Why AI alone will not transform manufacturing

Artificial intelligence has become one of manufacturing’s biggest technology priorities. Every week seems to bring another platform promising autonomous operations, intelligent automation or AI agents capable of transforming production. Yet despite the pace of innovation, many manufacturers continue to struggle to move beyond isolated pilots and demonstrations into AI that consistently delivers operational value on the factory floor.

Stephen Graham, Vice President Product & Technology for Hexagon’s Production Software division, believes the reason has little to do with the capability of the underlying technology. The real challenge is that manufacturing is fundamentally different from the environments where many AI systems have proved successful.

Factory operations are shaped by decades of engineering expertise, complex physical assets and constantly changing production conditions that cannot simply be reduced to data or algorithms. As AI becomes more deeply embedded in industrial software, the companies creating the greatest value will be those that understand manufacturing first and AI second.

“I think it’s absolutely everything,” Graham says when discussing the importance of industrial domain expertise. “A company that’s approaching this just from the AI pitch… it just doesn’t apply. It absolutely starts with the data readiness problem, but it’s also the practicality of having feet on the shop floor on a daily basis and understanding the real-world problems.”

That distinction is becoming increasingly important as industrial AI matures. Manufacturers are no longer asking whether AI can generate code, analyze data or automate repetitive tasks. Those capabilities are rapidly becoming expected. Instead, attention is shifting towards a more demanding question: can AI operate effectively within the realities of production, where decisions affect expensive equipment, product quality, workforce safety and delivery commitments? Unlike many business environments, factories combine digital systems with physical processes, meaning every recommendation ultimately has real-world consequences.

The difference between generic AI and industrial AI therefore extends well beyond software. Success depends on understanding how production works, how engineers solve problems, how experienced operators respond when conditions change and where automation genuinely improves performance without removing essential human judgment. Graham believes manufacturers should think of AI not as a replacement for engineering expertise, but as a way of extending it. The most successful deployments are those that accelerate existing workflows, help experienced engineers make faster decisions and allow valuable knowledge to be shared more effectively across the workforce rather than attempting to remove people from the process altogether.

Good AI starts with good manufacturing data

Few manufacturers would disagree that data sits at the heart of industrial AI. The greater challenge is making that data usable. Production information is often spread across machine tools, enterprise systems, engineering applications and decades of legacy software, creating fragmented datasets that are difficult to integrate and even harder to trust. Without that foundation, even the most sophisticated AI models struggle to deliver meaningful operational improvements.

Graham believes this remains one of the biggest obstacles preventing manufacturers from realizing AI’s full potential. “Data is definitely one of the biggest challenges,” he explains. “AI is nothing without data to feed from. The higher the hierarchy you go, the more data there is that’s available, but also the more fragmented it gets and the more spread across different systems. Being able to integrate your systems to bring the data together in some kind of consistent way is pretty much a prerequisite to getting full value out of AI.”

The problem takes different forms depending on the size of the business. Smaller manufacturers often lack the infrastructure and governance needed to capture production data consistently, while larger enterprises face a different challenge as information becomes dispersed across multiple facilities, software platforms and business functions. In both cases, AI is only as effective as the quality and context of the information it receives.

Bringing data together is only part of the challenge. AI also must earn the confidence of the people expected to use it. Engineers are unlikely to adopt tools that force them to abandon established processes or obscure the reasoning behind important production decisions. In manufacturing, trust is built through familiarity as much as technical performance.

“The most successful deployments of AI we’re seeing are the ones that fit into the existing workflow and support engineers in the way they’re working today,” Graham says. “They accelerate the tasks people already perform while leaving accountability with the engineers rather than trying to switch it over to the AI.”

That philosophy reflects a broader shift in industrial AI adoption. Rather than pursuing fully autonomous decision-making, manufacturers are increasingly looking for practical tools that remove repetitive work, shorten engineering tasks and improve consistency while allowing experienced people to remain firmly in control. For Graham, AI delivers its greatest value when it strengthens established manufacturing processes instead of asking organizations to redesign them around the technology.

