Many process manufacturers are currently exploring some version of predictive maintenance capabilities. And most plant operators are asking the same question: Is this going to deliver real value, or is it just another dashboard?

The honest answer is that it depends less on the sophistication of the technology and more on two things suppliers don't always lead with: whether the customer trusts the vendor enough to share operational data, and whether the vendor actually understands the specific failure modes of the customer's process. This means that as artificial intelligence (AI) enters the conversation, the human-expertise piece actually becomes more important.

Processing recently connected with Mike St. Germain, Customer and Market Insights Lead, Industrial Sectors, at Watson-Marlow Fluid Technology Solutions, to discuss this topic and more.

Q: Data trust and governance have become a real part of the AI adoption conversation. What does it take for suppliers to earn the access needed to deliver real value in this regard?

A: It really comes down to what you're offering in exchange for access. If all a customer wants from us is a recommendation such as, "here's what sensors to buy,” or “here's what to watch for," they don't have to give us anything. But if they want us to actually monitor their system and tell them when to act, that only works if we have access to the data.

The challenge is getting there. We work with customers to run our monitoring through their existing SCADA system, or we ask permission to install cellular connectivity so we can pull the data ourselves. Either way, it comes down to trust. Do they trust us enough to let data leave their site?

If a company owns and runs its own pumps in-house, and its own people are the ones looking at the data, then that's not an issue since nothing is leaving the building. The friction shows up when a supplier is the one asking to see the data from someone else's plant.

What earns that trust, ultimately, is being upfront that we're only looking at data relevant for our own equipment and then proving the value (such as reducing downtime costs or freeing up plant staff from manual monitoring) quickly enough that the access feels worth it.

Q: Why is human, process-specific expertise still essential to determining what is actually worth monitoring for a given application?

A: An algorithm doesn't know what it doesn't know. AI can process signals faster than any of us and spot patterns in data we'd never catch otherwise, but it can only learn from what it's told to look at in the first place. Someone still has to make the call on what actually matters for a given process, and that boils down to a team’s experience. It comes from people who've seen enough pumps fail, in enough different applications, to know that a rising suction pressure in one process means a clog is forming, while in another process, pressure isn't the signal that matters at all.

That's the piece that doesn't go away as AI comes in; it actually matters more. If you hand a system the wrong thing to watch, it'll monitor it very precisely and still miss the failure. So there always has to be someone with real process experience deciding what gets fed to the algorithm before the algorithm can do anything useful with it. The technology executes, but the person still has to know what's worth executing on.

Q: What is the payoff of trusting a supplier with operational data?

A: The biggest payoff is avoiding downtime, plain and simple. Downtime costs vary enormously depending on the industry. We had a cement facility in Germany that was looking at six to 10 hours of unplanned downtime from a clogged suction line. By catching the early pressure signal and folding a flush into their next scheduled maintenance window instead of an emergency stop, they avoided roughly $30,000 in losses from that one event.

So the real value is about giving customers the ability to plan. An unplanned failure means pulling people off other work, calling in pipefitters and dealing with the chaos of an emergency stop. A scheduled one means you handle it on your terms. When customers see that math clearly enough, sharing the data stops being a hard ask.

Q: What's the biggest misconception customers have about predictive maintenance technologies?

A: That it's a purely technical solution. The piece people underestimate is how much human, process-specific understanding still has to sit behind it. You have to actually know a customer's process well enough to know what failure mode you're trying to predict, or the sensors you install won't tell them anything useful. AI can process the data faster than we can, but it still needs people who understand the application to point it at the right problem in the first place.