Every morning an onslaught of newsletters about manufacturing topics hits my mailbox. I am not complaining, I signed up for them in the somewhat naïve hope to thoroughly peruse them on a daily basis.
I don’t – who does? – but I skim them and occasionally find an article that I want to comment on, like this one recently published by IBM. “How is AI being used in manufacturing?” talks about use cases and benefits as well as challenges of AI in manufacturing and that’s the section I want to comment on based on my experience in the field, deploying AI-based solutions on the manufacturing shopfloor.
Challenges of AI in Manufacturing – Our Perspective
Here is the list of challenges of AI in manufacturing published and our perspective from implementing solutions on the shopfloor.
“Data quality and availability: AI relies on high-quality data, but manufacturers often lack the clean, structured and application-specific data needed for reliable insights. This is especially true in areas like quality control, where incomplete defect data can impact model accuracy.”
Well, yes and no. Data availability and quality is definitely the Achilles heel of many AI applications in manufacturing and elsewhere. Our assumption always is that the company does not have the needed data readily available but we need to collect it before we start training the AI model. However, data for quality control applications is generally not the issue. Unless you operate in a very low throughput environment, you can collect enough data within a few weeks or even days to start training the model to recognize defective products. Over time the it will learn to classify different defect types and even identify new, previously unknown ones. In our experience, AI models can achieve very high accuracy (>99.99).
Now, predictive maintenance is a different issue. There, it is often not clear which data is needed, whether it has been collected so far, and if so, where it is stored and in what format.
“Operational risks: Manufacturing requires high accuracy and reliability, yet some AI models, such as generative AI, are still maturing. Current models can lack the precision needed in production environments.”
Again, yes and no. Generative AI is definitely still maturing, prone to errors (who hasn’t argued with ChatGPT and angrily told it “You are making stuff up!”) and even hallucinations. But this is not true for what is now referred to as classical or mature AI. The good old machine learning algorithms are mostly deterministic. They don’t make mistakes, the same input will result in the same output, there are no hallucinations and creative interpretations of what a big scratch in the middle of your polished surface or an imperfect welding seam means.
The lesson learned here is that you need to select the proper tools for the job. Gen AI is not ready for the shopfloor, mature AI is.
“Skills shortages: There’s a scarcity of professionals with expertise in AI, data science and machine learning. This shortage makes it challenging for companies to fully use AI without investing in workforce development.”
Yes, but … Data scientists and machine learning (ML) experts are definitely not a dime a dozen and, let’s be honest, the ones out there aren’t exactly banging on the doors of manufacturing companies.
But, trying to hire highly educated and therefore expensive people to operate highly complex and finicky systems is the wrong way to tackle this challenge. The solution is to develop AI-based products that are easy the use and implement by the engineers and technicians already on the shopfloor so they can take over managing and maintaining the AI-solution after an initial training. It means to put a lot of upfront thought and work into the solution to make it user-friendly. It can be done, in fact we have done it.
“Cybersecurity concerns: AI integration increases digital connectivity, opening more potential points for cyberattacks. Manufacturers need advanced cybersecurity measures to protect sensitive systems.”
True, this is a real, valid and wide-spread concern and one of the main challenges of AI in manufacturing. The situation is somewhat similar to the adoption of Cloud solutions: occasionally the door needs to be cracked open in a managed and secure way. What makes the challenge manageable is to limit the time that door is open by running the AI-based systems on premise, behind the firewall and only connect to the outside world when software updates need to be installed or some other intervention is needed.
Of course, all the standard security measures need to be applied to AI – just like to any system.
“Change management: Virtually 100% of organizations surveyed deemed at least some level of impact from AI and automation. Integrating these technologies can meet resistance from employees concerned about job security. Clear communication and retraining can help ease this transition.”
Absolutely, in fact change management issues are – in our experience – the number one reason why projects never get off the ground. Job security can be one reason for concern, though in manufacturing the problem is often finding enough people for jobs like quality inspection. Resistance to change overall and a lack of understanding what AI is, what it can and cannot do and what the best first use cases are from the executive level on down ranks high on the list of challenges of AI in manufacturing.
Change management, of all the challenges, is the most difficult one to address and overcome. However, the value AI can bring to manufacturing is just too convincing and early adopters who overcome that roadblock will reap the benefits and gain a substantial competitive advantage.
“Implementation costs: AI adoption requires a large upfront investment in technology and infrastructure, which can be a barrier, especially for smaller companies.”
It can but it doesn’t have to. It comes down to the old build vs. buy question. Building a solution from scratch might sound like the way to go (“we’ll just hire a couple of recent grad computer scientists”) but isn’t. An AI solution that integrates with the manufacturing environment, is robust, reliable, accurate, fast and user-friendly requires more than training an ML algorithm with some data. In all of our years in this business we have yet to see a successful in-house developed AI solution in production. In summary, developing your own AI solution is both expensive and risky.
However, buying a solution custom-build for manufacturing addresses both of these challenges: a proof-of-concept ensures that the solution does the job and knowing the cost upfront allows you to calculate the ROI and make sure you are getting a big enough bang for your bucks.
We are tooting our own horn here a bit by suggesting you give our solution, The Accella Quality Box for visual inspection a look. We have put years of work and decades of experience in manufacturing into building an AI-based platform that is custom-tailored to the needs of manufacturers. It is designed to make smart manufacturing a reality with the people you have and at a price point that ensures a fast return on your investment.
Please contact us with any questions.
