Why industrial AI struggles to reach production
03/11/2026 Beckhoff Automation Ltd
Moving machine learning from a promising trial to the production line ~
Only two per cent of UK manufacturers say AI is widely embedded across their operations, despite the attention and investment the technology continues to attract. Here, Beth Ragdale, product manager at automation and control specialist Beckhoff UK & Ireland, looks at why industrial AI projects can struggle to make the journey from pilot to production, and what manufacturers can do to get them over the line.
There is an old rule among engineers: if a machine is running, don't touch it. It's understandable, since a machine has been commissioned to make a product at a particular quality and rate. Making changes introduces risk and, potentially, downtime. Expecting an engineering manager to interfere with a perfectly good machine because an experimental AI model might improve it is a big ask.
That perhaps explains some of the gap between industry's enthusiasm for AI and what is actually happening on factory floors. Make UK's 2026 research found that just two per cent of manufacturers considered AI widely embedded across their operations, while skills shortages remain a significant barrier to further adoption.
That reflects what I see when speaking to manufacturers. There is no shortage of conversation about AI, pilots and university projects. What I see much less frequently is an AI project that has made it all the way onto a machine and become part of normal production.
Before AI comes data
Part of the problem is that it's tempting to jump in straight with AI but several less exciting things need to happen first. You need the right data. On an operational site, that might mean information from machines of different ages, manufacturers and sampling rates. In some cases, simply getting useful information out of an older machine can be difficult.
Then there's quantity. I've had conversations that essentially begin with, “we want all the data from the machine.” Once you start doing the maths, high-frequency logging across dozens of variables can create enormous data sets and memory requirements.
There is also knowledge that sensors don't capture. An experienced operator may instinctively know how to get the best from a machine. We can add another sensor, but capturing years of operator experience in a data set is much harder.
Bridging the skills gap
Industrial AI also sits between traditionally separate skill sets. The data scientist understands the model, while the automation engineer understands the machine, where the model could realistically be incorporated and what could go wrong. Then there's the operator who knows what actually happens during a shift.
Many machine builders don't have an in-house data scientist. Beckhoff's TwinCAT Machine Learning Creator is designed to help bridge that gap, allowing automation engineers to create machine-learning models from their own data without specialist AI expertise.
Computer vision is a good example. In one Beckhoff application, around 200 images of metal workpieces were labelled as OK or not OK and used to train an image-classification model. The resulting model correctly classified the workpieces in more than 95 per cent of the cases considered.
The lesson here is to start with a tightly defined problem. Instead of setting out to create a futuristic lights-out factory, identify one inspection, classification or process decision where AI could deliver measurable value.
Getting out of pilot mode
Deployment isn't the end of the story either. Machines change. Components are replaced, motors age, materials vary and new product variants are introduced. As the real process changes, the data can change with it and the model's accuracy can deteriorate - something known as model drift.
That means new production data may need to be fed back into the model. But this creates another question: who decides what is right and wrong? For example, if a vision system generates another thousand images, someone may still need to establish which represent acceptable products before they become useful training data.
That's why getting industrial AI out of pilot mode isn't simply about developing better algorithms. Manufacturers need useful data, clearly defined applications and people who understand both the process and what the model is telling them.
For manufacturers struggling to get started, I'd resist the temptation to tackle the whole factory. Find one process that is well understood, establish what good data looks like and prove that AI can solve a genuine problem there. Sometimes, getting out of pilot mode means starting smaller.
Find out how Beckhoff’s TwinCAT Machine Learning tools can help bring AI into real-world machine and process applications here.
Print this articleFor more information, please contact:
Rachael Duckworth
Beckhoff Automation Ltd
Videcom House
Newtown Road
Henley on Thames
Oxon
RG9 1HG
Tel: +44 (0)1491 410539
Email: r.duckworth@beckhoff.co.uk
Web: https://www.beckhoff.co.uk
Share article:
Process and Control Today are not responsible for the content of submitted or externally produced articles and images. Click here to email us about any errors or omissions contained within this article.