AI Use Cases in CPG: A Detailed Portfolio Transformation Case Study

AI Use Cases in CPG are best understood through the decisions they change, not through a catalog of algorithms. Consider the experience of Meridian Consumer Brands, a representative composite manufacturer created from recurring patterns in large food, beverage, household, and personal-care programs. Meridian operated 2,840 active SKUs across six markets, sold through grocery, mass, club, convenience, and e-commerce channels, and generated approximately 4.6 billion dollars in annual net revenue. Its brands were growing, but assortment complexity, volatile demand, rising trade spend, and packaging disruptions were eroding margin and service.

The transformation team used a portfolio approach to AI Use Cases in CPG, linking each model to an existing stage-gate, TPM, S&OP, IBP, RGM, or retail-execution decision. Over 18 months, the program reduced weighted forecast error by 19 percent, lowered forecast bias by 6.4 percentage points, improved case fill rate from 93.1 to 96.0 percent, and redirected 31 million dollars of trade spend. These figures describe a realistic composite case rather than the disclosed performance of any single company, but the operating lessons are directly applicable to manufacturers with the scale and portfolio complexity of Nestlé, PepsiCo, Unilever, or Mondelez International.

The Starting Point for AI Use Cases in CPG

Meridian entered the program with three connected problems. First, SKU proliferation had outpaced planning capacity. More than 38 percent of active SKUs generated less than 5 percent of gross profit, yet many required unique packaging components or constrained production time. Second, customer teams were planning promotions in separate spreadsheets and entering only approved events into TPM. Finance could see total trade spend, but RGM could not consistently distinguish true incrementality from subsidized baseline sales, cannibalization, or retailer inventory loading.

Third, the consensus demand plan reacted slowly to external signals. The statistical forecast used shipment history and limited seasonality, while retailer point-of-sale, inventory, weather, digital activity, and syndicated category data reached planners through separate reports. During packaging shortages, supply planners manually reallocated finished goods without a consistent estimate of customer-level lost sales or substitution. Reconciliation consumed most of the weekly demand review, leaving little time for decisions.

The executive sponsor rejected a broad platform-first proposal. Instead, the team established a value tree covering net revenue, gross margin, working capital, case fill rate, waste, and planning effort. Every initiative needed a named process owner, a measurable baseline, and a clear intervention. Four use cases passed the first screen: demand sensing, promotion evaluation and optimization, allocation during constraints, and portfolio rationalization.

Phase One: Building CPG Demand Forecasting AI Around Planner Decisions

The first release covered 620 high-volume SKUs for two grocery customers. The model produced weekly forecasts at customer-SKU-distribution-center level for a 13-week horizon. It combined shipment history, retailer sell-through, retailer inventory, distribution, pricing, promotions, holidays, weather, and known availability events. Rather than replacing the statistical baseline everywhere, it generated an alternative forecast and a confidence range, then highlighted combinations where the expected value of review exceeded a defined threshold.

The team designed evaluation around planning reality. Stockout periods were censored or reconstructed so constrained point-of-sale did not teach the model that demand had disappeared. Promotions were represented by mechanic, depth, display support, feature status, timing, and event history. New SKUs inherited priors from comparable products based on category, pack, price tier, channel, and launch pattern. CPG Demand Forecasting AI was measured by forecast value added relative to the existing baseline and the final consensus plan.

After a 16-week parallel run, weighted absolute percentage error improved by 14 percent at the eight-week horizon and 23 percent at the two-week horizon. Positive forecast bias fell from 9.8 to 4.1 percent. More importantly, planner touches declined by 27 percent because the exception engine suppressed low-value reviews. The team learned that automation rates were a poor target by themselves: planners needed to spend less time on stable SKUs and more time examining innovation, promotions, distribution changes, and constrained items.

Phase Two: Connecting AI-Powered Revenue Growth Management to TPM

The next set of AI Use Cases in CPG addressed promotion economics. Meridian reconstructed 46,000 historical events across three years, linking customer plans, TPM records, shipment data, point-of-sale, distribution, retailer inventory, trade deductions, and product cost. The analytical layer estimated baseline sales and separated incremental consumer demand from forward buying, cannibalization, halo, and post-promotion dips. Events with incomplete execution evidence were assigned lower confidence rather than treated as clean observations.

AI-Powered Revenue Growth Management gave account teams a range of expected volume, net revenue, gross profit, and retailer margin for candidate mechanics. The optimizer respected customer calendars, display availability, minimum event spacing, production capacity, and annual trade budgets. It did not automatically approve promotions. Recommendations entered the normal customer-planning workflow, where sales, RGM, finance, category, and supply representatives could adjust assumptions before commitment.

