Picture a 70-year-old gum company, the one whose comic strip wrappers you collected as a kid, deploying agentic AI workflows to forecast demand, develop new candy formulations, and deal with tariff-driven supply chain disruption. That's Bazooka Candy Brands in 2026, and it's one of the more instructive AI stories you'll hear this year, precisely because it has nothing to do with Silicon Valley.
Sankar Karuppasamy, Bazooka's CIO, has quietly built one of the more thoughtful AI deployments in consumer goods. His team uses Anthropic's Claude in agentic workflows across demand planning, formulation development, and supply chain decision support. The results are concrete: Bazooka pushed its forecasting accuracy from around 60% to 90%. A 30-point accuracy gain rewires how a manufacturer plans, buys materials, and staffs production lines.
Mid-market executives in manufacturing, food and beverage, distribution, or any capital-intensive industry should pay close attention to this one. Not because Bazooka is flashy, but because it isn't.
The "We're Not a Tech Company" Problem
Most AI case studies are written about companies that were already technical. Google, Salesforce, Goldman Sachs. The implicit message is that you need a large engineering organization, machine learning PhDs, and years of data infrastructure work before agentic AI makes sense for your business.
Bazooka's story blows that up.
The company makes Ring Pop, Push Pop, Baby Bottle Pop, Juicy Drop, and its namesake bubble gum. Founded in 1938, it manufactures candy for retail shelves globally. This is a $320M business operating in 8 countries, not a software company with a data moat. Karuppasamy and his team got AI working across core business functions and produced documented, measurable results.
RSM's 2025 AI survey found that 91% of middle-market firms are now using generative AI in some form. But 92% of those companies also hit major roadblocks during rollout, and 53% said they felt only "somewhat prepared" to implement it. The challenge isn't access to AI tools. It's knowing what to do with them and where to start.
Bazooka's CIO had a clear answer: start with your dirtiest operational problems.
Step One Was Not AI. It Was Data.
Before a single Claude prompt was written, Karuppasamy's team did the hard work that most executives skip. They got the data house in order. Bazooka consolidated fragmented data across six different business functions onto a single connected planning platform, using Anaplan as the foundation.
That's unglamorous work. It involves convincing finance, supply chain, operations, sales, and demand planning teams to share data and align on definitions. What counts as a confirmed order? How do you handle promotional lift? When does a SKU get retired? These are the debates that happen in conference rooms before any model sees a single row of training data.
The biggest blocker to AI-driven demand forecasting is almost never the AI. It's fragmented, inconsistent, siloed data. Fix that first, even if it takes 12 to 18 months of integration work. The AI deployment becomes faster and more reliable once the plumbing is clean. Bazooka's data foundation work is what made everything else possible, and it's the step that most executives either skip or underestimate.
Use Case 1: Demand Planning That Actually Works
Seasonal demand is one of the hardest planning problems in candy. Bazooka's calendar is dominated by Valentine's Day, Easter, Christmas, and Halloween. Miss the demand signal by 10-15% in either direction and you're either leaving money on the shelf or destroying margin with markdowns and excess inventory.
Karuppasamy's team built ML-driven forecasting into its connected planning platform to predict retailer demand for seasonal products based on historical patterns. With Anthropic's Claude layered in as an agentic component, the system doesn't just produce forecasts. It supports operational decisions on top of those forecasts.
The jump from 60% to 90% forecasting accuracy is the headline number, but the downstream effects matter more. Better forecasts mean more precise manufacturing plans. More precise manufacturing plans mean less waste, better working capital management, and stronger fill rates with retailers. In a thin-margin business like candy, those points add up fast.
McKinsey's analysis of AI in CPG found that a well-executed digital and AI transformation can deliver 6-10% incremental revenue uplift and 3-5 percentage points of EBITDA improvement over three to five years. Demand creation and demand management consistently rank as the highest-value domains in that analysis. Bazooka's numbers fit.
For manufacturers and distributors thinking about where to start: demand forecasting is usually the highest-ROI first use case. The data is often cleaner than you think, the business question is well-defined, and the financial impact is directly measurable.
Use Case 2: Faster Formulation Development
This is the use case that gets less attention but is more interesting for the long run.
Developing a new candy formulation, or reformulating an existing product for natural colors, new ingredients, or changed supplier specs, is a labor-intensive R&D process. Ingredient ratios, texture outcomes, shelf life testing, regulatory compliance, supplier ingredient availability. Each iteration requires time in the lab, review cycles, and cross-functional sign-off.
Karuppasamy identified formulation development as a priority use case for AI at Bazooka. The team is using AI to speed up the research and recipe development cycle, running faster iterations, surfacing relevant precedents, and giving food scientists more leverage per hour of work.
This maps to what McKinsey has documented more broadly: AI in R&D can double the pace of innovation and potentially unlock up to half a trillion dollars in annual value globally. For CPG companies specifically, McKinsey cites proofs of concept showing up to 30% reduction in time spent on product research.
