The Bottleneck For AI In CPG: Data Reconciliation, Trust And Context
Samuel Martinez is the managing director at SDG Group USA.
gettyConsumer packaged goods (CPG) companies are racing to use agentic AI for pricing, promotion and replenishment decisions. Adoption is still early, though: Only 17% of organizations have deployed AI agents so far, compared to the 60% expected within two years, according to a 2026 Gartner report.
The gap between that ambition and today’s reality is in the data a CPG agent needs, which moves at different speeds, and few organizations have built the infrastructure for an agent to tell the difference.
The data an agent needs doesn’t just live in different systems; it also flows at different speeds and frequencies. For example, syndicated market data is typically reported on a two-week lag, reflecting how the current POS collection and aggregation process works. Syndicated data carries an additional limitation: It’s built from participating retailers and is only projected to represent the full market, so it doesn’t fully capture reality. Some retailer-direct platforms, however, are moving in the opposite direction.
Every serious CPG company is already reconciling this with people doing the work manually. The problem isn’t awareness; it’s that few organizations have built the semantic data layers necessary to send temporal metadata and lineage context to their AI agents, so the agent knows how to properly use and balance these inputs. Without that layer, an agent will surely reason across data that was never meant to be compared directly or ignore coverage nuances.
Consider a typical CPG brand attempting to integrate shipments, retail distribution withdrawals and POS consumption data to support more advanced agentic reasoning. Each data set comes from a different source, in a different format and on a different refresh cycle. Shipments reflect what left the manufacturer’s warehouse, withdrawals reflect what moved from the retailer’s distribution center out to individual stores and consumption reflects what shoppers actually bought at the register. Getting an agent to reason correctly across all three takes real, sustained work: building a data layer that carries freshness and lineage metadata, consolidating and harmonizing disparate data sources, and training an agent to interpret what each source represents and how to integrate the data to drive insights and decisions.
McKinsey research found that roughly 72% of U.S. trade promotions lose money, a reminder that this is a hard problem even for experienced teams working with full context and human judgment. The causes are varied, and it’s precisely why adding agentic AI into the mix aggravates the problem. Agentic AI can genuinely increase productivity, but without the right data foundation and context, it can generate confident recommendations that may be wrong. Human-in-the-loop review isn’t a temporary fix; it is the foundational boundary that allows teams to scale AI adoption safely without handing over the keys to pricing strategy.
Leaders need to move beyond governing agents in the compliance sense and start asking the following: Does your organization know how fresh a data input is, how completely it covers reality and where it came from? And more importantly, does your agent know that, too?
• Treat data feeds as domain data products. Wrap raw POS, shipment and inventory streams into purpose-built data products. Each product should publish metadata alongside its payload (such as refresh cadence, known lag metrics and coverage confidence scores), so downstream AI engines receive full context.
• Embed temporal metadata into the semantic layer. Ensure your semantic data layer or RAG framework reads lineage and freshness tags before executing reasoning. If an agent evaluates a syndicated file, the system must weight that input as a broad trend signal rather than as a trigger for real-time inventory ordering.
• Establish exception-based escalation boundaries. Define clear variance thresholds where an agent must defer to human judgment. For instance, if real-time store withdrawals diverge from projected POS consumption beyond a set tolerance, the system should route the recommendation to a category manager rather than executing an automated adjustment.
The common thread underneath all of this is context. Agents need not only clean data but also context on how to interpret it correctly: what each source represents, how current it is and how much weight it deserves relative to everything else feeding the decision.
There is a simple test you can run before launching your next pilot. Ask yourself the following: Would you trust this agent’s output to define a price, a promotion or a shelf recommendation without a human checking it first? If the answer is no, the gap probably isn’t in the model. It’s in the data and the context you’ve given it.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
