What It Takes to Make Smart Manufacturing Tools Work with What You Already Have

There’s a particular kind of disappointment that shows up about six months into a lot of manufacturing AI projects. The model works fine. Everyone agrees the demo was impressive. And yet somehow, the thing isn’t actually delivering the value that got it approved in the first place. If you dig into what’s going on, it’s almost never the AI itself. It’s that nobody quite figured out how to get this smart new tool talking to the twenty-year-old ERP system, the PLM database with its own quirky naming conventions, or the three different spreadsheets someone in quality still maintains by hand because “that’s just how it’s always been done.”

This isn’t a niche problem. It’s arguably the central problem in bringing AI into manufacturing, and it doesn’t get talked about nearly enough, mostly because it’s not a fun thing to talk about.

The Gap Between the Demo and the Floor

Vendor demos are, understandably, built to show off what a tool can do under ideal conditions. Clean data, a well-scoped use case, everything wired up neatly in the background. What they don’t show you is what happens when that same tool has to pull real inventory numbers from an ERP that was implemented two decades ago, heavily customized by several different consultants over the years, and documented mostly in one retired engineer’s head.

That gap between demo conditions and an actual production environment is where a lot of manufacturing AI projects quietly stall. This isn’t a sign that manufacturers are behind the curve, or that the technology doesn’t work. It’s simply that manufacturing environments tend to be genuinely messier and more layered than the clean data pipelines most AI tools were originally built and tested against.

Where This Shows Up Most Clearly

Look at where AI is actually delivering value in manufacturing right now, and where it’s disappointing people, and the pattern usually comes down to how well the new tool connects to what already exists.

Take generative design and automated engineering documentation. These tools sound almost magical in a pitch: describe your constraints, get back a design candidate. But a generative design tool is only as good as its access to actual CAD history, simulation results, and the specific material and cost constraints an engineering team has accumulated over years. If that data lives in a system the AI tool can’t reach cleanly, or it’s scattered across formats that don’t play well together, the result is a tool producing interesting output disconnected from what a plant can actually build. The AI didn’t fail. The plumbing wasn’t there to support it.

The same pattern shows up with predictive maintenance, quality inspection, and essentially any AI application that needs a steady, accurate stream of information from systems that weren’t designed with AI in mind. The tool is rarely the bottleneck. The connection is.

Why Manufacturing Makes This Harder Than It Looks Elsewhere

This isn’t unique to manufacturing, but it does tend to bite harder here than in many other industries adopting AI.

Part of it is age. Many manufacturers run enterprise systems that have been in place for a decade or more, customized along the way in ways that made sense historically but complicate any attempt at standard integration. Part of it is fragmentation — ERP over here, PLM over there, a separate quality management system, a historian database tracking sensor data that few people outside maintenance ever look at closely. And part of it is that the stakes are simply higher. A recommendation engine getting something wrong in retail is a minor inconvenience. A manufacturing AI system working from stale or incorrect engineering data can create a genuine safety or cost problem, not just a poor customer experience.

None of this makes the situation hopeless. It just means the work of connecting systems properly deserves far more attention than it typically receives during early project planning.

What Actually Helps

Manufacturers who avoid the worst version of this problem tend to do a few things differently, and none of them are particularly glamorous.

They start by honestly assessing what existing systems can and can’t do before committing to a specific AI tool. That sounds obvious, but it’s surprisingly rare — most projects begin with “we want to do X with AI” and only later ask whether the underlying systems can actually support that. Reversing that order tends to prevent a lot of pain later.

They also stop treating integration as something to be sorted out during implementation, almost as an afterthought to the “real” project. Giving it a dedicated budget, timeline, and planning attention upfront changes how the whole initiative gets scoped, and produces far more realistic expectations about how long the work will actually take. In practice, this is often the point at which manufacturers bring in specialists such as ERP Consulting for Manufacturers to assess, ahead of any commitment, whether existing enterprise architecture can genuinely support the real-time data flow that a proposed AI system will need.

Finally, manufacturers who get this right tend to test integration during the pilot phase, not just model accuracy. It’s easy to run a pilot that proves a model performs well on a curated dataset. It’s considerably more useful to run one that also proves the model can reliably pull real data from production systems and write results back into them, without someone quietly reconciling spreadsheets behind the scenes.

The Practical Takeaway

The exciting part of AI in manufacturing — the model, the algorithm, the capability itself — is genuinely the easier half of the work. The harder half is the unglamorous task of making sure that capability can actually see and use the data a plant already has, sitting in systems built for a different era. That work rarely photographs well for a case study, but it’s what separates the AI projects that quietly become part of how a plant runs from the ones that fade from leadership conversations within a year.

Author Profile

Adam Regan: Deputy Editor

Features and account management. 7 years media experience. Previously covered features for online and print editions.

Email Adam@MarkMeets.com