected. They lack access to all the contextual knowledge a company has built up over decades. They’re a brain in a sealed box.
Intelligence alone isn’t enough. Hiring a young, inexperienced but bright employee is not a bet on immediate productivity. It’s a bet on the individual’s ability to combine intelligence with the ability to learn how the business works. If that smart new employee had no access to the company’s data or files, the employee’s intelligence largely would go to waste because there’s nothing for it to process.
AI is no different, and it’s why many companies purchase off-the-shelf chatbots but then find disappointing staff engagement. It’s not that employees aren’t AI-forward enough or that if they attend one more seminar, they’ll suddenly “get it.”
It’s that if the AI isn’t integrated into the systems the team uses, it’s not going to do useful work for them. They don’t need a brain in a sealed box. They need it to be connected to the company’s data and systems, including its ERP, maintenance system, quality system and network-shared drives. That’s when it starts saving time. Employees use systems that save them time, and don’t use those that don’t.
Mining Tribal Knowledge
“It goes back to accessibility of information,” said Hardin. Fabrik connected its AI to its shared drives. It annotated, categorized and learned millions of pages of content in its memory store. “A lot of our tribal knowledge is captured within our drives: Molding machine manuals, mountains of articles and secondary brochures, processing guides, RJG studies.”
Fabrik’s AI also mined all of its maintenance records from its maintenance management system. “The real heavy value is in all the work orders we’ve done in IQMS over the years. Now we type a question in, and it grabs the 10 work orders from press 12 when we had heater band issues. And here’s the work center notes where our process technicians and supervisors had issues with heater bands on that press,” said Hardin.
Its AI researches the content, finds patterns that match the current problem and uses them to recommend corrective action steps based on what’s happened in the past.
Hardin explains with a real-world example: “So a process engineer on second shift says he is having trouble with random short shots on a given polycarbonate job.” The AI finds where that’s happened before in work center notes and maintenance logs, cross-checks it with processing guides from SABIC and other reference sources, and generates a troubleshooting guide on the fly for that specific part, machine and problem.
That model of employee-AI interaction, called employee AI augmentation, is proving to be the winning path for AI adoption. When large language models initially were commercialized, some believed in a different approach – employee replacement. The idea was that an employer
could look at a single person’s job and have an AI replace it in totality.
While it’s obvious that this doesn’t work for roles with a physical component, like a technician, it’s also been increasingly clear that this isn’t the right model for pure knowledge work either.
The “Jagged Frontier”
Erik Brynjolfsson, a Stanford professor who has extensively studied digital technology’s impact on the workplace, co-authored a paper that explains why. The paper found that jobs are not made up of a single task repetitively performed. They are made up of bundles of different tasks that the employee applies skills to perform.
That combines with the idea of the “Jagged Frontier,” a term coined by Wharton Professor Ethan Mollick to describe AI capabilities. AIs are not humans, and their capabilities do not neatly map to human capabilities. Instead, they perform at super-human levels on some tasks, while performing worse than a toddler on others. Today, there is no human in the world who can beat the latest AIs in programming competitions, and no human chess player who comes close to AI performance. But a toddler is more dexterous at manipulating small, physical objects in a dynamic environment. That’s what Mollick means by the Jagged Frontier – there is not an even frontier of capabilities, but a surprising landscape of extreme performance mixed with pockets of near-total inability.
Moreover, this Jagged Frontier is becoming even more jagged. The current path of increasing large language model (LLM) capability is due to heavy investment in a technique called reinforcement learning. Reinforcement learning means that researchers give a carrot or a stick to a model during its training – the carrot when it performs a task correctly, the stick when it doesn’t. This only works for tasks that are verifiable; that is, if it shows whether the AI’s answer is correct, it is known whether to offer the carrot or the stick.
It can be verified whether two and two equals four. It can be verified whether a piece of computer code works. But it cannot be verified whether a novel stirs the soul. That’s why AIs have shown incredible gains in writing software and doing mathematics, but have shown little gain in writing long-form content. It’s why AI “slop” writing is recognized and associated with lesser value.
Long, content-light internal emails are a real downside risk of AI use, and the Jagged Frontier explains why. LLMs can produce the “form” of communication without the substance, because writing with substance is not as easily verifiable with the reinforcement learning techniques used in their training.
In contrast, AIs can use math and code to improve quantitative reasoning at a company. They’re especially competent when they work in a loop, writing code, getting quantitative feedback and iterating on an algorithm. Give it access to an order book, production history, machines and constraints, and the AI will work in a loop, developing an algorithm to cut mold changeovers while optimizing the capital cost of inventory.
With access to historical quotation, part, production and scrap data, along with the bill of materials, an AI can build a model to reasonably estimate the cost of a newly quoted part based on actual company experience with similar parts in the past. By holding out a test group of already-produced parts and grading the AI’s performance in a loop, the AI can iteratively optimize its approach in an iterative loop until it produces good schedules.
Connected to historical production records, the AI can analyze capacity loading and identify which tonnage class of machine is most lightly loaded. It then can build a model that applies marginal cost analysis and demand-driven machine-rate pricing. This helps a processor win business with aggressive quotes on lightly filled tonnage classes while avoiding margin erosion on high-utilization presses.
Those examples all use strengths of AI – the spikes of capability in the Jagged Frontier. That frontier’s unevenness means that jobs aren’t replaced. They are transformed. The AI can be superhuman at certain tasks but outclassed by humans in others.
To become an AI-native manufacturer that outperforms the industry, a company must accomplish two initiatives:
- Build a single, cohesive AI into all its core systems infrastructure.
- Redefine its job roles so that AI-augmented employees can use that integrated AI to accomplish more, work faster and with better quality.
Derek Moeller is the co-founder and president of CognitionWorks, a company that provides generative AI transformation for the manufacturing industry using SprocketAI, its AI tribal knowledge and analytics platform. Prior to founding CognitionWorks, he was the president and owner of an injection molding and extrusion company, and before that, the founder of a medical media company. He studied economics at Northwestern University.
More information: www.cognitionworks.ai