AI on the Factory Floor: Why Most Initiatives Fail and What Leaders Must Fix First (Growth Files Ep. 6)

Your plant just invested in an AI-powered predictive maintenance pilot. Six months later, it is still running on one line, but no one can clearly say whether it worked because success was never defined before the project started.

The vendor says the data looks promising. Your operators don’t trust it. Your CFO wants ROI numbers. And the rest of your equipment still runs on the same preventive maintenance schedule it’s been on for 15 years.

This isn’t a hypothetical. It’s happening right now across mid-size manufacturers in every sector, from automotive parts to pharmaceuticals to food processing. AI has become the technology that every executive is under pressure to adopt, but few organizations are prepared to implement it successfully.

In this episode of Growth Files, we sit down with Mike Nager, who has worked with Festo Didactic, Metz Connect, and Phoenix Contact at the intersection of workforce development, industrial automation, and AI adoption. He’s also the author of The Smart Student’s Guide to Smart Manufacturing. His perspective: the problem isn’t the technology. It’s the gap between buying it and knowing what to do with it.

Episode TL;DR

• Many AI failures in manufacturing stem from choosing the tool before defining the business case
• Many successful AI applications (facial recognition, chatbots, vision inspection) are no longer even recognized as AI, skewing failure statistics
• The average manufacturing workforce is unprepared, not just in skills, but in underlying data infrastructure and connectivity
• If you spend $1M on AI technology, budget an equal $1M for workforce development and organizational change
• Pilot programs work, but only when success criteria are defined before they start, you end up in “pilot purgatory”

In this conversation

• Why the “80% of AI projects fail” headline deserves more scrutiny than it gets
• Choosing tools before business cases is the most common and most expensive mistake
• Why do data infrastructure problems block AI before the algorithms even run
• What psychological safety has to do with successful AI implementation
• The workforce skills that will actually matter in manufacturing’s next chapter
• How to run pilot programs that scale instead of stalling

The Reality Check: Is AI Actually Failing in Manufacturing?

Sathish: There’s massive investment happening in manufacturing AI right now, but many initiatives fail. From your experience, what’s really going wrong?

Mike: The topic deserves nuance. You see a lot of headlines in the technical press saying 50% or 80% of AI projects fail in manufacturing. And you have to ask yourself two things: who’s publishing those headlines and why, and whether the premise is actually accurate.

Here’s what most people miss: there’s a lot of AI that’s been sliding under the radar for the last 10 years, successfully incorporated into organizations but no longer recognized as AI.

The AI that’s no longer called AI:

• Chatbots that answer customer calls? That’s AI. But we just call it “the phone system.”
• Facial recognition used every single day across facilities? That’s AI. We call it “security.”
• Inspection cameras with embedded algorithms that detect defects in real time? That’s AI. We call it “quality control.”

As soon as AI becomes successful, it gets renamed into something else and loses credit for being AI. So when someone says most AI projects fail, they’re ignoring the vast number of quiet successes that have already been absorbed into standard operations.

The failure statistics aren’t wrong. But they’re incomplete.

Most manufacturers don’t realize their AI investment is failing until the budget is already spent. This conversation breaks down how to spot the warning signs early and what to fix first.

▶ Listen to the full episode on Spotify to hear Mike’s complete breakdown of what separates AI pilots that scale from ones that stall.

Why AI Initiatives Actually Stall: The Two Friction Points

Sathish: So if it’s not the technology itself, where do things go wrong?

Mike: When you look at AI being employed in manufacturing, there are two friction points that are genuinely hard for firms to overcome.

Friction Point 1: Choosing the Tool Before the Business Case

AI has created enormous pressure on CEOs, presidents, and boards of directors to announce that they’re doing something with AI. There’s pressure to act. And organizations often don’t care what the action is, as long as they can say they have a strategy.

But there’s a real disconnect between having an AI tool and using it to perform an existing business function. There has to be recognition upfront of what problem you’re solving, in typical business language: cost savings, quality increases, better customer satisfaction.

That clarity is lacking in a lot of big AI projects.

Friction Point 2: The Skills Gap from Top to Bottom

Even when leadership formulates a solid plan, who implements it? Education and training institutions are struggling to catch up. They’re not producing enough graduates who can take AI from theory into practice.

So you have a problem at two levels: leadership that doesn’t always know how to frame the business case, and a workforce that hasn’t been equipped to execute the plan.

Mike’s analogy: Imagine you had a time machine and went back to 1890, to a factory running on steam engines. You bring with you a massive electrical control panel and plop it inside the factory. The workers come in and ask, “What is this thing?” You say, “It’s electricity. It’s a super powerful tool. It’s going to fix our business and make us rich.” They ask, “What can it do?” And you say, “We’re not sure.” They ask, “How much does it cost?” And you say, “We have this thing called watt-hours, and we’re going to charge you per watt-hour.” And they say, “What’s a watt-hour?”

