AI’s speed in supply chain frees up time

KANSAS CITY — Artificial intelligence (AI) quickly gathers information from commodity reports, production lines and customer orders, opening up more time for buyers and purchasers. More advanced AI programs may give recommendations, too, but successful AI implementation in supply chains will require the right program or programs, the correct data and human oversight.

AI reduces the time needed to gather information and maintain spreadsheets, said Erin Nazetta, an executive adviser for agriculture, commodities and global risk.

“It is freeing up time for buyers, purchasers and the decision-makers to focus a little bit more on relationships and decisions,” she said.

Nazetta, who has experience across investment management, banking and global agribusiness and previously worked for Rabobank and Bunge, spoke at the Sosland Purchasing Seminar, held in June in Kansas City. She electronically fielded a poll on AI in a session that had over 100 attendees. Forty-eight percent said their companies were experimenting with AI through pilots with no commitment, compared with 24% not using AI yet, 18% deployed in real workflows and 10% embedded in daily work.

“Things are moving fast,” Nazetta said. “But again, in our industry, which is still very physically driven and (where) some pieces of the information and the processes are not digitized, it’s a little bit slower than in other industries, I would say, to be able to adopt AI.”

Research from Aptean, a supplier of vertical AI and industry-specific software, showed that 23% of food and beverage organizations said that AI was essential to their workflows and decision-making, said Katherine Parr, senior food and beverage solutions consultant at Aptean.

“Daily commodity reporting is exactly the kind of activity that’s moved from a once-a-week manual pull to something people expect AI to help with every single day,” Parr said.

AI may track baking ingredients such as flour, sugar, cocoa and butter or other fats, she said.

“AI-supported reporting can help a bakery track how a flour lot’s protein content or a swing in cocoa prices are likely to affect dough yield or formulation cost, so a team can see that impact before it hits the production schedule, not after,” Parr said.

AI may reduce the time that lines are shut down to clean out allergens.

“You don’t want to spend all your shift time with your lines down for washouts,” Parr explained. “You want to start with the least allergenic items and then move on to the items with allergens. This can also be relevant for colors of items or flavor profiles. Having an AI tool can help intelligently schedule items in an order that helps minimize those changeovers and downtime.”

Weather and tariffs are the next frontier for AI in sourcing, she said.

“Predicting exactly how a drought affects a wheat crop, or how a new tariff schedule hits one specific imported ingredient, means layering external weather and trade data on top of operational data,” Parr said. “That’s the direction this is heading, and it’s where I expect the most value to unlock next for buyers managing ingredient risk.”

AI programs: a closer look

Large language models (LLMs) and agentic AI are two examples of programs in the AI realm.

“So the way I’d put it is, a large language model is reactive,” Parr said. “You ask it something or give it a prompt, and it gives you an answer, a summary or an analysis right then. Agentic AI actually goes and does something. It can run a multi-step process on its own, like watching a vendor’s on-time performance, flagging when it slips, and routing an approval to the right person, with a human still signing off along the way.”

LLMs are fast and conversational with minimal setup, she said, but agentic AI would be better for tasks that need to happen the same way every day, such as flagging inventory nearing expiration and routing it for markdown.

“That distinction between technologies matters,” Parr said. “Our research found 63% of food and beverage organizations are using general-purpose AI, but only 46% have put in AI built specifically for their industry.

“The ones using industry-specific tools are consistently more likely to see improvements in competitive positioning, workforce morale and forecast accuracy. A general-purpose tool can answer a quick question just fine. It’s the industry-specific, workflow-aware tools that actually move the business metrics that matter.”

Aptean compared generic AI programs with industry-specific AI programs. Forecast accuracy, after switching to a generic AI program, increased by 19%, but switching to an industry-specific AI program increased accuracy by 29%.

Nazetta said industry-specific platforms might cover commodities or lengthy legal documents.

“What is it that you’re trying to solve, and what’s the best tool to get the best outcome for that task?” she said, adding, “The winners aren’t necessarily going to have the best model. The winners are going to be able to combine their proprietary data with their commercial judgment and have the ability to act on it.”

Anyone selling full automation without a human checkpoint is selling risk, said Marc Losito, vice president of regulatory solutions for FoodChain ID.

| Photo: ©ARSENII – STOCK.ADOBE.COM

Two models in one

An LLM model answers a question, and agentic AI takes an action, said Marc Losito, vice president of regulatory solutions for FoodChain ID, which helps customers mitigate supply chain risk, strengthen audit-ready compliance and accelerate product innovation.

