Every manufacturing executive faces a common problem: experienced older technicians are putting in their retirement notice, but it’s becoming increasingly difficult to find their replacements.
It’s not specific to a particular geography, nor is it limited to an industry. Companies in industries from injection molding to chemical processing to chipmaking have the same report: tenure is down, skilled talent is scarce and trainers face the harsh reality that as soon as they invest in developing talent, a competitor poaches the newly skilled hire away, cutting the returns on apprenticeship.
The result is that when a robot stops, it stays down longer. When a PLC throws an error code at 2 a.m., hours of production are lost. Output numbers slip. Executives get upset. CFOs raise their eyebrows. Shareholders frown.
Automation, which is usually the answer for manufacturing labor problems, instead exacerbates the problem: automation reduces overall labor, but needs more highly skilled labor.
But could recent developments in an AI technology called Large Language Models (LLMs) revolutionize how manufacturers approach technical knowledge retention?
For some, it will be a critical tool. But others will face a competitive disadvantage.
What is this new type of AI? It’s different from the sorts of machine learning systems that manufacturers have used before, which are typically used for machine vision or data analytics.
It’s built from neurons trained on written human language, using billions of words in dozens of human languages. It turns out that human language reflects the way that humans think and reason, and so when researchers added enough neurons to them – hundreds of billions of them – they began, unexpectedly, to think and reason on their own.
When scaled to be big enough, the capabilities of this new type of AI (called a ‘transformer’) startled researchers. Without being explicitly trained to know law, it passed the bar exam. Without being trained to know chemistry, it got a 5 on the AP Chemistry test. On general reasoning ability, it placed at the 96th percentile of all high school seniors.
Ask it to write a program, and it produces a reasonably good draft. It’ll write business plans, emails, and proposals, too.
But ask it to fix a problem on a production line, and it fails. Ask it what the quality standards are for a product, and its response is generic and unhelpful. Ask it how to follow a process at a plant, and it won’t have the slightest clue what the user is asking for.
That’s what any manufacturing engineer who has tried out one of these models, such as the one powering the popular ChatGPT, finds when they try to apply it to their facility.
And why would it be any good? After all, it does not know how a particular facility’s machines work. It hasn’t trained on equipment manuals that aren’t publicly available, which is nearly all of them. It doesn’t know a company’s internal process documentation. Nor is it aware of what spare parts belong to a machine.
Being smart is not enough: to be useful, AI must be both smart and knowledgeable. Not that different from a human employee.
But other industries are showing how to use this new type of AI. McKinsey, the global management consulting company, wanted to empower their consultants with the institutional knowledge built up over decades of projects. And so, in August, they took a large neural network and plugged it into the vast library of work the firm had done over decades.
Then, when a consultant had a question about some obscure topic for a client, they could simply text the AI, and it would respond with every relevant piece of information the firm knew.
That consultant would learn and apply knowledge from another consultant who had left the firm years prior – and whom they never even knew.
It was a fundamental shift in how knowledge in organizations works. Before, a consultant would ask a more experienced colleague, or a subject matter expert within the firm. But the AI aggregated all the knowledge from all the firm’s experts, including those who had left the firm, and made it instantly available to any consultant at any time of day.
For anyone who’s ever watched Star Trek, that might sound like they created a centralized Borg hive-mind. It’s an apt metaphor.
McKinsey named their new AI ‘Lilli.’ Harvard economists conducted a study of a similar system at the Boston Consulting Group. The conclusion was that their consultants increased the quality of their work by 40%, while improving speed of output by 25%. The benefits were greatest among those consultants with the least experience: it made consultants’ performance more uniform, by bringing up the average closer to the highest performers.
Another study, conducted by Stanford economist Erik Brynjolfsson, found that a call center that handled technical questions from business customers achieved 35% increases in problem resolution speed for novice workers using an AI trained on call center chats.
Management consulting companies and call centers are not manufacturers. But they share a common challenge: how to make an organization’s tribal knowledge persistent and easily available to employees, so they can deliver results faster and more accurately.
In manufacturing, attempts to fix the problem of retaining tribal knowledge have been full of frustration. Process engineers, who wisely understand that a process can’t be trained if it’s not documented, invest their time to formalize production processes and then document them into standard operating procedures or playbooks.
The problem is that these procedures or playbooks sit in a binder. The binder sits in a cabinet that might as well be a tomb, for how often it’s opened. If they’re digital, they sit in a shared drive, and then almost everyone promptly forgets the file exists.
