Ask a room full of food manufacturing executives who’s holding up their AI rollout, and most will point to the plant floor. That’s not where Jared Helenic sees the resistance.
As an AI Product Specialist at Infor, Helenic has spent the last three and a half years working with food and beverage manufacturers and distributors across dairy, agriculture, ingredients, and protein, on everything from new ERP builds to their existing install base. What he’s found is that the people running the line tend to welcome AI. The pushback comes from a layer most companies don’t expect.
In this conversation, Helenic walks through where AI adoption stalls in food manufacturing, why a stale dataset once told one company to overstock a warehouse that no longer existed, why food safety should keep AI as a companion rather than a decision-maker, and where a manufacturer with budget for exactly one AI project in 2027 should spend it.
Q. When you’re talking to food manufacturers about AI strategy, is there a pattern that surprises you, something you’d never expect to be a blocker to success?
Jared Helenic: Most people are already asking questions like, “Do we have the right data strategy? Is our data siloed or centralized?” Those are top of mind. What actually surprises me is the reaction we get from different levels of the organization.
At the top, executives love AI because it lets them grow without adding headcount. That’s an easy win. At the individual contributor level, people don’t mind it as much as you’d think, either. If AI takes the mundane work off their plate, like reading through documents, that’s a huge win for them.
The quiet blocker is actually the middle management layer. A lot of what AI does is analyze data, find trends, and suggest what to do next: which suppliers to buy from, what quantities to order. That work happens at the manager and director level. There’s a quiet concern among those employees that AI is taking over their function. Individual operators, on the other hand, know their jobs are safer, because someone still has to run the line and make physical decisions.
That’s the blocker I didn’t expect. I assumed it would be the people on the line. It’s actually not. They tend to be more pro-AI, because it makes their lives easier and takes away the guesswork.
Q. Are those attitudes shifting over time? When AI first came onto the scene, where was the fear then, and where is it now?
JH: In the beginning, the fear was concentrated at the individual operator level, because people assumed AI would take their jobs. At some point, once physical AI (robotics on the line) becomes more of a commodity, that could change, but we’re still years away from that, and it requires serious capital investment. That conversation about the ROI of humans versus physical AI will likely run for the next five-plus years at minimum.
Once we found the actual use cases, though, and saw who’s really affected day to day, it became clear that most of the decisions, trends, and analysis AI touches are aligned to that middle management layer. So the concern has shifted there. Now the harder conversation is: how do we make these employees more AI-fluent so they don’t feel at risk, or do we genuinely need as many layers of middle management? That’s not typically a conversation I’m involved in, but I sometimes have to kick it off.
Q. Some experts recommend starting with one narrow AI use case and scaling from there. Others say tackle a whole function, like operations or finance. What’s your take?
JH: Both camps have a valid argument, but I’d take a bit from each. Devin McCarthy, VP of IT at Eagle Family Foods, has a phrase I like: think big, start small, move fast. That sums up the right approach.
You don’t want to pick a narrow use case in a system that isn’t mission-critical, like HR. For food manufacturers, the business succeeds or fails in operations and finance, so that’s where you want to focus. But starting there directly is hard. If your goal is “increase financial workflow efficiency,” that’s a huge ask on its own.
Start small within that function instead. Document automation is a good example. There’s a constant flow of trade invoices, supplier invoices, and purchase orders. Solving that pattern paves the way for bigger use cases.
And move fast. Analysis paralysis is real: people spend six, nine, 12 months debating use cases on a whiteboard and end up with nothing. Focus on operations and finance, pick a small use case that adds value, and go. Then keep stacking wins over time.
Q. Can you give a concrete example of a manufacturer that went from a small pilot to full production and scale? What did the path look like, and what did it take to get there?
JH: Eagle Family Foods is a public reference for Infor, so I can share this. The use case we started with was forecasting. They were ingesting consumption data and spending real money on it, but their supply chain team couldn’t turn that data into better forecasts. For six months, they struggled to justify the spend. They knew the data had value, but couldn’t integrate it into their forecast to find trends and anomalies.
We worked with their team and the Infor Data Science team on an initial pilot. Could we integrate Nielsen IRI data into their forecast, at both the warehouse and item level, and improve accuracy? At the item level, we improved forecast accuracy by double-digit percentages in some cases. That’s significant on its own, but the bigger benefit was what came next. A forecast feeds into how you buy inventory. Inventory feeds into labor scheduling. Labor scheduling feeds into how you load your trucks. That one pilot became the highway for a string of additional use cases as value from AI compounds.
It started with a simple question: how can we use this data better to make better decisions, because right now we know there’s value in it and we’re not finding it? That was the hypothesis, and it held up.
Q. Now that agentic AI can act on a company’s behalf, which agents are earning trust the fastest with stakeholders?
JH: The ones earning trust fastest are the ones where you can easily check the work, and where mistakes aren’t costly. I wouldn’t put an agent in charge of fully autonomous inventory decisions right off the line. If that agent is wrong, it’s costly, not just in revenue, but potentially in human safety.
Document automation is the safest place to start. This builds on robotic process automation and optical character recognition that have existed for years: extracting information from an order form and matching it to your ERP, or matching supplier invoices to open purchase orders. That work is low risk and high volume. Most people don’t want to spend their day staring at paperwork, they’d rather do something more valuable. And if the agent gets something wrong, or flags something for a person to review, it lands in an exception folder, by design. The business keeps running, and a person is right there to catch it. That’s where I’ve seen the fastest, most reliable value.