Manufacturing knowledge cannot be automated away

One of the biggest misconceptions surrounding industrial AI is that manufacturing expertise can simply be captured, codified and handed over to software. Graham argues the reality is far more complicated. Every factory develops its own ways of solving problems, shaped by the experience of engineers who understand not only how machines are designed to operate but how they behave after years of continuous production.

“There’s a lot of stuff that’s uncodifiable,” he says. “The practical reality of understanding how a machine works, the quirks of this machine versus that machine and the daily reality of how people work on the shop floor – all that deep domain expertise is something that’s very difficult for AI to get hold of.”

Many manufacturers now face a demographic challenge that extends well beyond recruitment. As experienced engineers retire, companies risk losing decades of practical knowledge that has never been written down because it has been learned through years of solving production problems on the shop floor.

“Manufacturing companies are trying to hang on to people in their late 60s and early 70s because they’ve got decades of experience of how this really works on a day-to-day basis,” Graham explains. “AI getting up to speed is a little bit like fresh engineers coming out of college and getting them up to speed as well.”

The distinction is important because manufacturing combines digital decision-making with physical execution. Unlike many office-based environments, production systems operate in conditions where unexpected failures, material variations and equipment behavior frequently require human judgment. An engineer standing beside a machine can draw on years of experience to recognize subtle warning signs or improvise a solution that would never appear in a training dataset.

That reality also explains why Graham sees augmentation rather than replacement as the industry’s most realistic path forward. AI can recommend actions, accelerate programming, identify patterns and surface relevant information, but accountability ultimately remains with the people responsible for quality, productivity and safety. The objective is not to remove engineers from the process; it is to give them better tools to make faster, more informed decisions while preserving the practical knowledge that continues to underpin successful manufacturing.

The future belongs to engineers, not autonomous factories

For all the discussion surrounding autonomous manufacturing, Graham believes the industry often confuses automation with autonomy. Modern factories have relied on highly automated production equipment for decades, but genuine autonomy requires something very different. It demands systems capable of interpreting changing conditions, responding intelligently to unexpected events and making decisions that manufacturers are prepared to trust.

“Automation lends itself very well to situations where things are highly repeatable and well understood,” he explains. “Autonomy requires some degree of intelligence and some degree of feedback loop to respond to what’s happening. There’s a spectrum from pure automation through to full autonomy where AI becomes part of that decision-making loop.”

The challenge is that production environments rarely behave as predictably as software developers would like. Equipment failures, material inconsistencies, changing customer requirements and countless small variations are part of everyday manufacturing. Those exceptions are precisely where experienced engineers add the greatest value and where AI still has the most to learn.

Trust also becomes far more significant once AI begins influencing production decisions rather than simply providing recommendations. Manufacturers may be prepared to accept occasional errors from software that assists with administrative tasks, but the consequences are very different when mistakes affect expensive machinery, production schedules or product quality.

“If it gets it wrong once, you can cost hundreds of thousands of dollars on the shop floor,” Graham says. “For anything that’s really going to make a difference, you’re always going to want that verification piece. I doubt we’ll get to the point where we fully delegate responsibility on the shop floor to AI.”

That does not mean progress will be slow. Graham sees significant opportunities for AI to remove repetitive engineering work, shorten programming times and help experienced specialists apply their expertise more consistently across multiple facilities. The manufacturers creating the greatest operational value are unlikely to be those pursuing fully autonomous factories. Instead, they will be the ones that use AI to strengthen engineering capability, improve production readiness and enable better decisions long before the first machine begins producing parts.

AI succeeds when people succeed

Industrial AI will not succeed because algorithms become more powerful or computing resources become cheaper. It will succeed when it fits naturally into the realities of manufacturing, supporting engineers rather than replacing them and building on the knowledge that already exists across the factory floor.

Graham believes that is where the industry is heading. “There’s massive opportunity for acceleration and scaling expertise,” he says. “But I can’t see the point where we fully delegate responsibility on the shop floor to AI. The practical reality of manufacturing means you’ll always need people in the loop.”

The manufacturers that gain the greatest competitive advantage will therefore be those that treat AI not as an autonomous replacement for human expertise, but as a practical tool that helps experienced people make better decisions, improve consistency and solve problems faster across every stage of production.