Artificial intelligence has the power to transform manufacturing — but the industry’s current strategy for implementing it is “deeply flawed.”

That’s the contention of Matthew Hart, founder and chief executive officer of Soter Analytics, a maker of AI-driven wearable devices.

The market size for AI in manufacturing is expected to rocket from $3.2 billion in 2023 to $64.63 billion by 2030, according to a report by the consultancy and tech services firm SkyQuest Technology. Why, then, are only one in five generative AI projects deployed by manufacturers today considered to be a success?

Hart references the mining industry, in which he has extensive experience, as one sector struggling to fully embrace AI. A report by Deloitte cites multiple challenges within traditional mining companies, including poor testing methods for training AI models, a culture that’s resistant to innovation, a limited understanding of operational and financial returns, and a lack of advanced technology skills. The underlying problem, Hart says, is a massive amount of legacy technology and equipment that can’t be easily retrofitted to work with the latest AI innovations. Companies are attempting “to lug all of that into an AI solution that effectively automates some kind of decision-making.”

In theory, AI technology is just the thing for companies looking to streamline operations in line with demands for greater adaptability and efficiency. “What they find,” says Hart, “is that the scope just creeps and creeps. And legacy systems don’t do what they need them to do. So implementation [of AI] ends up being quite complicated.”

Legacy systems that don’t mesh well with AI include enterprise resource planning, material requirements planning and manufacturing execution systems. In addition, closed-circuit camera systems installed decades ago often have no way to transmit the data they’re collecting to a centralized depository. An expensive server upgrade might be required.

There’s the additional problem of AI itself not quite living up to the hype. GenAI continues to suffer from “hallucinations” and the dispensing of advice that’s just plain wrong. AI models have a way to go before they’re able to sort through massive stores of information, scraped from all over the internet, and extract proper answers on demand.

The very nature of machine learning, an essential aspect of AI, is that it improves with experience. But Hart says some of the data needed to train models remains trapped in legacy applications and can’t be formatted to be of use to modern systems.

Enthusiasm for AI can lead to overly optimistic expectations of functionality. Adopters believe the necessary data can be fed into the system for quick results. “Then the reality kicks in.”

Hart says managers make a mistake when they “turn this into a huge process. We’ve refused to be part of some projects where we can see that it’s just so complicated.” The very scale of the strategy ends up undermining the success of the initiative.

The better way to approach an AI implementation is to do it in phases, keeping humans in the loop along the way, Hart says. They’re still needed to make ultimate decisions about such issues as safety, quality, productivity and auditing.

Hart expects people to be remain relevant in manufacturing organizations for years to come. An AI application that relies on past data, no matter how well it’s trained, isn’t necessarily prepared to deal with events that haven’t happened before. That’s a strength of human creativity. “We’re really good at coming up with new systems.”

That said, Hart believes AI will play an increasingly vital role in multiple aspects of business operations, especially where high volumes of data are involved. Workplace safety compliance, requiring knowledge of thousands of ever-changing regulations, is one area that’s ripe for monitoring by AI. Another is equipment maintenance, for which GenAI combined with computer vision can process status data “at a scale that’s never been done before.” Automated systems are able to run multiple scenarios and workflows in parallel, leaving it to human engineers to determine the best path forward.

Looking to implement AI in manufacturing successfully? Start by use applications like ChatGPT to obtain “relatively decent answers, but at a very basic level,” Hart advises. “Then you can build GenAI workflows that replicate human ways of thinking.”

It’s equally important to be specific about the issue you’re trying solve with AI. “When you break it down to a small, specific problem,” Hart says, “you can build very high-quality solutions.”