That’s where we are with AI. It looks powerful. It is powerful. But most organizations are still figuring out how to actually grasp and implement it.

The Data Problem No One Wants to Talk About

Sathish: How big of a challenge is data quality and infrastructure in manufacturing AI?

Mike: We’re unprepared on the people side, but we’re also unprepared on the underlying data structure of the plant.

There’s a lot of manufacturing out there running production equipment as islands of automation without much connectivity between those islands. There’s also limited data harmonization. Different names, different tags used to describe the same data. And just trying to throw all of that into some AI program and have it make sense of it is probably not going to work.

What “getting your data house in order” actually means:

Until you can describe a process manually, organize all the data, and call everything by the same name, it’s very hard to automate a data system that doesn’t have a clear workflow.

Factories today throw out terabytes of data every day. Every sensor, every piece of equipment, every camera is producing data. But is it tagged properly? Can a human read it and understand it?

There’s a concept in industrial automation called the Unified Namespace (UNS), which is an attempt to standardize data from all the different producers and consumers inside a network. The idea is that data is structured in a way that almost self-explains itself when it appears: which plant, which city, which piece of equipment, which sensor the data came from, all inherent to its structure.

Once you have that, then tools including the large language models like ChatGPT and Claude might be able to do remarkable things with it. But the old computer science principle still applies: garbage in, garbage out. AI doesn’t erase that concept. The more organized you are, the better your results will be.

What an AI-Ready Manufacturing Workforce Actually Looks Like

Sathish: How unprepared is the average manufacturing workforce for AI?

Mike: Pretty unprepared. AI implementation is part of an overall digital transformation, and because it’s a technology that not many people knew about until recently, it certainly wasn’t being addressed in most engineering programs until very recently.

But here’s the nuance: the manufacturing workforce has already been exposed to AI. A lot of it has been distilled down and incorporated into discrete devices.

AI at the edge is already happening:

Inspection camera manufacturers have started embedding AI algorithms into their image feeds. The AI is being pushed further toward the edge, into specific pieces of equipment. When you peel that back, we’re not talking about large language models. We’re talking about machine learning: how can this camera count objects, determine what it’s looking at, or perform a quality check?

When you narrow the scope like that, things become more manageable. There’s a clear business goal for each use case, and there are relatively straightforward tools that can be implemented by a technician or engineer without a PhD in computer science.

This is going to continue. Just like how every device eventually got a Wi-Fi connection, both in personal and professional life, we’re going to see small kernels of AI and machine learning embedded inside more and more devices.

But the same PR problem returns: people might not recognize it as AI, so it doesn’t get credit for being a success.

AI in manufacturing isn’t failing because the technology doesn’t work. It’s failing because organizations buy the tool before they define the problem.

This conversation breaks down where AI actually delivers ROI on the plant floor and what has to change before it can.

▶ Listen to the full Podcast

The Talent Gap and Reshoring: Why This Problem Is About to Get Bigger

Sathish: Are students signing up for manufacturing and AI courses, or does it still need a push?

Mike: Manufacturing in the United States has had a bad reputation for a long time: dirty, dangerous, and dull. That reputation has stuck.

But when people learn that machine learning algorithms are part of the work, and that robots are going to handle the physical labor, it opens up a lot of doors. It changes the perception of what a manufacturing career actually looks like.

The timing matters. The U.S. has now decided that reshoring manufacturing is part of a national initiative, both for economic livelihood and national security. That means we’re going to need significantly more people with these skills than we currently have.

At Festo Didactic, we started building machine learning and AI curriculum about five years ago, incorporating it at the high school, community college, and university levels. We started with no-code programming to ease people into it without scaring them away with complicated math. Then we introduce complexity in a measured format as they progress.

But there’s always a lag. With any workforce development program, you’re talking a minimum two-year delay between when a program is installed and when you actually have people who’ve embedded that knowledge into their professional abilities. The pipeline is being built, but it’s not filled yet.

The ROI Mistake Manufacturers Keep Making

Sathish: What do manufacturers misunderstand about ROI when investing in AI?

Mike: It comes back to the ROI not being clearly defined before a project starts. If you don’t know where you’re going, how do you know if you’re lost?

The goal can be modest. It doesn’t have to be some transformational moonshot. It can be a 5% or 10% reduction in waste, or time savings on a specific process. That would be a very nice win, and critically, it should be measurable.

The problem is that many plant floors don’t have systems in place that can tabulate those metrics so you know whether you’ve met the goal. There are platforms emerging that can take tens of thousands of sensor data points and calculate things like true OEE (Overall Equipment Effectiveness) or takt time for a production run. But you have to have those measurement concepts in place before you start.

