Cutting Through the Hype to Reveal What AI Really Means for Manufacturing

**This story originally appeared in the April 2026 edition of AMT Magazine**

Part of the problem in truly understanding and acting upon artificial intelligence (AI) lies in defining the term, data scientist Mikhail Golovnya says.

He tells AMT that ‘artificial intelligence’ has become “a highly overloaded term,” a catch‑all explanation for almost any digital change – and his first move is to narrow its definition.

“What we name today as AI is usually generative AI, which is a narrow part of AI… but in the broader terms, especially in manufacturing context…AI simply means effective use of modern data analytics and machine learning techniques to address specific manufacturing and business needs,” he says.

For him, the most recent inflection point is large language models (LLMs) and the transformer architecture that finally let people ‘talk’ to machines.

“We have experienced the ability to communicate with machines using human language,” Golovnya says.

“Here comes the revolutionary part, and [here] also comes the entrapment. Some people are being fooled into thinking that we’ve somehow discovered the secret of intelligence.”

Under the surface, he adds, it remains number‑crunching at scale: GPUs, vectorised representations, and massive datasets. “Yes, it is a revolution, but it’s just another development in computer science.”

While some companies are trying to use generative AI in marketing and administrative roles, Golovnya says he doesn’t see a role for generative AI in manufacturing beyond some project management work.

“I still see a lot of potential to conventional, traditional machine learning…to unleash the power of digital transformation there,” he says.

To bring the conversation back to shop‑floor realities, Golovnya describes a layered model of digital transformation that manufacturers can put into action.

“We all need to start at level zero,” he says. That means clarifying the needs of a business and establishing a data foundation that directly supports it.

From there, “we go to level one. We have the data, we describe it. Then…level two. Okay, let’s diagnose, and apply our own intelligence and domain expertise to understand what conclusions we can draw from the data. Why did that machine fail?…Why did that thing go out of spec?”

Level three is predictive analytics: “Can we use [the past] to build models to predict what will happen in the future?” Then comes level four: “We apply human intelligence trying to decide, okay, what do we learn from this? What changes do we introduce into our business to optimise the outcomes?”

In practice, he sees many organisations – including Australian businesses – still at the foundation level, where others are trying to use data to develop control charts and carry out process control.

His caution is simple: don’t get trapped either collecting data for its own sake or “introducing innovations for the sake of innovation,” such as installing sensors that do not create durable value.

“We do need to start collecting the right data to address business [needs] – because if we don’t collect the right data then the whole digital transformation is irrelevant,” he says.

“We also need to properly invest into reliable sensors…they need to be utilised a lot more, because right now, they are underutilised, and that’s where the real business impact comes in.”

When considering whether to invest in an innovation or introduce a novel digital transformation technique, Golovnya advises manufacturers to ask themselves two questions.

“Number one, what is the long-term impact on the bottom line, the revenue – ‘long-term’ being the key word…Number two, what is the impact on customer satisfaction and the reputation of us as a company?” Reputation, he stresses, compounds slowly and can be lost in a day.

“We want something that will satisfy our customers instead of causing an unnecessary annoyance, [otherwise] eventually…they’ll start thinking we are scamming them for money, and we don’t care about them anymore,” he says.

To ensure that the right data is collected, Golovnya advocates for the building of operational models—digital twins—that mirror the real process and can be experimented with safely.

“Some associate it with AI, I do not. To me a digital twin is simply use of a computer simulation that mimics what is going on in your actual business.

“They need time to set up the system, but once it’s up and running, the benefits will start flowing in – so it’s this type of mentality that I would love to see more and more on the industrial side, as opposed to this nebulous idea of agentic AI and job cuts and eliminations, which in my view isn’t going to happen.”

That will require a workforce that has become AI literate, he says, with an understanding of what the technology can and cannot do.

Around that literacy, he sees new roles: technicians to install and maintain sensors, specialists in data security and data fidelity, digital twin managers, and troubleshooters who can clean up failed automations.

“Sometimes…AI may save us 100 hours of work and create 1,000 hours of cleaning up the mess…I’m not worried about [jobs in future], it’s just shifting,” he says.

Once manufacturers fully understand the data needs of their businesses and implement collection procedures for that data – whether it is real-time monitoring or the inputs needed to build a digital twin – Golovnya says there is more to do.

“Once you have that data available, don’t stop there. Go do the actual data analytics… and then create a loop, a feedback loop… Work with the data to see what you need or don’t need, adjusting strategies, adjusting processes,” he says.

The institution of a fully autonomous or “agentic AI” that makes its own decisions – is where he becomes sceptical.

“I’m personally not very keen on that, because there’s a lot of conceptual issues there… that’s kind of like a holy grail… I just don’t see it happening anytime soon.”

Golovnya says that he sees a lot of industry interest in adopting agentic AI systems as soon as they possibly can – but he instead urges caution.

“And I would say, that’s the last thing you want to do… There’s a lot of great potential with conventional machine learning and predictive analytics and all of that doesn’t have to go to the level of an agent. You may not even need that,” he says.

That is to say nothing of the regulatory and ethical challenges that an agentic AI system would create.

While Golovnya says “the genie’s out of the bottle” when it comes to AI regulation, he urges authorities to concentrate effort where the downside risk is catastrophic.

“It would be foolish to put an autonomous agent into a system responsible for nuclear response, or ICBMs, or anything of that nature…They can make a mistake. You can make an error. And then there are certain areas where a cost of even a single error is unacceptable,” he says.

“If you study the history, you’ll discover how many times having a human in the loop prevented an actual nuclear disaster…It’s fascinating when it’s a human intelligence overrode that final move.”

“So, we have to be very careful not to put that in place. It has nothing to do with intelligence, AI and all of that. It’s just how we put automation, or its modern, agentic form of automation, into these vital systems.”

“For manufacturing the biggest warning is to make sure that we protect trade secrets, that we do the data security and that we also guard against adversarial attacks,” he says.

“Because the agentic AI could be a good thing, or it could be a clever way to unleash a computer virus that pretends to be an agent to get that type of stuff.”

Golovnya also mentions emerging ethical concerns with large language models (LLMs) including potential “biases or imbalances” in analytical systems, and the training of the model on existing, copyrighted data and potentially untrustworthy sources from the wider internet.

“When the LLM model has no mind of its own…it may give you hallucinations that may sneak into the area of knowledge as if it’s actually real.

It’s for that reason Golovnya predicts that within the next five years the industry will shift focus onto what he calls “amplification of intelligence” – an understanding that AI systems empower or amplify human intelligence, but do not replace it.

“In 2030 people will say we underestimated complexity, we overestimated speed, and we forgot accountability. And because of that, we only moved the cost instead of eliminating it,” he says.

“I honestly believe agentic AI as a hot topic will disappear and it will yield to something like responsible AI, and the human intelligence will come back in vogue.”

Golovnya says that a part of that, as wider industry moves to adopt AI gather steam, is to collaborate with industry peers for future planning.

“Be human. Do not be afraid, but also do not be a fool…there’s no free lunch, and we need to know each other. We need to help each other navigate these treacherous streams of AI,” he says.

“For so many years people have been worried about machines outsmarting us, but in reality it seems to me that sometimes we are degenerating to the level of mindless machines.”

“I don’t want us to sell our intelligence cheap, ever… let’s be masters of it, and not turn ourselves into slaves,” he says.