AI in Manufacturing, Revisited: Where It’s Paying Off on the Floor

Two Years Later

In December 2024, Tory Bjorklund wrote about how AI was transforming manufacturing. He covered the different types of AI, the risks worth taking seriously, the use cases with the most potential for return, and five steps manufacturers should take to get ready. It was written for leaders who were still deciding whether AI belonged in their operations at all.

A lot has changed in two years. Generative AI tools have become far more capable and accessible, and most of the manufacturing leaders we talk to have stopped asking whether to use AI. Now the question is where it will actually pay off, and how to get there without joining the long list of AI projects that never met expectations.

That makes it a good time to revisit the original article. Much of the advice has held up well. But one use case wasn’t on the list, and it’s one of the most practical opportunities we see for AI in manufacturing today: helping the people on the plant floor fix problems faster.

What Held Up From 2024

Tory’s five preparation steps were build data maturity, automate processes, start small and scale fast, upskill your workforce, and manage organizational change. Looking back, four of them aged almost perfectly, and the fifth has grown in a direction we didn’t fully anticipate.

Start small, scale fast is still the best AI advice we can give. Organizations that prove value on one well-defined problem tend to get much further than those launching enterprise-wide AI initiatives. The project we’ll describe below started as a focused proof of concept for exactly that reason.

Upskilling and change management have held up too. Tory argued that AI would change jobs more than it would replace them, and that’s what we see in practice. The AI tools that get adopted are the ones that make people better at work they already do.

Process automation remains a prerequisite for most planning and forecasting use cases. AI can’t optimize a production schedule that lives in a spreadsheet.

Data maturity is still the biggest single factor in whether an AI project succeeds, but what counts as “data” has widened. In 2024, the conversation was mostly about structured data: sensor readings, ERP transactions, inventory records. Generative AI can work with unstructured content like equipment manuals, work instructions, and technicians’ repair notes. That means your documentation is now data, and its quality matters just as much as the numbers in your ERP.

That shift is what opened the door to the use case we want to focus on.

The Use Case Missing From the List

The original article highlighted four use cases: predictive maintenance, quality control, demand and supply chain planning, and raw material price forecasting. All four are still valuable. But they share a pattern. They run on machine or transaction data, and they mostly serve planners, engineers, and executives.

None of them address what happens when a line goes down and someone has to figure out why.

Picture how that usually plays out. An operator calls a technician. The technician checks the binder by the machine, searches the shared drive for the right manual, and tries to remember whether this has happened before. If it’s a problem they haven’t seen, they call the one senior tech who probably has. If it’s second shift, that person might be home. Every one of those steps adds minutes while the line sits idle and product isn’t moving.

In my recent post on tribal knowledge, I wrote about how expertise that lives in one person’s head becomes a ceiling on growth. On the plant floor, that same problem shows up as downtime. What makes it frustrating is that the answer often already exists. It’s on page 212 of an OEM manual, in a work order from 18 months ago, or in an SOP nobody bothered to dig out and check. The challenge isn’t missing knowledge but finding it under pressure, and that’s a problem generative AI is well suited to solve.

Inside the Troubleshooting Assistant

We’re currently working with a packaging manufacturer on a proof of concept for an AI troubleshooting assistant built on a technology partner’s platform. Victoria Fide’s role has been helping the client select this use case, implementing the assistant using the client’s own documentation, and leading training for the people who will use it.

The assistant is designed to be used from a tablet or phone right at the line. An operator or technician describes the problem in plain language, the way they’d explain it to a coworker, and the assistant walks them through likely causes and next steps.

It draws on three sources of content:

• OEM equipment manuals, which explain how each machine is supposed to work and how the manufacturer recommends troubleshooting it.
• The plant’s SOPs and work instructions, which reflect how this particular operation runs its equipment.
• Historical work orders and maintenance logs, which may be the most valuable source of the three. A manual tells you what could go wrong. Work orders tell you what has gone wrong with this specific machine in the past and what fixed it.

