AI as an Early Warning System

This interview is part of our extended Q&A series exploring AI and food safety.

AI is beginning to change when food safety teams act, not just what they do. Instead of confirming problems after they happen, companies can now identify patterns that suggest risk is building. In this Q&A, Jeff Varcoe, vice president of quality assurance and food safety at The J.M. Smucker Co., explains how AI is being used as an early warning system and why it should guide decisions, not replace them.

How would you describe the current role of artificial intelligence in food safety systems? Are we still in the experimental phase, or is it already becoming operational in some parts of the industry?

In general, I think AI is still experimental. The questions you’re raising are real—how much weight do you put on AI versus the human who has years of experience and more context than AI does? In many cases, it’s still new, and we’re trying to figure it out.

That said, there are some areas where it’s becoming more operational. That’s more on the outside look, where there’s less internal risk to a department or team. If you look at recalls and situational issues in the global market, that’s where you’re starting to see AI being used more operationally and starting to drive value.

Internally, though, at least for us, we’re still experimental. There are smaller pockets where we’re using AI in food safety, less for predicting issues and more for demonstrating compliance—for example, documentation. We have thousands of suppliers and hundreds of co-manufacturers, and you can imagine all the documents coming at you. So, we are using AI more on the rudimentary compliance side. Did they check yes here, how do you map that, and so on. So, it’s kind of all over the board depending on how you look at it.

Where do you see AI potentially generating measurable value in a food plant?

Predictive defects and environmental situations that could result in a contaminating event—and I’ll say contamination broadly, not necessarily just microbial. What runs through my head is maintenance activities, preventive maintenance, [and] sanitation activities. Can we get to a point where we can predict foreign material issues or failures? That’s where the value is going to be driven.

I don’t think we’re going to lose the human interaction, at least not in my lifetime or the next generation because [humans are] always going to have context that AI doesn’t have. AI may be helpful in predicting, but the value is in seeing issues before they occur and being able to look around the corner.

At the end of the day, there’s going to be accountability. Right now, we’re very reactionary. I take a swab, I get a result, and then it’s, “Oh my gosh, it’s bad.” If I know ahead of time that environmental conditions are trending in a certain direction—traffic patterns, prior positives—that’s where we’ll be. That’s how we’re starting to look at it.

How do you think about AI in the context of environmental monitoring and recurring issues like Listeria?

That’s really what’s behind my comments on both foreign material and the environment.

What are my sanitation activities? Are we using quaternary ammonium boosts and all those other inputs? You can start putting all that data together, and all of a sudden the system can say, “Hey, you’re trending in the wrong direction.” Ultimately, it should save time and money, whether that’s by responding to an issue or preventing one.

It’s giving me indicators that we’re headed toward something bad—but before you get to catastrophe, you should have reaction criteria built in.

Jeff Varcoe

The J.M. Smucker Co.

How should companies think about using AI in operational decision-making, such as when deciding whether to shut down a line?

For me, AI should be likened to statistical process control. It’s giving me indicators or feedback that we’re headed toward something bad. But before you get to catastrophe—and before you get to your million-dollar problem—you should have reaction criteria built in.

If we go up one sigma or one statistical deviation, here are the actions we start taking. There may be other thresholds as well.

That’s where AI can help, acting as a kind of statistical process control for whatever process you’re managing. Because if I have to shut down due to a catastrophe, I’ve already paid for it. If I can prevent that, that’s where the payoff is.

Food safety systems have historically relied on verification and testing. How does predictive analytics change the way companies approach risk?

It gives you a more measured approach to managing risk. We’ve got a couple of pilots we’re starting to put together, and one of the questions is: Do I still need a person running to the line or swabbing every X amount of time? Or do I have a verification process for the AI itself within my food safety plan?

We rely heavily on time-based verification—“do it at two hours, do it at four hours”—but bacteria don’t care about time. Foreign material doesn’t care about time. It should really be based on activity or a driver of activity. Maybe I need to verify after a maintenance activity or a shutdown. That’s where I think AI is going to help us—moving away from time-based controls toward activity-based controls.

If an AI system fails to detect a contamination risk or produces a false signal, how should responsibility be handled?

I think it’s addressed through root cause investigation and how much diligence you put into making the process work. If AI is part of your work process, then it’s no different than a failure from a lab result or an online piece of equipment. You have to understand why.

Root cause is going to be central. If AI misses something, it could be because it wasn’t trained properly or because there’s a variable influencing it that wasn’t accounted for.

Where should human judgment remain central, even as AI systems become more sophisticated?

Humans are still essential throughout the process. This is just a better tool in the toolbox.

For the foreseeable future, you’re going to need humans evaluating the data and making the risk decisions. AI doesn’t have the full context of everything going on in a facility. Accountability is still going to come back to people—how well you understand the system and how well the process is designed.

Ultimately, when something fails, it’s not just the person or the system—it’s the work process. What failed in that process?

How do you think about AI in relation to the human errors that already drive many food safety problems?

The food industry is still heavily reliant on people in quality. That means we’re going to have issues driven by human judgment, usually with good intent, but still mistakes can occur.

That’s where I see the value of AI. I want to reduce reliance on administrative controls that depend on people and instead bring in more technology and innovation. If I can minimize crises and reduce the number of decisions people have to make, that’s where the value is.

We’re not that sophisticated yet. I always think back to flipping burgers at McDonald’s. You hit a button, and it tells you exactly what to do. Food safety systems aren’t there yet.

Is there anything else food safety professionals should be thinking about as AI becomes more integrated?

We’re going to start seeing more situations where data tells us a trend is developing, even though food safety hasn’t been compromised yet. The question becomes: How do you react to that? My view is that AI should be treated as an early warning indicator. We already have early warning indicators in place—AI is just a more advanced version. The response should be the same: Recognize that food safety is still intact but take action in advance based on the signals.

This interview has been edited for clarity and brevity.