There’s no shortage of AI in manufacturing. There is, however, a shortage of AI that works when things get complicated—AI that can move the needle.

If you spend any time in industrial service environments, where assets go down and the fix is buried in five different systems and 4,000 pages of technical documentation, you already know this. The problems that actually cost real money and burn serious time aren’t the ones AI is solving well—yet.

The debate about whether AI will replace workers is a distraction from the more immediate question: Can AI make the knowledge these workers carry more accessible before they’re gone?

AI isn’t replacing the shop floor

There’s a persistent narrative that AI will replace experienced workers. That’s not what’s happening on the shop floor.

Some of the most valuable knowledge in manufacturing, such as pattern recognition, instinct, and physical awareness of equipment, doesn’t live in documentation. It comes from years of experience, and it’s simply not something AI can replicate.

What AI can do is reduce how often that level of experience is required. In practice, that means changing how often teams must rely on a small number of people to solve every complex problem.

Every service organization has one: The person whose phone rings before anyone opens a manual. The 25-year veteran who can diagnose a fault by sound, who knows which workaround works on the third-generation model, and who half the team depends on for anything that isn’t straightforward. That person is invaluable. They’re also a single point of failure.

By making knowledge easier to access and apply, AI enables less experienced technicians to solve problems that would have previously required a call to that person. It helps teams move faster without routing every complex issue through the same small group of experts.

The real problem is complexity

Where AI in manufacturing succeeds or fails almost always comes down to one thing: complexity.

Most AI systems work well in controlled environments. They can answer simple questions and automate repetitive tasks. That’s useful for productivity hacks, but in complex manufacturing environments, it misses the mark.

The real challenges show up when something goes wrong. A machine goes down. A technician can’t immediately diagnose the issue. A ticket gets escalated, but the answer isn’t in one place. Rather, it’s spread across manuals, schematics, service logs, and past tickets. Sometimes it depends on context that was never formally documented: what happened on the last shift, what workaround someone used, or what condition the equipment was in at the time. In many cases that knowledge lives with a handful of experienced technicians, not in a system.

Resolving that kind of problem requires connecting multiple sources of information and reasoning through them step by step.

There’s also a meaningful difference in the types of information involved. Historical service tickets and past case notes are typically text-based. They’re easily searchable, and most AI tools can handle that content reasonably well. But the documentation that actually tells you how a piece of equipment works, how to diagnose it, and how to fix it, looks very different.

Schematics, wiring diagrams, exploded parts views, functional block diagrams: This content is visual, highly structured, and dense. The complexity is the point.

An AI system is required to understand spatial relationships, component hierarchies, and contextual references across pages. Most general-purpose AI tools weren’t built for this. They convert visual content to text and then try to reason over it, which is where accuracy breaks down.

Fast answers are easy. Right answers aren’t.

A big part of the issue comes down to how several AI systems are designed. Most are optimized for speed and incentivized to provide their user with an answer, even when that answer can’t be dutifully verified. That’s mostly OK in low-risk, productivity-based scenarios. It’s entirely unacceptable on the shop floor.

In complex environments, one of the biggest challenges is something the industry has started to talk about more openly: hallucination, when AI systems generate responses that sound confident and technically correct but are partially or entirely wrong.

In manufacturing, that’s not just a minor flaw. It’s a real risk.

Most general-purpose AI tools top out around 40–60% accuracy on complex technical documentation. That might sound like a starting point. In practice, it means nearly every other answer is wrong, incomplete, or misleading. A confidently incorrect answer can lead to misdiagnosed issues, unnecessary part replacements, extended downtime, or even safety concerns. And just as important, it erodes trust.

Once technicians encounter a few answers that don’t hold up in real-world conditions, they stop relying on the system altogether.

This is why so many AI initiatives stall out during their pilot phase. They perform well in demos, where problems are simplified and outcomes are controlled. But they break down in real-world conditions, where problems are multistep, data are fragmented, and accuracy matters more than speed.

The real value: Making existing knowledge usable 

Where AI does work well is much more specific and practical.

It’s not about replacing technicians or automating everything end to end. It’s about making existing knowledge usable.

In most organizations, the answer to a problem already exists somewhere, buried in hundreds of pages of documentation, stored in past service tickets, or scattered across different systems.

The issue isn’t the absence of knowledge. It’s the time and effort required to find and apply it.

AI delivers value when it can bring those sources together and help people reason across them quickly: Not just retrieving information, but synthesizing it into a clear, usable answer tied to the specific equipment and issue.

That means a technician facing a fault code on a specific model gets a step-by-step resolution that refers to the actual schematic for that unit, not a generic answer that might apply to three different machines.

That’s what reduces time to resolution, improves first-time fix rates, and prevents every complex issue from being escalated to the same small group of experts.

The real risk isn’t AI. It’s knowledge loss.

The stakes are only getting higher.

As experienced employees retire, organizations aren’t just losing people; they’re losing the knowledge those people carry with them. With an estimated 2.1 million manufacturing jobs projected to go unfilled by 2030, that gap is only widening.

Historically, companies have tried to address this through documentation and training, but both have limits. Documentation is often incomplete or difficult to use, and capturing knowledge takes time from the same people responsible for keeping operations running.

AI doesn’t solve that problem on its own. It can’t create knowledge that doesn’t exist.

What it can do is make existing knowledge far more usable and ensure that new knowledge, generated through service and problem-solving, is captured and available the next time it’s needed.

For manufacturers, the opportunity isn’t to automate everything or replace expertise. It’s to make knowledge accessible at the moment of need, be it on the floor, during a failure, or simply when time is of the essence.

The real risk was never that AI would fail to live up to the hype. It’s that the person whose phone rings every time an asset goes down is about to retire, and everything they know is going with them.