What to measure depends on the use case, but two areas consistently show strong ROI:

Quality control: Especially in sectors like semiconductors or electronics where people are looking through microscopes to check solder joints or wafer integrity. Machine learning-based inspection is already proving itself here.

Predictive maintenance: Instead of looking at product quality, you’re monitoring the health of equipment and trying to eliminate unplanned downtime while avoiding the cost of routine preventive maintenance done purely on a time interval. Preventive maintenance keeps machines running, but it’s expensive because you’re overpaying to maintain them on a schedule rather than based on actual condition.

Real-world example: Festo Automation’s platform, Festo AX, lives on the plant floor taking information from sensors and actuators at a very fine level. It monitors something as simple as how fast a cylinder piston extends and retracts. Depending on the timing and acceleration, you can judge the health of that component and get warnings like “this cylinder is at 85% of its lifespan and should be swapped out in the next two weeks.”

Most people would never think of that as AI. But it’s exactly the kind of targeted, business-case-driven application that delivers real ROI.

The $1M Rule: Match Your Tech Investment with People Investment

Sathish: What’s the biggest mistake leadership teams make with AI adoption?

Mike: You have to get buy-in from the organization. That’s a leadership skill. How do you bring people into a shared vision of what AI is going to bring?

The very first thing you have to address is the elephant in the room: fear.

There’s a lot of fear about automation, both robotics and AI. People think: “Am I going to implement this system and then lose my job? I still need my job. I have kids in college.”

You have to provide emotional and psychological safety first. It’s like Maslow’s hierarchy of needs. At the very bottom of the pyramid is safety. Nothing else is going to happen unless people feel secure.

If AI is going to change a job role, which it inevitably will, you need upskilling and reskilling programs in place to provide that psychological safety to the people on the plant floor.

Mike’s rule of thumb: If you’re implementing a million-dollar AI system, you should have about a million dollars also set aside for workforce development and organizational change. By the time you bring in consultants, training, implementation support, and reskilling programs, you’ll need roughly an equal investment in the human side.

If you just buy the cool shiny technology and don’t invest in people, that’s where failure happens. People rush into the technology too fast without making an equal investment in organizational change.

Regulated Industries: A Warning from the FDA

Mike: Some industries have to be especially careful. In regulated sectors like aerospace, food and drugs, and pharmaceuticals, there’s a whole additional layer of scrutiny.

I just read this morning that in April 2025, the FDA issued its first warning to a pharmaceutical company for misusing AI in its production runs. The company used AI to conduct a manufacturing operation but didn’t ensure regulatory compliance. Their excuse was that the AI didn’t tell them. The FDA said that’s no excuse.

This was the first instance of its kind and marks a significant regulatory action regarding AI in drug production. In regulated industries, you have to be extremely careful about how AI is applied and documented.

Pilot Programs: How to Avoid Pilot Purgatory

Sathish: Should manufacturers start with pilot programs or go straight to full-scale implementation?

Mike: Pilot programs are probably the easiest way to go. But they have to be well-constructed.

There’s a concept called pilot purgatory: you run a pilot and it never becomes more than just a pilot. It just sits there, neither scaling nor shutting down.

How to structure a pilot that actually scales:

Before the pilot starts, define in writing: what does success look like? Be specific.

Example: “We’re going to install a vibration-sensing AI system for predictive maintenance on this production line. We expect 20% fewer unplanned stoppages over the pilot period.”

Then the pilot either hits the target or it doesn’t. If it works, you have a clear case for scaling it to every similar machine. If it partially works, say a 12% reduction instead of 20%, you can evaluate whether that ROI is still worthwhile. Maybe 12% is good enough to justify the investment.

But if you don’t define success criteria upfront, the pilot can technically succeed and no one knows. That’s how you end up stuck.

The Skills That Will Define Manufacturing Jobs in the Next Five Years

Sathish: What skills will be in demand among manufacturing professionals?

Mike: Here’s a funny thing: soft skills are making a huge reemergence in importance in manufacturing.

Why? Because we’re taking the job of being a robot out of a person’s job.

For a long time, we hired people as living robots. “What’s my job? My job is to pick up these bricks from this place and put them into this other place all day long.” Or: “My job is to take purchase orders written by customers and retype them into our order entry system.” We treated people as robots because the technology didn’t exist to automate those tasks.

When you remove that robotic work, what’s left? The skills that machines still can’t do well:

• Teamwork and communication: More collaboration across functions, not less
• Cultural awareness: Many manufacturers run multinational operations with diverse workforces
• Decision-making at every level: The job is increasingly to make decisions, not perform repetitive actions
• Leadership from the floor: Not just from management

The CNC operator example: People who ran CNC machines were traditionally called “operators.” The job was defined as placing metal into the machine, closing the door, and pressing a button. The metric was pieces per hour.

