AI on the Shop Floor: Ignore the Hype, Here's What's Real

There's a pitch that every manufacturing leader has sat through at least once. A technology vendor, sometimes a startup, sometimes a household name, presents a platform that will transform your operations. Inspection: solved. Supply chain: solved. Logistics, procurement, RFQs through to order matching: solved, solved, solved.

None of them do what they promise. They might work on the happy path. They might demo beautifully on a single part, in controlled conditions, with clean data. But manufacturing doesn't run on the happy path. Factory management is exception management. The job isn't dealing with the boring day. It's dealing with everything that deviates from it.

Take procurement. Tools that promise to manage your entire supply chain, from RFQ through to order, matching, logistics, work if you have a limited set of suppliers who always deliver on time, whose data arrives in the same format, and who you deal with in the same structured way. That's not the reality of manufacturing life. The edge cases are where manufacturing businesses actually operate, and the edge cases are exactly where these tools fall apart.

So where does AI actually deliver value on a shop floor today?

The honest answer is less exciting than the pitch decks suggest. The biggest win right now isn't AI analysing data. It's AI enabling people to build the tools that capture and surface data in the first place.

Most manufacturers don't have sophisticated digital teams. They have problems, a 5S process that lives on paper, quality audits that get filed and forgotten, inventory that's tracked in someone's head. AI has compressed the time to solve these problems from months to days. A team can go from paper-based data collection to digitised, tracked, and trended information in a matter of days, not quarters. The analysis layer on top, finding patterns, surfacing insights, is genuinely powerful. But for most manufacturers, it's the bonus round. The real step change is getting data in front of the right people at all.

This is the thing the AI conversation in manufacturing gets backwards. The question isn't "which model should we use?" It's "can we actually capture this signal reliably?"

The use cases that get the most airtime, vision and inspection, predictive maintenance, are real. They're also really difficult to get right.

Vision and inspection isn't complicated in a lab. Demonstrate on one part, in one set of conditions, that machine vision can detect a defect. Fine. Now scale it to a hundred SKUs, with engineering tolerances that change, stored across PDFs, Excel sheets, and drawings buried in a PLM platform. Put it on the factory floor, where it works at 11 o'clock in the morning but fails at 4 in the afternoon because the lighting changed, or there's vibration, or dust on the lens.

Then wait six months. New product introduction comes along. Do you have the skill set in-house to retrain the system? Maybe. More often, the integrator who built it is long gone, and the low-code platform that promised easy updates doesn't deliver. There's a reason people doing serious machine vision work in Python and sophisticated ML libraries, not drag-and-drop interfaces.

Predictive maintenance follows a similar pattern. When organisations implement a predictive maintenance platform, most of the benefit, maybe 80%, comes from simply having a modern digital CMMS. It's not the prediction that transforms maintenance. It's the fact that they finally have a coherent system. The AI layer adds value, but it's sitting on top of a much more fundamental upgrade.

Underneath all of this is a data problem that's bigger than most people realise.

There's the data thread everyone talks about: PLM to ERP to MES to the shop floor IoT network, SCADA, machine-level sensors, and back again. The "digital twin" conversation. That matters.

But there's another layer, a hidden chasm, that gets far less attention. All of the messy, ancillary data that actually runs a factory day-to-day. Where does your logistics data really live? Your ERP has some of it, but not all. Your MES captures physical operations on a product, but does it capture the 5S inspection of the build area? The quality audit? The 8D problem-solving? The data you need for certification, regulatory compliance, or continuous improvement?

That disconnected data, less sexy, harder to access, spread across paper forms and shared drives and someone's notebook, matters just as much as the clean digital thread. Until it's connected, AI has nothing useful to work with.

When AI tools do make it onto the shop floor, the people question is simpler than most change management consultants would have you believe.

An AI-based tool should not be distinguishable from any other piece of manufacturing technology. It either works or it doesn't. Bringing in an AI solution is no different to bringing in a digital measurement system or a conveyor. It's a tool to improve the process.

People get on board if it works and it works for them. When it behaves unpredictably, they lose trust, and they lose trust fast. Worse, they don't just lose trust in that specific tool. They lose trust in the entire category. Try an AI-based inspection system that fails, and the next ten AI initiatives become harder to land. Not because the technology is wrong, but because the first impression was.

Where is AI in manufacturing in five years? Honestly: no idea. And anyone who tells you they know is selling something.

The pace of change in the last twelve months is so far beyond what anyone could have predicted that five-year forecasts are worthless. What's clear is that AI will fundamentally transform how things get made. The question for every manufacturer isn't whether it will affect them. It's whether they're at the front of the curve or one of the first to be swept away by it.

The way to stay at the front is unglamorous. It's having individuals on shop floors and management teams who are pushing boundaries, trying things that make people a bit uncomfortable, taking ownership of staying current. Not waiting for a vendor to show up with a solution. Not waiting for a strategy document. Just building, testing, learning.

This is a tidal wave. There are two options: get smashed by it, or try to ride it as long as you can.

Tom d'Arcy is Head of Manufacturing & Deployments at AUAR, where he builds robotic construction systems. Previously Rolls-Royce, Edwards Vacuum, and BCG.