The food and beverage industry is notoriously slow to adopt technology. Only 2% of F&B brands describe themselves as fully digitized, according to recent industry surveys. Any company that moves faster on AI-assisted R&D gains real competitive ground, particularly in markets where reformulation demands come every year, not every decade.
Use Case 3: Supply Chain Decision Support Under Fire
Bazooka didn't build its AI infrastructure during a calm period. It did it while navigating serious operational turbulence.
Tariff pressure on candy manufactured in China and other Asian countries forced Bazooka to rethink its entire supplier network. The company made four core changes to its supply strategy: it shifted from a pure price-for-volume model to a genuine partnership model with suppliers, committed to tracking supplier performance with defined KPIs, mapped its full value chain to identify waste and cost savings, and began formally recognizing supplier wins to build collaborative relationships.
Karuppasamy piloted agentic AI specifically on freight management and tariff analysis within this context. When supply chain decisions are this complex, with multiple variables shifting simultaneously, scenario planning required across geographies, and the cost of a wrong call measured in millions, an AI system that synthesizes information and supports rapid decision-making is worth a lot.
This is also where the switch to Anthropic's Claude becomes notable. Karuppasamy moved his team away from OpenAI's ChatGPT toward Claude, deploying it with governance guardrails and business outcome testing. Not choosing a model based on hype, but validating business results before scaling.
Agentic AI is particularly valuable in supply chain because the work is naturally multi-step and decision-heavy. An agent that can pull freight rates, cross-reference supplier lead times, model tariff scenarios, and draft a recommendation is doing the work of an analyst in minutes rather than days.
What Made This Viable
Bazooka's deployment didn't happen because Karuppasamy had unlimited budget or an engineering team of 200. Four factors made it work, and each one any mid-market operator can copy.
The starting point was operational pain, not an AI strategy. Demand accuracy was broken. Formulation cycles were slow. Tariffs created supply chaos. Karuppasamy identified those as the problems to solve, then found AI tools that addressed them. Most enterprise AI initiatives do the opposite: they acquire a tool and then hunt for use cases.
The data foundation investment was the longest part of the journey, and the most critical. Consolidating six business functions onto a shared platform, with shared definitions, took significant organizational effort. Without it, agentic AI has nothing meaningful to act on. This step deserves more budget and more patience than most companies give it.
Karuppasamy didn't deploy AI everywhere at once. He piloted in freight management and tariff analysis, built governance guardrails, tested business outcomes, and scaled from there. Promising experiments tend to die in pilot purgatory when they aren't tied to a measurable outcome from the beginning. The governance layer and the business metric have to come before the scaling decision.
Finally, the shift to Claude wasn't arbitrary. It reflected a deliberate choice to use a model suited to agentic workflows and enterprise governance. Tool selection based on business requirements rather than marketing momentum is a discipline that matters more as these systems get more deeply embedded in operations.
The Broader Pattern: Non-Tech Companies Are Getting There
Bazooka isn't alone. Unilever, Newell Brands, and Levi's are all building agentic AI workforces, an approach Consumer Goods Technology has described as "internal multiplicity," where AI agents augment human teams across functions. McCormick & Co. is running AI-driven autonomous operational planning with OMP. Century Pacific rolled out AI-powered demand and supply planning with Blue Yonder.
Agentic AI in CPG is particularly powerful because the industry runs on interconnected decisions. A demand signal triggers procurement triggers production scheduling triggers logistics. When any part of that chain is off, the ripple is expensive. An agent that monitors and optimizes across that chain continuously, not just when a planner runs a report, is a different category of operational leverage.
A Framework for Non-Tech Operators
If your business has a demand planning problem, a slow R&D cycle, or supply chain complexity that burns analyst time, the ingredients for a Bazooka-style deployment are probably already there.
Identify which operational bottleneck costs you the most, in dollars, not gut feel. If it's forecast accuracy, that's your first workflow. If it's product development cycle time, that's where you pilot. The specificity matters because it gives you a success metric before you spend anything.
Run the math on that number before any tool conversation. A $200M manufacturer that improves forecast accuracy by 10% might avoid $5M in excess inventory and markdowns annually. That's your business case. Present that figure, not a vision of the future.
Before hiring a data scientist or signing an AI contract, assess the data situation in that area honestly. If it's fragmented or inconsistent, the integration work comes first. Bazooka spent roughly 12 to 18 months on the data foundation before the AI layer went in. That timeline is probably not going to shrink for most mid-market manufacturers.
The "we're not a tech company" objection had some merit five years ago, when deploying AI in a non-technical environment required substantial custom engineering. That gap has largely closed. The tools are more accessible. The implementation patterns are better documented. The business cases are published, including in Bazooka's Fortune profile, which ran not as a tech story but as an operations story.
If your operations problems are painful enough, the data investment pays. Bazooka's answer was clearly yes. For most mid-market manufacturers, the honest answer probably is too.