In the original article, Tory flagged hallucinations and accuracy as major risks of using AI, and two guardrails address those concerns directly. First, every answer cites the source document it came from, so a technician can check the manual page or past work order before acting. Second, technicians can flag answers that are wrong or unhelpful, so those answers can be reviewed and the underlying content corrected. Over time, the assistant should improve as the people using it engage with it.

The project is still at the proof-of-concept stage, but based on what we’ve seen so far, we expect the assistant could help technicians diagnose issues, complete repairs, and get lines running again roughly twice as fast as they do today. The next phase will put that estimate to the test on the floor.

Transformation is not easy, but it doesn’t have to be impossible. Take control of your project’s success today and schedule a free 30-minute consultation to find out how Victoria Fide can equip you for transformational success.

What We’re Learning Along the Way

The technology tends to be the first thing people ask about. Honestly, it hasn’t been the hardest part. Here’s what stands out so far as the assistant moves toward the floor.

The content is the project. Most of the effort goes into gathering, cleaning, and organizing the source material, not into the AI itself. Manuals are sometimes outdated or missing. SOPs don’t always match how the work is actually done. Some work orders say little more than “fixed it.” An assistant built on that content will give answers that are just as thin, so preparing the content is where data maturity, Tory’s first step, really shows up for a project like this. We also expect a side benefit: once technicians see their work orders being put to use, they’ll have more reason to write better ones.

Trust is earned one answer at a time. Experienced technicians are skeptical of a tool that claims to know their equipment, and they should be. That’s why source citations matter so much. Instead of a black box giving orders, the assistant becomes a faster way to find something the technician can verify for themselves. The ability to flag bad answers matters for the same reason, because it gives technicians a voice in how the tool improves.

Design for the floor, not the office. A tool that works well at a desk can fall apart next to a running line. Answers need to be short, practical, and organized as steps someone can follow while working on a machine, and the tool has to be available on a device that’s already in the technician’s hand.

It should change the job, not replace it. Newer technicians should be able to troubleshoot with more confidence instead of waiting for help. Senior technicians should get interrupted less, and over time their hard-won knowledge will make its way into the work order history where everyone can benefit from it. That’s why training and change management are part of the plan from the start: involving technicians early, positioning the assistant as a tool that supports their judgment, and setting realistic expectations about what it can and can’t do.

Signs You Have a Knowledge Problem AI Could Solve

Line troubleshooting is one example, but the underlying problem shows up across nearly every manufacturing operation. Knowledge exists, but it isn’t reachable at the moment someone needs it. The same approach can help customer service reps answering product spec questions, quality inspectors checking tolerances, or planners working through scheduling exceptions. If several of these sound familiar, you may have a good candidate for an AI assistant:

• The same handful of questions reach the same few people every week, usually through a phone call or a tap on the shoulder.
• The answer is usually written down somewhere, but finding it takes longer than asking someone who already knows.
• Work slows or stops while someone waits for a response, whether that’s a line, a customer quote, or an order on hold.
• Answers come noticeably slower on second shift, on weekends, or at satellite sites than they do on first shift at the main plant.
• Putting together a single answer means checking several places, such as shared drives, binders, ERP notes, and old email threads.
• Ask two experienced people the same question, and you’ll often get two different answers.
• Useful knowledge gets captured after the fact in places nobody goes back to read, like work order notes, case histories, or meeting recaps.
• When a key person is out, certain decisions simply wait until they’re back.

You don’t need to solve all of these at once. Start with the one where waiting costs the most, whether that’s downtime, delayed shipments, or frustrated customers. That’s usually where a focused pilot will show value fastest.

Let’s Find Where AI Pays Off in Your Operation

Two years ago, the biggest question manufacturers had about AI was whether it was ready for them. Today, the better question is whether they’re ready for it, and which problems are worth solving first. Victoria Fide’s AI Business Strategy services help manufacturers identify the AI use cases worth pursuing, prepare the processes and content behind them, and implement tools that people on the floor actually want to use. If downtime and scattered knowledge are holding your operation back, let’s talk about where AI could help.