Manufacturers are doing AI. Or they think they are. While the pressure is real and the tools are moving fast, there’s still a problem most companies don’t want to admit. They assume their data is ready when it isn’t.
In most cases, they’re just layering powerful tools on top of data that was never built to support them. While everyone is racing to do something with AI, there’s a huge gap between having it deliver value and where most manufacturers currently stand both in their data state and in their connectivity to the equipment on the shop floor.
These tools can bring real value quickly. But to do that, the datasets need to be structured, unified and put in context so that the large language models can use them correctly. Otherwise, you’re just going to be tossing money out the door.
Finding structure
Every vendor has created their own data structure: how it’s worded and how it’s made available to a networked system. Even across newer equipment, each one is different. Ask for current draw across your CNC machines and you’ll get it back five different ways with five different names. There’s no consistency, no way to query across it reliably and no way to use AI and receive anything meaningful back.
And it isn’t just the machines. There’s a whole layer of asynchronous activity around the shop floor that has to come in alongside machine data. Think temperatures, humidity, compressor pressure, voltage fluctuations. All of it matters. All of it needs to be quantified and unified in the same structure because that’s what gives AI the context to find patterns. Without it, you’re missing half the picture.
The fix is a unified naming structure. You create organization across all your data sources – every machine, vendor, signal – so that when you query for something, you get it back the same way every time. You know it came from the CNC area of your plant and all your vendors are now organized in the same manner. No gray areas about what you’re asking or what you’re getting back.
Every source named differently, coming through a different medium, OPC, analog signal or whatever is dirty data. The source is disorganized, non-uniform and you can’t put it in context or make decisions from it. Without that foundation, it’s garbage in, garbage out. That isn’t a new concept but AI makes the consequences a lot more expensive.
People point to ERP as proof that AI works in manufacturing. It does work there and the reason is straightforward. ERP is a financial control system: transparency for tracking orders using standardized business practices, all driven from an IT structure. When you say those words, you know there’s a lot of structure there.
The shop floor is a completely different animal. The floor is chaotic. It’s an ever-changing battleground. If you try to instill standardized business practices for how you want your floor to run, more than likely it doesn’t align well and is ignored or implemented incorrectly. The data was never the point. Making parts was.
MES is what bridges that gap. On the shop floor, MES and SCADA are the sweet spot for where data is aggregated to make decisions, add schedules and provide insight. MES is the ERP of the floor. If a manufacturer has moved through and matured into a modern MES layer, they’re three-quarters of the way down the road. The problem is that MES deployments typically start with assets deemed higher value and end up less complete than they should be. The variability problem doesn’t disappear but just becomes smaller. And you still have to do the normalization work.
Real obstacles
What actually kills most of these projects isn’t the technology, it’s the people. Any change that comes down to the shop floor is read as a threat and AI reads as a bigger threat than most. Many times, an operator perceives any change coming down as a threat to their job.
Until you can articulate how it will improve the organization as a whole and won’t threaten individuals at any level, you’ll struggle from the start. It comes down to clear communication of the goal, the problem and what you need from everybody to solve it. Resistance means inconsistent inputs and inconsistent inputs mean dirty data. AI initiatives fail for reasons that have nothing to do with AI.
Assembly operations are where it’s the hardest. On the machine side, you can pull data directly from the equipment; the machine generates it regardless. Assembly work is still very much dependent on human dexterity, and that’s where you lose the connection. There’s variability built into it.
When you go from open loop, where an operator can do whatever they want without any connectivity or guidance, to something with structure and control, you change what’s possible. That structure can include digital work instructions and control of the tools and devices operators use.
On the assembly side, an operator has a screen in front of them and is guided step by step with pictures, arrows and video. The screen might instruct the operator to hand-start four bolts. In the next step, the DC tool is enabled with the torque, angle and fastener count and the screen shows them the order to run them down. With that structure, you connect to that information much more easily and consistently. That’s the only path to reliable data on the assembly side.
What’s AI actually doing on the shop floor today? It’s making individuals, including supervisors, leaders and managers, more efficient. It allows them to get done what they normally would, but quicker. What we aren’t seeing yet is that freed-up time being redirected into improving how the operation runs as a whole. That step requires the data foundation to be in place first.
The edge bridge
Edge computing is the other piece manufacturers underestimate until they’ve felt the pain of not having it.
The cloud makes sense: scale, redundancy, centralized data, AI working across everything from one place. But the floor can’t tolerate a 2-sec. response time. Think about an operator sitting in front of a screen being guided through a process, pressing an acknowledgment that says they’ve completed a step. If there’s an internet hiccup, some EMI, whatever, and that round trip takes 2 sec., multiply that across hundreds of operators and you have a real problem.
Edge computing is the bridge between the real-time demands of the floor and the cloud. The edge executes in real time, no latency. It snaps back a response immediately and reports up to the cloud as appropriate. Someday there won’t be a need for it but for the foreseeable future, it’s the reality.
Epicor is building toward something most vendors aren’t attempting: a turnkey solution for manufacturers from ERP all the way through to the shipping dock. Information flows seamlessly and AI can be leveraged across all of it in a consistent way, instead of picking vendors and dealing with the finger-pointing that comes with it.
Only a couple of vendors, if that, are actually pushing in this direction. Every integration point between vendors is a place where data becomes inconsistent and your single source of truth starts to fracture. One platform eliminates that.
Manufacturers that normalize their data, build the right structure and connect their systems will be in a fundamentally stronger position. They’ll make decisions based on consistent information, quote delivery dates they can actually meet, and have insight into their supply chain the way they have insight into their own floor. They’ll see correlations between variables that were invisible before and use them to continuously improve. Higher quality, more throughput, more visibility, faster deliveries.
Whether you skip the data work entirely or just assume it’s already done, the outcome is the same. AI will expose what’s actually there. And the companies that don’t deal with it are going to be sadly mistaken.
AI is changing the way manufacturers approach production. Dig deeper into the topic by visiting our archive.