Q. Any cautionary tales, something that went into production and caused a costly mistake, or came close?
JH: This didn’t happen at Eagle, but I’ve seen it elsewhere, mostly around inventory decisions. During data discovery with one company, they gave us historical data from two or three years back. Fine on paper, except their warehouses and distribution strategy had changed since then. We produced outputs based on data that no longer reflected reality.
When we told them to increase inventory in a location because it looked underutilized, the CIO said, “That’s because it doesn’t exist anymore.” Which was a pretty good reason. The lesson isn’t dramatic, but it matters: the data you feed AI has to be current. We all sit on a decade or more of historical data, which is great, but you need something closer to what’s happening today. Feed AI stale data and you get garbage out. Near-real-time data is what makes the output both accurate and relevant to where your business is at today.
Q. Beyond being recent, what else needs to be true about a company’s data before agents can work with it responsibly?
JH: It’s less a technical problem than a people-and-process problem. You need your procedures mapped out first, not just the steps in a process like procure-to-pay, but who owns each step and who’s responsible for saying no. There’s a framework for this called RACI, responsible, accountable, consulted, informed, and it’s worth having in place before you bring agents in. If that’s not already in place, your agents don’t have a reference point. They’ll go do their own thing, potentially in a way that doesn’t match your actual procure-to-pay or order-to-cash process, and you’ll be left wondering what happened.
If you haven’t mapped your processes, and you don’t have human checkpoints built in, it can turn into a mess. None of that has to do with the technology. It’s about understanding how your business actually operates day to day, not how it looks on a whiteboard: who does what, who says yes or no. Once that’s in place, agents get context: this is how the business actually runs, not how I think it should run. Those are two different answers for an agent to figure out.
Q. Beyond inventory, are there other use cases you’re telling food manufacturers to stay away from right now?
JH: Food safety, specifically fully autonomous food safety decisions. Given the recalls we’re seeing across the industry, I’d push hard for a clear human in the loop at every step, someone who’s actually accountable for the call. Use AI as a companion, not a decision-maker.
Computer vision is powerful. Production lines move too fast for a human to catch every defect. But when a defect shows up, don’t let an agent classify it and move on unsupervised. Have a human make that judgment call and own it. AI can flag the issue and raise it, but full autonomy on food safety decisions is a risk I’d caution against, because AI can still be wrong. And regulators don’t care that “AI told you so.” Under FSMA 204, the traceability rule coming into effect, that’s not a defense. You still need a person accountable at the point of decision. AI can be the sidecar on food safety. It shouldn’t be the driver.
Q. Vendors promise AI can compress the experience gap, so a new hire performs closer to a veteran, faster. What holds up in practice, and what’s still fuzzy?
JH: What holds up is pattern detection within documented workflows. A new hire doesn’t know what they don’t know. AI is genuinely good at getting them from that point to a working knowledge of “here’s how we do things,” whether through Copilot or a chatbot they can ask questions of. That works well because it’s rooted in something already documented and approved.
What partially holds up is diagnosis. AI can surface issues, similar to the food safety point, which compresses the time it takes to understand how a line operates and where its quirks show up. What doesn’t hold up is judgment. AI won’t know that every third Tuesday, flour from a particular supplier runs a bit more moisture, so you need to bump your oven temperature by 30 degrees. That kind of judgment call still falls on the human, and so does the responsibility for the outcome.
In summary, AI is strong at surfacing trends and getting new workers onboarded to documented procedures. It’s decent at flagging outliers and exceptions. It falls short on judgment calls and knowing who should do what in a given scenario, because unless that knowledge is digitized and measured, AI doesn’t have the context to make that call without a human in the loop.
Q. If a food manufacturer has budget for exactly one AI project in 2027 and they’re starting from scratch, what would you recommend, and how would they know a year later if it worked?
JH: For a company starting from scratch, document automation. So many documents run through a business. If you’re not scanning them, capturing the information digitally, and putting it into a centralized repository where you can layer a Retrieval-Augmented Generation (RAG) model on top to ask questions of your own documents, that’s a low-risk, high-reward starting point.
One of the values is, at the end of the year, you can ask which suppliers have the friendliest payment terms, and more importantly, which ones penalize you hardest for paying late. Some suppliers charge no penalty at all, which means you can hold that capital an extra week and put it to work elsewhere. Having that information digitized and available for an agent to learn from is super valuable.
That said, a lot of companies have already digitized their documents. If you’re one of them, I’d suggest what Eagle Family Foods did: start with demand forecasting. A better forecast cascades into nearly every part of the business, from inventory to labor scheduling to trade spend decisions. Ingest as much relevant data as you can, and that cascading effect is close to limitless.
Not sure where your one AI project should go? Infor’s ebook, The Simplest Path to Achieve Value With AI, breaks down how to pick that first use case and turn it into a repeatable win.
Q. AI is moving fast. A year from now, what part of running a food plant will a leader no longer have to do manually?
JH: It’s going to look different for everyone, but if a company has digitized its documents and has a forecast running on machine learning, I think the start of a plant operator’s day will already be laid out for them. Yesterday’s OEE, where the downtime happened, who’s scheduled to work today: all of that will show up in a single report, because that information is already in your MES, your ERP, and your warehouse management system.
Once you train agents on the key data points that matter each day, I think operators, managers, and directors will simply wake up to an email: here’s what happened yesterday, here’s what happened on the previous shift, here’s what to focus on today. That’s the shift I expect, how people start their day, and how much of that ramp-up gets compressed. I don’t think it’ll take three hours into a shift to learn a line went down. AI can help close that gap, today and in the near term.