When a problem occurs, the answer is documented in the operating procedure. But for a mechanic trying to fix a servo drive problem late at night, or an electrician troubleshooting a temperature controller, it’s hard to find which operating procedure has the necessary information, or to remember that the procedure even exists.
It’s not just the standard operating procedures, either. The equipment manuals, which can be more than five hundred pages for a single system and have thorough troubleshooting guides, often go unread, and staff may not even know where they are.
Can manufacturers train this new transformer-type of AI to create a ‘hive-mind’ that knows every technical detail of every machine, every process document, and the maintenance history of every machine?
Early adopters are already moving on it. One example is McConkey Company, an injection molder and thermoformer in Washington State that makes agricultural products. Like many other manufacturers, they faced challenges in finding skilled workers, and process documents which got written but then filed away.
With turnover, there was also a problem of retaining institutional memory. An example: a PLC faulted, a mechanic troubleshot and then fixed it, but when the PLC threw the same fault again a year later, nobody knew about the prior fix. And so expensive hours were consumed re-troubleshooting the same problem.
So McConkey tried something new. The company used an AI platform to feed tens of thousands of pages of equipment manuals, process documents and twenty-five thousand maintenance records into a neural net with billions of neurons.
After ingestion was completed, employees could ask it how to set up a robot, or how to fix a machine error code, or what to do to resolve a quality problem, by chatting with the AI. With access to equipment manuals, process guides and historical work orders, the AI could fuse knowledge from many different sources into a single answer, providing step-by-step guides to fix the user’s problem. And the user could have an ongoing chat with the AI to clarify parts they didn’t understand.
As the management team found new blind spots of areas that were undocumented, they added more process guides and tips, which the AI learned from. And then it used them to help employees.
Many workers found that they preferred asking the AI, because they could get questions answered without having to find a manager or interrupting a coworker in the middle of another task. New employees loved it because they could ask any question, no matter how silly, and never be afraid of feeling like it was a “dumb question.”
How do these AI expert systems work under the hood?
The first step is turning all the documents provided to it into their underlying meaning. This is called a semantic meaning. ‘Semantic’ meaning refers to the concepts represented by words, rather than the text of the word. The words ‘dog’ and ‘canine’ look very different, and with a normal database, a search for one wouldn’t pull up the other. But with large language models’ understanding of language, they understand that these words are synonyms.
The semantic meaning is critical. It allows the system to later fetch relevant records by finding related concepts. If a user searches for “servo,” it knows that a “servo” is related in concept to terms such as “encoder” and “motor,” and that helps it pull relevant pieces of documentation.
Since semantic meaning is no longer relying on the text of a word, it works across different languages too. An AI can receive a question in Spanish, match it against English documents, and then respond, accurately, in Spanish. This capability was not intentionally programmed into large language models. It just happened. During training, it ‘read’ a lot of books and websites in different languages, and then it was suddenly able to effortlessly translate between every world language.
When a user asks a question, the semantic search system acts as a super-fast librarian that pulls relevant pages or paragraphs from the library of all a manufacturer’s equipment manuals, process documents, maintenance records, troubleshooting guides, quality reports and anything else available. This process is fast, usually clocking in at around a tenth of a second.
Then, this speedy librarian supplies these excerpts to the AI. These excerpts can come from many different sources: for example, it could be three paragraphs from a manual, two maintenance records, a page from a processing guide, an excerpt from a standard operating procedure, and a mention in a quality report.
The AI evaluates it, considers the user’s question, and then provides an answer to the user’s question based on the excerpts provided to it.
This approach has several advantages for manufacturers. It’s what AI engineers call “grounded”: it relies on concrete facts provided to it, which virtually eliminates the problem of hallucination. Hallucination happens when AIs invent facts out of thin air, usually because an AI is asked to provide answers but with no relevant information.
Using these excerpts provides another benefit: it allows for citations, so the operator can see the actual document snippets the AI used in its response, and, if they want to dig further, zoom to the relevant page in a manual or corresponding maintenance record.
But not all manufacturers will benefit equally from this new generation of AI. For some, it’ll help insulate from a tumultuous labor market. For others, they’ll find these systems to be of more limited benefit.
Here’s how to be prepared, and the three steps manufacturers need to take for a successful implementation.
#1 If it’s not written down, it’s not helping anyone
AI trains on text. So if there is no underlying documentation, a manufacturer will be unable to train an AI expert, and won’t be able to achieve productivity lifts. Manufacturers that don’t have a culture of documentation will fall behind.
That has always been true, but AI will significantly accelerate the importance of documentation.