In the first full planning cycle, the system identified 74 events that generated volume but destroyed gross profit after cannibalization and trade cost. It also found 29 underfunded events with strong display execution and attractive incrementality. Meridian cancelled, redesigned, or shifted the first group and increased support for the second. Across the pilot categories, trade spend fell 3.2 percent while incremental gross profit rose 8.7 percent. The crucial lesson was that promotion lift alone is not a value measure; the decision must account for what consumers would have purchased without the event.

Phase Three: Using Decision Agents in IBP and Constrained Allocation

The program then moved into supply response. Packaging resin volatility and intermittent closure shortages were forcing weekly changes to production schedules. Meridian built an allocation service that ranked deployment options using customer inventory, projected consumption, distribution, substitution, service agreements, margin, strategic importance, and shelf-loss risk. Recommendations were refreshed when demand, production, or inbound-material assumptions changed.

The technology team worked with an enterprise AI agent developer to create bounded workflow agents for evidence gathering and exception routing. One agent assembled customer-SKU exposure for the supply review; another drafted allocation alternatives and documented the constraints behind them. Neither could alter the production schedule, promise customer quantities, or release an allocation without authorized approval.

Generative AI for IBP was introduced as a decision-preparation layer. Before the monthly executive meeting, it summarized changes from the prior cycle, linked demand and supply assumptions, identified unresolved gaps, and generated scenario narratives grounded in approved planning data. Leaders could trace every material statement to its source. Meeting preparation time fell from roughly 320 person-hours to 190, but the larger benefit was consistency: commercial, finance, and supply teams entered the meeting with the same version of the constraint story.

During the next major closure shortage, case fill rate for priority customers remained above 96 percent, compared with 91 percent during a similar earlier disruption. Finished-goods write-offs fell 12 percent, partly because allocation recommendations considered shelf life and substitution. The lesson was not that an agent should run S&OP autonomously. It was that agents can compress evidence collection, surface conflicts, and preserve a decision trail while accountable leaders retain control.

Phase Four: Extending AI Use Cases in CPG to Innovation and Assortment

With commercial and planning foundations in place, Meridian turned to portfolio complexity. Category and portfolio teams combined syndicated demand, household panel behavior, retailer assortment, product attributes, manufacturing constraints, packaging commonality, and SKU profitability. The model identified products whose apparent revenue contribution largely transferred from adjacent packs. It also highlighted regional items with small national volume but strong local incrementality, preventing an indiscriminate tail-SKU purge.

The analysis supported removal or consolidation of 186 SKUs and recommended investment in 43 packs aligned with distinct consumer need states. Changeover hours fell 9 percent on two constrained lines, obsolete packaging inventory declined by 6.8 million dollars, and weighted on-shelf availability improved 1.7 percentage points in affected categories. Customer teams received retailer-specific assortment narratives based on shelf productivity, category incrementality, price points, and service implications.

Generative AI for CPG also supported the early stage-gate process. Brand and innovation teams used governed consumer insights, complaint themes, sensory findings, and product attributes to develop concept territories and research questions. Formulators used it to retrieve prior learning and identify potential ingredient or packaging constraints. Generated material remained a starting point: claims substantiation, sensory research, regulatory assessment, label approval, supplier qualification, and plant trials continued through formal controls.

How Meridian Governed Scale and Measured Results

The steering team treated AI Use Cases in CPG as changes to operating processes. Each release had a product owner from the function receiving the recommendation, a data owner, a model owner, and a control owner. Monitoring covered data freshness, drift, forecast bias, recommendation acceptance, overrides, realized economics, and subgroup performance. When users rejected an output, they selected a reason such as missing event, customer intelligence, distribution change, capacity risk, or model disagreement.

The team also used holdouts and phased rollouts wherever practical. Promotion recommendations were compared with similar untreated events. Forecast improvements were assessed against both the original statistical baseline and the planner consensus. Allocation performance was measured through service, margin, waste, and customer impact rather than a single optimization score. Finance validated benefit calculations to prevent gross modeled opportunity from being reported as realized value.

Across 18 months, direct benefits exceeded program costs by an estimated 4.3 times. The largest contributions came from reduced trade leakage, lower finished-goods and packaging write-offs, and improved availability on priority items. Results varied by category: stable household products delivered the highest forecast automation, while snacks delivered greater value from promotion and assortment optimization. That variation reinforced the need to scale by decision archetype, not by copying one model across every brand and market.

Conclusion

Meridian's case shows that successful AI Use Cases in CPG connect demand signals, promotion economics, portfolio choices, and supply constraints to the forums where decisions are already made. The program created value because models entered TPM, demand review, S&OP, executive IBP, stage-gate development, and customer assortment planning with clear owners and measurable outcomes. For manufacturers ready to build on that foundation, Generative AI for CPG can accelerate evidence synthesis, scenario preparation, and controlled content generation while preserving the human approvals required for product quality, claims, retailer commitments, and financial decisions.