The company has built FoodChain ID Scout to do both. The LLM layer reads and generates language: It summarizes human-curated regulatory and commodity data or drafts a report when asked, Losito said. The agentic layer chains multiple steps together without a person prompting each one. The agentic layer pulls the data, checks it against a proprietary rule set, flags the exception and routes it to the right person.

FoodChain ID Scout connects proprietary data from FoodChain ID with partner data and public data streams such as commodity futures, weather, and geopolitical and trade feeds.

“AI does the connecting work across all of it,” Losito said. “Most solutions confirm a food safety problem after it shows up as a border rejection or a failed audit. Scout is built to catch the earlier signal — a fertilizer market shock, a crop forecast, a tariff change — and map it against the documented fraud and compliance patterns that follow, specific to a company’s own ingredient and supplier portfolio.”

AI turns commodity tracking from a manual literature review into continuous, intelligent monitoring.

“Instead of an analyst pulling CME futures, USDA WASDE data and weather feeds by hand every morning, FoodChain ID Scout ingests all of it in real time and flags deviations outside historical norms,” Losito said. “The report stops being a data dump and becomes an exception list: what moved, why and what it means downstream. Purchasing teams get the hours back and catch signals days or weeks before a manual scan would surface them.”

He gave an example of FoodChain ID Scout analysis after the Strait of Hormuz closed due to US conflicts with Iran. Since the strait carries roughly a third of global seaborne fertilizer trade, Chicago Mercantile Exchange urea futures spiked within weeks, Losito said.

“For the market, that was read as an energy story, not a food story,” he said. “But fertilizer scarcity forces farmers to cut application rates and shift acreage away from fertilizer-intensive crops, which tightens wheat, corn and soy supply, the exact inputs behind flour, starches and syrups in baked goods.”

Tightened grain supplies historically trigger economically motivated adulteration such as dilution, species substitution and mislabeled origin, he said.

“AI connects a commodity futures shift to that documented fraud pattern and gives a bakery’s sourcing team a three- to four-month head start on supplier requalification and incoming testing, instead of finding out from a border rejection,” Losito said.

Scout also is designed to watch weather anomalies against crop calendars, tariff announcements against harmonized tariff schedules and currency moves against sourcing geography.

Avoiding “garbage” data

The need for accurate data is found in a term heard often in the AI industry: garbage in, garbage out.

“So, I would say this comes back to something I tell people all the time: You don’t get to good reporting without a good data backbone first,” Parr said. “A lot of the businesses I talk to are still on QuickBooks or spreadsheets. So, before you can even think about what AI can do with commodity pricing, you need your data in one place. Once that’s true, a planner can ask a plain-language question, like which ingredient costs moved the most this week, and get an answer instead of pulling numbers out of five different systems.”

An Aptean 2026 survey found that 81% of food and beverage decision-makers said data quality and access was the biggest challenge to successfully implementing AI.

Nazetta added that data might not be perfect.

“If you have a junior analyst and still you want to go in and double-check their work, I think this is probably a similar situation,” she said.

Humans needed

After implementing AI programs, companies still need buyers, sellers and hedgers, Nazetta said. AI pulls the data together quickly, “but at the end of the day, you still need that experience and that wisdom (that) comes from within the industry and understanding what questions to ask and then ultimately being the one that pulls the trigger.”

Parr said AI can find a risk or make a recommendation, but planners or buyers can make the final call on financial decisions like purchasing commodities.

“This could look like AI making purchasing suggestions based on market conditions, vendor compliance and lead times,” she said. “Then, an experienced person can take a look, then confirm those suggestions. What we want the AI to do is take on some of the effort and time it takes to compile all those outside factors and then have a person spend time on confirmation and action.”

Generic AI is fast at pattern detection but is “tragically poor” at judgment, Losito said.

“It can tell you a signal is unusual,” he said. “It cannot tell you with certainty what it means or what to do about it, especially when the decision touches supplier relationships or regulatory nuance that shifts by jurisdiction. Every credible AI deployment in this space keeps a human in the loop for validation and the final call. Anyone selling full automation with no human checkpoint is selling you risk.”