Five questions to ask before approving an AI purchase at a consumer products portco.
The CPG AI vendor pitch in 2026 leans heavily on Kraft Heinz, Unilever, and PepsiCo case studies that don't translate cleanly to a mid-market portco. The questions below are the ones the case studies don't survive. Ask any one of them on a vendor call and the honest answers separate real solutions from deck-only ones.
Question 01
"How much working capital does AI inventory optimization actually release for a mid-market CPG brand?"
Ten to twenty percent of inventory dollar value within 12 months of deployment, measured against a controlled baseline. For a $150M CPG portco carrying $30M of inventory, that's $3M to $6M of working capital freed. The recovery concentrates in slow-moving SKUs and seasonal categories where the legacy forecast was structurally wrong, not in fast-moving staples where the existing forecast was already reasonable.
Why most vendors get this wrong: they quote the enterprise case study (Unilever, Kraft Heinz) where the working-capital base is $500M-plus. The percentage lift is similar; the absolute dollars are an order of magnitude smaller for a mid-market portco. The investment case has to be built on the mid-market math, not the enterprise headline.
Right answer pattern: the vendor agrees to a 90-day baseline measurement on the actual inventory base, with the working-capital release computed against that baseline by an independent reviewer. If they only quote percentage lift without a dollar number for your specific portco, the projection isn't defensible.
Question 02
"What's the real ROI of AI demand forecasting and over what time window?"
Five to ten percent forecast accuracy improvement on weekly granularity within 90 days, on top of the existing baseline. Kraft Heinz with o9 published 11 percent monthly and 14 percent weekly improvement plus 20 percent safety stock reduction across 7,000 SKUs. The financial impact at portco scale is $1.5M to $8M annually depending on revenue band, mostly from working capital release and reduced markdown.
Why most vendors get this wrong: they quote MAPE improvement on the existing forecast as the ROI. MAPE improvement is not money. The money comes from reduced safety stock, reduced markdown, and reduced stockouts. Ask for the conversion model from accuracy improvement to dollars, in writing.
Right answer pattern: a model that converts the forecast accuracy improvement to working capital release, markdown reduction, and revenue recovery from fewer stockouts, with each component tied to your specific cost-of-capital and gross-margin assumptions. If the vendor can only quote accuracy, they don't know how the money gets made.
Question 03
"Which AI vendors lead in CPG trade promotion optimization, and how do we choose between them?"
Anaplan is the dominant trade promotion management platform when integrated with finance and S&OP planning. o9 Solutions leads on connected demand and supply planning at the enterprise tier. RELEX Solutions is the strongest pure-play for retail-led CPG forecasting. Tellius and other analytics specialists handle the ROI-attribution layer after the fact.
Why most vendors get this wrong: each vendor pitches their tool as a complete solution. None of them is. The right architecture for most mid-market CPG portcos is a planning layer (Anaplan or o9), an execution layer (often a different tool), and an attribution layer (Tellius or built in-house). Buying one vendor's full-stack pitch usually means accepting weaker capability on two of the three layers.
Right answer pattern: the vendor choice depends on whether your pain is planning, execution, or post-event attribution. Diagnose that first, then choose the tool. A 60-day data-and-process diagnostic with a named consultant beats a vendor RFP every time on this category.
Question 04
"Whose data trains the model, and what's the contractual line on shared learning across your customer base?"
Multi-tenant AI vendors get smarter the more CPG brands they have. Your portco's POS data, trade calendar, and retailer relationships are competitive assets. If they flow into a shared training set, the portco is paying for the privilege of educating its future competitors, including the brands sitting next to it on the same retailer shelf.
Why most vendors get this wrong: they conflate "your data is private" with "your data doesn't train the model." Those are different statements. The first is about access; the second is about model weights. Many SaaS contracts permit the second under "aggregated and anonymized" clauses that aren't really anonymous when only a handful of brands compete in a category.
Right answer pattern: a clean contractual line saying model weights derived from your data stay in your instance and don't propagate to the shared base model. For a category-leading brand the sponsor wants to sell, this matters double because the data is part of the exit asset.
Question 05
"What is agentic commerce in CPG and should we be planning for it now?"
Agentic commerce is when an AI agent acting on behalf of a consumer or a retailer's planogram team makes purchase or assortment decisions without a human in the loop. NielsenIQ's 2026 forecast says 40 percent of enterprise applications will be powered by task-specific AI agents by year-end, up from less than 5 percent in 2025. For PE-backed CPG, the brand needs to be machine-discoverable and machine-orderable within 18 to 24 months or it gets routed around.
Why most vendors get this wrong: they sell agentic commerce as an immediate buying signal. It isn't yet. The right framing for 2026 is a 2027 boardroom conversation about how the brand shows up to AI agents (structured product data, machine-readable claims, programmatic media), not an immediate platform investment.
Right answer pattern: a 12-month roadmap to make the brand's product data machine-grade across every channel, plus a 24-month plan for how the brand competes when shoppers and retailers increasingly let agents pick the assortment. Start with the data layer. The agent layer will come whether you're ready or not.