How CPG Leaders Are Making AI Work in Practice: Adoption, Data, and Human-in-the-Loop
Insights from Altria, General Mills, and Bel Group on making AI practical in CPG: adoption, data readiness, and keeping humans in the loop.
Key Takeaways:
Design for adoption, not demos:
embed AI into existing workflows so usage isn’t optional.
Treat trust as a deliverable
governance, training, and guardrails make outputs usable.
Don’t wait for perfect data
Start now, then strengthen ownership and processes.
Keep humans in the loop
oversight and customer context remain essential.
AI in consumer goods is past the “someday” phase. The real question CPG teams are tackling now is simpler—and harder: how do we make AI useful, trusted, and adopted in the flow of day-to-day work?
This post distills practical takeaways from a panel moderated by Ben Peart (XTEL) featuring leaders from General Mills and Bel Group—speaking candidly about vision, governance, data realities, and what actually drives adoption.
If you want to know where the world is heading, follow the money. That was the framing Ben Peart, Director of Sales, North America at XTEL, used to open a recent panel on the state of artificial intelligence in consumer goods. Citing a KPMG study, Ben Peart set a clear tone for the conversation: “Over $330 billion has been invested in AI in a single quarter. That’s doubling the total investment in one quarter. AI is no longer what’s coming; it’s here.”
To unpack what that actually means inside large CPG organizations, Peart was joined, at the 2026 Promotion Optimization Institute Spring Summit in Chicago, by leaders at different stages of the journey: Gaurav Yadav, Senior Director, Digital and Technology at General Mills; and Karine Bergeron, Global Director, Business and Commercial Growth Capabilities at Bel Group.
The result was a refreshingly candid discussion that moved past the hype to the hard realities of vision, data, and adoption.
Three visions, one direction
Each organization framed AI through its own lens.
General Mills has been building AI muscles for years. Yadav described a deliberate, three-pillar strategy spanning core AI and machine learning, generative AI, and the topic of the day, agentic AI. The company was “probably one of the first companies to launch our own generative platform internally,” and Yadav previewed a launch that would replace the corporate intranet’s search box with AI: “anything that you search will be AI-driven.”
Bel Group positions AI as a force multiplier rather than a replacement. “The way we view AI is it’s not to neglect human judgment, but it’s really as an accelerator,” said Bergeron, describing a tool to speed augmented decision-making and free teams “to be focused on more strategic tasks.” Critically, Bel is “working on two fronts”: capability building alongside “upskilling our people, and change management and adoption.”
The adoption gap
Peart cited a sobering statistic: across organizations, “only about 30% adoption rate is actually taking place.” The challenge isn’t enthusiasm; it’s friction and trust.
General Mills‘ answer is to meet users where they already are. “For a user to go to a generative platform, it takes effort to go there,” Yadav noted, so the strategy became removing the effort entirely. By embedding AI directly into the intranet search box, the math changes: “100% of people have to use it and will use it. There’s no other way.” As Peart summed it up, “If you only give the group or the organization one option, that’s forcing adoption.”
Bergeron‘s research points to a different root cause. After running surveys, her team found resistance was “not by lack of interest or lack of curiosity,” but rather “relating to a lack of confidence or a lack of trust.” The remedy is targeted, function-specific training and use cases that make AI “relevant to the people,” because, as she put it, “managing change and people… one cannot succeed without the other.”
The myth of perfect data
When Peart asked whether perfect data exists, the answer was unanimous and instructive. “There’s no such thing as perfect data, but I think you can’t wait till you have perfect data before you start using AI,” said Bergeron, who emphasized that breaking down silos through “cross-functional collaboration” and strong, embedded processes is what ultimately “increase[s] confidence in the outputs.”
Yadav agreed, noting the sheer scale of data now flowing into a company selling in about 92 countries. His prescription: “Keeping that as a baseline, now work on how we make it better,” using agents and clear ownership of “what data moves from where to where.” The underlying philosophy was simple and recurring throughout the panel: get started, then improve.
As Peart distilled it, “It doesn’t matter where you start, but start… if you’re not already doing AI, you are way behind. No longer are we talking about what if, it’s how well.”
Humans aren’t going anywhere
Asked whether AI might one day run the show entirely, the panelists were aligned. “Coming from a business standpoint, you’re always going to need the human relating to the customer,” said Bergeron.
Yadav offered a more provocative, tech-forward view, pointing out that in investment banking, “40% of trades are not touched by humans,” while conceding that customer- and consumer-facing work will remain human-centered far longer. The shift, all agreed, is in how people interact with technology, not whether they do.
What’s next
The use cases that excited the panel most were squarely operational: closed-loop promotion management that tracks sell-in against the original plan, AI-driven price-pack architecture simulations, and autonomous replenishment agents that detect low warehouse stock and place orders across ERPs.
When Peart polled the room on the single most important enabler, data, leadership, or cross-functional alignment, data drew the most hands, validating XTEL’s own conviction. “Data at the foundation is what we start with,” Peart said, describing the company’s approach of harmonizing data across pillars before layering on AI. As one of XTEL’s core beliefs and the panel’s shared message make clear: you can’t engineer perfect conditions before you begin.
You simply have to start.