AI use in industry is on everyone’s minds. How will it help? What will it change? Will it solve the perpetual problem of the skills gap, or accelerate a loss of institutional skills? Deloitte’s recent AI Institute report, The AI Dossier, breaks down AI use cases for physical assets, infrastructure, employee safety and more. Plant Services spoke with Tim Gaus, principal at Deloitte Consulting, about the report’s findings on the role of AI in predictive maintenance and quality control.

A quick definition: AI in the report refers to “machine learning, deep learning, large language models and agentic, even physical AI,” Gaus says. Establishing a definition is important, since AI use can mean anything from autonomous robots to using Claude to try to tackle a thorny engineering problem during your lunch break.

The results of AI deployment vary. One Deloitte client, a company in the metal industry, reduced its mean time to repair by 45% using an AI agent-driven maintenance routine. On the other hand, Gaus says, some AI projects don’t change anything at all. The difference, he says, is setting yourself up for success.

In order to do so, Gaus recommends the following best practices for using AI in maintenance and reliability.

• Pick a problem you want to solve and explore how you can apply AI to it, as opposed to applying AI for the sake of new technology. 
• Pick an AI project for which you have confidence in the data. The more data you provide to the AI, the better it performs. At the same time, don’t wait for the perfect data before you try something new. 
• Always keep the human worker in mind.
• Consider repeatability and scalability.
• Don’t be afraid of your IT department but recognize you’re the OT expert. It takes both to succeed. 

Ultimately, it comes down to balance. Gaus provided the example of a client in the distribution space that implemented an advanced AI solution while also keeping human needs in mind.

“They maintain a huge asset base, and they've put in a very sophisticated solution that both allows the AI to predict downtime, but also combines it with agents and other elements, so that the predictions are being brought to the users,” Gaus says. “They also allow the users to refine and tune, so you don't get all the noise. Because sometimes these AI things turn into a bunch of annoying pings that don't actually make any sense.”

That prediction element makes AI especially useful for maintenance, enabling workers to move from detection to anticipation.

Industrial AI agents can also be used for continuous sensor monitoring and anomaly detection, root cause analysis, automated work order generation and scheduling, plus simulation and reinforcement learning. AI agents can also work with human-centric teams, validating outcomes, prioritizing alerts and presenting their findings in natural language.

“If you always struggle to get access to the machine data and compare it to your maintenance records, that's another great [AI] use case for maintenance and repair that's shown a lot of value,” Gaus says.

According to a survey from the Manufacturing Leadership Council in early 2025, nearly one-quarter (22%) of manufacturers expect to use generative and agentic physical AI in the next two years – more than double the number using it today (9%).

AI can bridge the knowledge gap between new and retiring employees

According to Deloitte’s research, AI adoption is scaling, but “most of the usage to date is within various constrained spaces,” Gaus says.

AI is particularly useful in transferring institutional information to newer employees that might otherwise be lost when workers retire. Some companies try to reduce the loss of expertise by turning retiring workers’ knowledge into AI agents that can mentor new employees.

“I'm working with a company who has a lot of their senior engineers, in this case capital engineers who do a lot of the implementations of their programs, start to think about retirement or are retiring,” Gaus says. “They've got decades of knowledge, not only in their heads, but also in their emails and in their files. This company used AI to both try and take out as much tacit knowledge as they could from an interview process, but also combine that with their files.”

The result is an AI agent with the collective knowledge of the engineering worker. Junior engineers can chat with the agent or have it automatically review documents. 

“That way, you have almost the senior engineer oversight without having the senior engineer there,” Gaus says.

AI combines with robots for autonomous inspection

One use case Deloitte highlighted was autonomous physical AI for infrastructure inspection. Autonomous infrastructure inspection is nothing new: vision sensors have been in use for 25 years, Gaus points out. However, the sensors have become easier to put in place and much less expensive, perhaps $20 instead of $40,000. Mobility has also changed, with robots conveying sensors on legs or wheels.

A robot can now patrol the facility, taking images and temperature readings or reading gauges. It can automatically cover work that is physically hard to reach or that employees tend to overlook.

Autonomous systems could also prevent the blame game. 

“I can't think of how many times I've seen a technician go over to a machine that's broken down,” Gaus says. “He says, ‘Didn't you see the oil leaking out of the XYZ?’ Well, maybe the operator did, maybe the operator didn't, but now it's too late.”

Robotics “allows you to combine those things to really help make the job better and more reliable, because you're doing those observations without having to have the human do all the hard work and the crawling and the looking,” Gaus says.

Working with robots will require new skills. Keeping humans in the loop and considering where the human sits within the AI future is “incredibly important,” Gaus says.

Humanoid robots have caught a lot of attention, too, but they introduce an extraordinary amount of technical complexity. They aren’t ready for large-scale adoption, Gaus says. Instead, the way we work with robots may change more than the form factor.

“I do think you're going to find a lot more robots collaboratively working in our workspace that will shift the way we think about work and that will help us be both more effective and efficient, but also allow the robots to have a little bit more, or a lot more, independent decision making.”

Naturally, working closely with robots also requires rigorous safety standards and safety control systems.

AI won’t solve every problem

When is AI not right for the job? In some cases, AI technology may not be mature enough, Gaus says.

“We need to have a whole new set of skills in order to properly maintain AI on devices that maybe we haven't had to worry about in the past, and so I think there's a lot of things where we have to be mindful of not just is it possible, but is it actually sustainable, maintainable and do I get the right ROI?” he says.

Another wrinkle in the financial aspect is that many cutting-edge robotics companies offer robotics-as-a-service. These services are often considered OpEx as opposed to CapEx, which changes how ROI is looked at.

“I'm not waiting for the robots to come and solve all our problems,” Gaus says. “I think we can collaboratively solve them together.”