New technology is tied to a real operating question: how do we improve production without learning every lesson the expensive way? Pick a problem that is already costing the business money. Then ask what has to change for the problem to improve.
getty
A check-engine light doesn’t fix the car. It tells you something under the hood needs attention. Maybe the problem’s small. Maybe it’s the first signal of a major failure that’s been building for months. Either way, once the light comes on, you know you have to do something.
Manufacturers should think about AI the same way. The technology can predict machine failures, improve quality and speed up production. But first, it’ll likely turn on the factory’s check-engine light, showing leaders where the operation is messy, fragile or running on workarounds.
That might be painful and expensive but it’s necessary. AI won’t fix manufacturing until manufacturers fix the work. And this is where many companies are getting stuck.
MIT NANDA’s 2025 GenAI Divide report found that 95% of enterprise generative AI pilots showed no measurable profit-and-loss impact. And manufacturing has an even more basic problem. A full 70% of manufacturers still collect data manually. That means many companies are trying to add advanced intelligence to operations still running on handwritten notes, side spreadsheets, and individual memory.
Here are six places that need attention before AI can deliver real value.
1. Start With a Problem Worth Fixing
BMW’s virtual factory work started with a practical production problem: how to improve the plant before disrupting the real line. The company used AI and simulation to create digital twins of its plants, allowing teams to test layouts, robotics and logistics before making physical changes.
That’s the right starting point. New technology is tied to a real operating question: how do we improve production without learning every lesson the expensive way? Pick a problem that is already costing the business money. Then ask what has to change for the problem to improve.
A good AI use case starts with operating pain, a clear owner and a result the company can measure. Otherwise, the project becomes another technology experiment looking for a reason to exist.
2. Get One Version of the Truth
AI is only useful if the operation has a shared view of what is happening. Bosch’s Bamberg plant shows what that looks like. Cameras, sensors and test stations feed production data into AI-supported analysis. Bosch says the system processes about 1 million data messages in 24 hours and helps spot deviations before faulty parts leave the line.
That’s only possible because the company has done the harder operational work first. The data has a place to go. The standards are clear. People can trust the information.
Most manufacturers don’t have that. They’re still living with multiple versions of the truth. One number in the ERP. Another in a spreadsheet. A third in someone’s head. AI will not reconcile that by magic. But it will force the company to decide which version is real. Bad data doesn’t become good because AI touches it. It just becomes bad data moving faster.
3. Fix the Breaks Between Teams
Lenovo’s AI-powered scheduling work is a good example because the problem was bigger than building a better schedule. It was getting the right information across the business fast enough to keep production moving.
Lenovo says the system cut schedule planning from two hours to two minutes and increased production volume by 19%. That kind of scheduling only works when teams are connected. Sales has to know what production can actually make. Purchasing has to know what materials are missing. Production has to know when priorities change.
AI can help make those decisions faster. But it can’t fix a business where every team is working from a different version of the plan. A factory can’t run this way, no matter how smart the software is.
4. Fix the Process Before You Automate It
Automation has a way of making bad habits permanent. That’s why manufacturers should pause before putting AI on top of a process everyone already knows is frustrating, slow or inconsistent. Speed is useful only when the work itself makes sense.
Before automating a process, ask the uncomfortable question first: why does the work happen this way? The answer is often some mix of old software, unclear ownership and workarounds that became normal because people got used to living with them.
AI should not become expensive duct tape for broken workflows. The goal should be to redesign the work so better decisions become easier to repeat.
5. Bring in the People Who Know the Shop Floor
AI projects built too far from the shop floor usually miss how the job actually gets done. The people doing the job know the difference between the official process and the real one. They know which data is wrong, which alerts people ignore and which “temporary” process quietly became standard practice.
Bring them in before the tool gets built. No AI project needs a giant committee. But every serious one needs people close enough to the work to know whether the recommendation is useful, realistic and safe to act on.
The model must learn from reality, not the version of reality leaders wish they had.
6. Treat AI Like Continuous Improvement
Continuous improvement is where manufacturers excel. They know how to look at a process, find what’s slowing it down, fix what’s broken, measure the result and keep improving.
AI should be treated the same way. AI has to earn its place in the factory. It has to solve a real problem, work for the people doing the job and make the operation better in a way people can see.
That is how AI becomes useful inside a factory. It becomes part of how the company solves problems, not a separate pilot living off to the side.
For many manufacturers, the first step will feel uncomfortable. AI will turn on the check-engine light. It will show where the work is confusing, where the numbers can’t be trusted, and where things are broken.
That’s where the repair starts. Before AI helps a factory run smarter, it needs to show leaders what to fix under the hood.