Most manufacturers who think they have an AI problem actually have a data problem. Getting the right model is often the easier part. The harder part is the data the model needs, and in most plants that data is missing: machine readings that were never captured, maintenance history sitting in a technician's notebook, and an ERP that cannot tell you which press produced which batch.
That is why so many pilots for AI in manufacturing look convincing in a demo and then quietly stall. Gartner predicts that through 2026, organizations will abandon 60% of AI projects that lack AI ready data. Factories are especially exposed, because they were built to make parts, not to record what happens while the parts are being made.
So when people ask what the first step to AI in manufacturing is, the honest answer has little to do with models or vendors. It is digitization. Digitization in manufacturing means recording what happens on the shop floor as digital data and connecting it across your systems, so a model has something reliable to learn from.
AI Readiness in Manufacturing Is a Data Readiness Question
Most AI in manufacturing learns from past examples. A predictive maintenance tool, for instance, gets good at spotting failures by studying what a machine's readings looked like in the weeks before earlier breakdowns. For that to work, two things have to be true. It needs a solid history of machine readings matched to what actually happened, and the data it sees day to day has to be recorded the same way as that history. If either is missing, its predictions get less reliable, and often nobody notices until one turns out badly wrong.
Most plants struggle with both, and old systems are a big reason why. In Kaufman Rossin's 2026 research, 55% of mid market manufacturers said connecting their legacy systems was their single biggest obstacle to AI. Across the mid market as a whole, the figure was 41%.
There are two less obvious problems as well. First, most plant data was collected to answer someone's question at the time, like a reason code for the monthly downtime report or a check sheet for a customer audit. Those records tell you what happened, but a model also needs the conditions leading up to it. A log entry reading "Line 3 down, 40 minutes, mechanical" says almost nothing about why the line stopped.
Second, even well run plants can have surprisingly thin digital data. Experienced operators catch trouble by sound, smell and feel, so the knowledge that prevents failures never gets written down anywhere a system can see it.
Digitization, Digitalization and Digital Transformation in Manufacturing
These terms tend to get mixed up in conversations about digital transformation in manufacturing, but they are easier to think about as layers.
| Term | What it means | Shop floor example |
|---|---|---|
| Digitization | Converting analog information into digital data | Replacing paper downtime logs with sensor readings and tablet entries |
| Digitalization | Using that digital data to change how a process runs | Triggering maintenance by actual machine condition instead of the calendar |
| Digital transformation | Rethinking operations and the business model around data | Running a connected smart manufacturing network across several plants |
Running a connected smart manufacturing network across several plants
For most AI used in plant operations, each layer depends on the one before it. Predictive maintenance, process optimization and production scheduling all lean heavily on connected data from machines and business systems. The order is not absolute, though. A vision inspection camera produces its own image data from day one, and generative AI can work through manuals and work instructions that already exist as documents. These can be good places to start, but on their own they usually stay small, one off wins until the rest of the plant's data is connected.
What Digitization in Manufacturing Looks Like on the Floor
Picture a mid sized plastics plant with six molding machines. Machine 4 starts turning out defective parts every few days. Here is where the information needed to figure out why sits today:
Machine data: temperature and pressure readings are stored on the machine itself and get wiped every few hours.
Maintenance: the last time a heating part was replaced is written in a paper logbook.
ERP: shows a batch of raw material arrived from a new supplier two weeks ago.
MES: records how many parts were scrapped each shift, but not how the machine was set up at the time.
Operators: the second shift lead knows the machine "runs cold in the mornings" and adjusts its settings by hand.
Every piece needed to solve the problem exists, and none of it is connected. Digitization here means saving the machine readings to a central database instead of letting them disappear, moving maintenance records into a digital system, linking material batches in the ERP to production runs in the MES, and giving operators a tablet to note any setting they change and why.
Once all of this data is tagged to the same machine and the same timeline, a model can check whether the defects line up with the raw material, the worn part, the room temperature or the manual adjustments. The plant had plenty of data all along. What it lacked was a way to connect the pieces, and that is most of what good manufacturing data management comes down to.