There’s five major types of documentation that an AI can train on. Although there are no strict requirements for what format the documents should be in, AIs perform better when the dataset is properly labeled at the point of learning. This should be automated and should require no preparation from the manufacturer.
First, there are standard operating procedures. These instructions provide a playbook for how the manufacturer’s team executes its processes. Training an AI on this material gives it the ability to help people know how to do the processes that are unique to that manufacturer.
Second, there are machine manuals. They usually come from the equipment vendor, but most machines also have manuals for their subcomponents – for example, a servo controller may have its own manual. These subcomponent manuals can be even more extensive than the machine manual. Machine manuals are the greatest use for an AI to train on, because there tends to be a lot of content that the manufacturer doesn’t have to write themselves. A single machine and its subcomponents could have two thousand pages of documentation or more.
Third, there are troubleshooting guides. Troubleshooting guides are like operating procedures, but rather than outlining how a process ought to work, it helps workers resolve a problem when something isn’t working. These resources are invaluable to train an AI on, because employees are most likely to consult with the AI expert system when something isn’t working.
And finally, there is maintenance documentation. At some plants, maintenance can be informal. Most informally, an operator finds a technician when there’s a problem, the technician fixes the problem, and there is no record that there was ever a problem. This doesn’t give manufacturing managers much to work with to understand recurring issues, and it doesn’t give an AI much to learn from either. Slightly more formally, ticketing systems can track when problems occur, and whether they were fixed. These tend to be lightweight and only serve as task trackers.
But to benefit the most from an AI system, use a maintenance tracking system to document what problem occurs, and then have technicians document exactly what they did to fix it. That allows an AI to learn both from the problem and the resolution. Once this becomes standard procedure, the AI can have thousands of problems and resolutions to learn from – and to begin providing automated intelligent recommendations from the plant’s own history.
#2 Continually Train the AI
Lifelong learners accomplish more than those who stop after school. AIs are no different. Successful AI knowledge management systems treat the AI’s expertise as a long-term asset that they build up every day, rather than a single project that’s done after implementation.
There are two approaches to this. Both are important. The first is setting up organizational processes to continually train the AI, and the second is setting up automations.
Setting up organizational processes means capturing learnings as they happen. Once a learning moment happens – a tool gets damaged, a machine breaks because of an incorrect operating parameter – it’s easy to just say, “don’t do that again,” and then ignore it until the problem recurs.
Instead, manufacturers can set up a process to document what happened, why it happened, and what can be done in the future to prevent it. Then submit this to the AI training system (which the system should make as simple as a text box with a form) and it’ll be memorized — and brought up in the future. Designate Subject Matter Experts as those who have responsibility for continually providing training material for the AI. And then turn it into a KPI, and track how much the organization continually contributes to its AI.
The other approach is automation. If the organization is collecting valuable information elsewhere, then set up an integration to pipe that information to the AI. Where can that information be found? Look at anywhere data is stored, especially if it’s data that could be relevant to institutional memory.
Examples include maintenance recordkeeping, project bids or RFQs, quality assurance test documents, and production notes. If records are already being kept, and it’s ever useful to refer to them, then it’s likely useful for an AI to learn them to incorporate into its knowledge base.
#3 Make it easy for staff to use
Training an AI and keeping it up to date is only part of the battle. The AI also must be easy to access. Fortunately, people will want to use these intelligent agents, because they make peoples’ jobs easier by giving them a fast way to find documentation. But it still needs to be accessible.
There are two ways to do this for employees on the floor. One, set up a computer out on the production floor, keep the conversational agent open, and anyone can walk up to it and ask questions.
Second, consider putting the AI agent on tablets that employees can take with them to the machines, so they can use the AI to answer questions and pull reference diagrams right at the machine. Physical proximity to the system is critical for adoption.
Engineering and office staff will likely already have the system at their fingertips, so adoption will be easier, so long as the information they need is integrated into the system.
Looking toward the next decade, today’s labor issues are likely to get worse. The population is aging, and there will be fewer young people joining industry. As manufacturing automation advances, manufacturers will need them to be more skilled to contribute. That will make the manufacturing executives’ job a difficult one.
Finding ways to increase the skillset of workers and to retain the knowledge of those retiring will be critical. There will be no silver bullet. But the advent of new technologies, including those of artificially intelligent large language models, will be an important tool for manufacturers to retain their expertise and empower new workers to improve operational outcomes.
Derek Moeller is the founder of CognitionWorks, a company that designs natural language AI systems for manufacturers, and has over a decade of experience in the manufacturing sector.