Now, robots can feed the machine. The machine can turn itself on. So what is the “operator”? Maybe the role becomes a “production specialist” whose job is to ensure all machinery is performing the right quality and quantity of work on their shift. That means reading screens, interpreting sensor data, watching for warning lights, and making judgment calls before small problems become big ones.

These jobs are changing. But they require the next level up in thinking, which most people are capable of. It just was never required of them before.

Where Manufacturers Should NOT Invest in AI

Sathish: Where are manufacturers wasting money on AI right now?

Mike: I wouldn’t presume to tell every company where not to invest. But the principle is straightforward: don’t invest if you don’t know what the outcome should be or could be.

There’s no point in it, because even if the AI is successful, you won’t realize it without metrics in place. And if it fails, you won’t know why or what to change.

What to do instead:

Get input from vendors, IT suppliers, and consulting firms. External perspectives can be very useful because every organization has blind spots, things people don’t talk about consciously or unconsciously. Someone who has seen a hundred similar operations can walk in and immediately spot inefficiencies that have become invisible to your team.

It’s critical to have an early win to build confidence. In this rapidly changing world, if your first AI project goes wrong and everyone gets scared to try the next thing, 18 months goes by. And 18 months in the AI world is like 20 years in the industrial automation world.

You need a balance: push the envelope and be willing to experiment, but tie every experiment to something tangible that can be measured.

What the Next Two to Three Years Look Like

Sathish: Where do you think manufacturing will be in two to three years with AI?

Mike: I think it’ll look pretty much the way it does today. What you’ll find is that we used to talk about “islands of automation.” I think we’re going to see “islands of artificial intelligence.”

Maybe there’ll be a huge success in processing purchase orders from customers and getting them into the ERP system automatically. Maybe predictive maintenance, something we’ve talked about for 20 years but has been hard to implement, will finally reach its potential because the tools and platforms are now good enough.

We’ll have to see. But it’s going to be interesting no matter what happens.

What B2B Manufacturers Must Do Now

AI in manufacturing isn’t failing because the technology is immature. It’s failing because organizations are deploying tools without defined business cases, into environments with fragmented data, without investing in the people who have to make it work.

The manufacturers who win won’t be the ones with the most advanced AI. They’ll be the ones who match technology investment with organizational readiness.

Here’s where to start:

1. Define the business case before choosing the tool

What are you trying to improve? Cost reduction, quality, throughput, uptime? Set a specific, measurable target. If you can’t articulate the expected outcome in one sentence, you’re not ready to buy.

2. Audit your data infrastructure

Are your production systems connected or running as islands? Is your data tagged, harmonized, and structured so that AI tools can actually use it? Investigate the Unified Namespace (UNS) concept for standardizing data across your plant.

3. Budget equally for people and technology

If you’re spending $1M on AI systems, allocate an equal amount for workforce development, change management, and organizational restructuring. The technology is only as good as the people implementing and using it.

4. Address fear and psychological safety head-on

Your teams are worried about being replaced. Acknowledge it directly. Create upskilling and reskilling programs. Show people how AI changes their roles rather than eliminates them.

5. Run well-defined pilots, not open-ended experiments

Define success criteria before the pilot starts. Measure results against those criteria. Decide in advance: what result means we scale, what result means we iterate, and what result means we stop.

6. Get external eyes on your operation

Bring in outside experts who have seen hundreds of similar facilities. Your team’s blind spots are someone else’s obvious opportunities. Compressed air audits, predictive maintenance assessments, and cybersecurity evaluations are all high-ROI starting points.

Ready to Understand How AI Answer Engines Perceive Your Manufacturing Brand?

While you’re optimizing operations on the plant floor, AI answer engines are shaping how your buyers discover and evaluate you before they ever talk to your sales team.

We run prompt-based AEO testing across your product families to show you:

• Which answer engines cite you vs. competitors
• Where citations come from (your site, distributors, third parties)
• What blocks visibility across your content ecosystem
• Which high-impact pages to fix first for maximum citation lift

Get your AEO visibility snapshot →

We’ll identify the top 10 revenue pages to optimize and show you exactly where you’re losing ground to competitors inside AI answers.

Keep the Conversation Going

This is Episode 4 of Growth Files by CommerceShop, inside stories and strategies from manufacturing and B2B leaders navigating the shift to AI.

Also Watch ←

• Episode 1: AEO & GEO in B2B Manufacturing
• Episode 2: Building Tomorrow’s Manufacturing Workforce