You Do Not Need to Replace Legacy Manufacturing Systems
AI in manufacturing often gets postponed because leaders assume it requires new machines, a new ERP or a full MES rollout. In most plants it requires none of these.
Older machines often already measure what matters through their built in controllers. That information can usually be read with small connector devices, without changing how the machine runs. Where a machine has no digital output at all, simple add on sensors for temperature, vibration or power use can be fitted, often during a scheduled shutdown.
The same goes for software. A 15 year old ERP or a custom built MES can stay in place. Integration tools can pull data out of them, put it into a consistent format, and send it to a modern platform where it can be combined with everything else. The old system keeps doing its job, and its data becomes usable for the first time.
This approach keeps production running and keeps costs down, which makes it much easier to get support from operations leaders who have seen big IT projects disrupt a plant before. A useful rule of thumb: if a system can be read, it can usually be integrated. Replace a system when it is failing or no longer supported. Wanting to use AI is rarely a good enough reason on its own.
A Practical Path to AI Readiness
Start with a problem that has a cost attached, such as unplanned downtime on a bottleneck line or scrap on one product family.
Trace the data that problem depends on across machines, maintenance, ERP, MES and operators, and note what is missing.
Digitize and connect only the machines and systems that problem involves, keeping existing equipment wherever you can.
Give the data a few months to build up, and start with simple dashboards. Visibility often pays for itself and exposes data quality gaps early.
Add AI once the data is trustworthy, then reuse the same foundation for the next use case.
Done in this order, digitization in manufacturing pays off in numbers plants already track: less unplanned downtime, lower scrap, faster troubleshooting and schedules based on how machines actually perform.
Where Expeed Fits
Most of what this article describes is the kind of work Expeed does every day. In the molding plant, capturing and storing the machine readings is industrial IoT work. Reading from the old ERP and tying it to MES production orders is legacy modernization and application integration. Turning the connected data into an early warning for worn parts is predictive analytics. We usually begin with one costly problem, connect only the data it needs, and build outward from there. If that sounds like where you are, we are happy to talk it through.
The Model Is the Easy Part
The plants getting real value from AI in manufacturing have usually done the unglamorous work first. Every machine, material lot, repair and manual adjustment gets recorded in a way that links to everything else. With that in place, choosing and training a model is a fairly small step. Without it, even a very good model has little to work with.
If your AI pilots keep stalling, the problem is probably sitting underneath them, in a controller, a logbook or an experienced operator's head. Smart manufacturing starts with getting that information out and into a system where it can be used.
Frequently Asked Questions
What should manufacturers do before implementing AI?
Pick one problem worth solving and get the data behind it in order. That usually means capturing machine signals, moving paper records into digital tools, and connecting systems like ERP and MES so their records line up.
What is the difference between digitization and digitalization in manufacturing?
Digitization turns analog information into digital data, like replacing a paper downtime log with sensor readings. Digitalization is what you do with that data, such as triggering maintenance when a machine's condition changes rather than on a fixed schedule. In most plants, digitization needs to be in place before digitalization can make much of a difference.
Can you use AI in manufacturing without digitizing first?
For some use cases, yes. A camera based inspection system creates its own data, and generative AI can work with manuals and SOPs that already exist as documents. Predictive maintenance, yield optimization and scheduling are a different story, since they rely on equipment history and context from several systems.
Why do AI projects fail in manufacturing?
The algorithm is rarely the weak link. Projects usually struggle because data sits in disconnected systems, older machines were never instrumented, records are incomplete, or there is too little history of the events the model is supposed to predict.
How do you digitize legacy manufacturing systems?
Start by finding out what each machine and system can already share. Many older machines have controllers whose data can be read with connector devices, and those without any digital output can be fitted with add on sensors. Older ERP or MES software can be connected through integration tools, so existing systems stay in place while their data becomes usable.
