Most software for consumer goods has been organized around tools. You buy a route planner, a shelf-audit app, a deductions module, a forecasting engine, and a field-sales kit, each from a different vendor, each with its own data model and its own idea of what a "store" or an "order" is. The result is familiar to anyone who runs a distribution business: a drawer full of point solutions that do not talk to each other, and a route to market that looks integrated on the org chart but fragments the moment information has to move from the shelf to the order to the truck.
We think the right way to organize an AI ecosystem for FMCG is not a longer feature list. It is a taxonomy. So we built one: sixteen agent types, grouped into five families, spanning the full route to market. Every solution in the FMCG Cloud marketplace classifies under exactly one primary type, and every solution must earn the FMCG Verified certification to carry that label. This post explains the structure and, more importantly, why a structure is the point.
Start with the unit. We organize around agents rather than apps because an agent names a job to be done, not a screen to look at. An agent suggests the order, audits the shelf, plans the route, or protects the margin. It acts and it explains what it did. A feature list answers the question "what does this software have." A taxonomy answers the more useful question for a route-to-market leader: "what work in my business can now be done by software that acts." Those are different questions, and only the second one composes into a system.
The five families follow the shape of the business. The Commercial family covers the four agent types closest to revenue: the Order Agent that suggests and builds orders, the Promo Agent that plans promotions and flags the margin-negative ones, the Pricing Agent that recommends price moves inside guardrails, and the Segmentation Agent that keeps customer segments current as behavior shifts. The Execution family handles what happens at the point of sale, with three types: the Shelf Agent that audits shelves from photographs and triggers corrections, the Field Sales Agent that briefs representatives before a visit and captures the outcome, and the Merchandising Agent that verifies planogram and point-of-sale-material compliance.
The Operations family moves goods, again with three types: the Route Agent that plans and re-optimizes routes, the Delivery Agent that orchestrates proof of delivery and arrival-time flows, and the Inventory Agent that watches stock and triggers replenishment. The Data family is the connective tissue, with the Demand Agent sensing signals and updating forecasts, the Market Intelligence Agent watching competitor pricing and promotion from field data, and the Data Quality Agent detecting and repairing anomalies across feeds. Finally the Finance family closes the loop with three types that most route-to-market software ignores entirely: the Deduction Agent that identifies and recovers invalid deductions, the Settlement Agent that reconciles claims and trade settlements, and the Credit Agent that scores outlet and distributor credit and flags exposure.
Sixteen types, five families, one route to market. The discipline matters more than the count. A taxonomy is closed and governed; a folksonomy is open and chaotic. Vendors cannot invent new types to make their product look novel. New types are added only by platform governance, deliberately and rarely. That constraint is exactly what makes the classification mean something. When a buyer sees a solution listed as an Order Agent, the word carries a definition that the platform stands behind, not a marketing claim the vendor minted last quarter.
That governance is enforced through the FMCG Verified certification. To earn the FMCG Verified mark for its type, a solution must satisfy six criteria: protocol compliance, so it consumes and emits on the shared FMCG Cloud data model and interoperates with other agents; observability, so every action is visible in the activity feed with an auditable decision trail; performance against the latency expectations of its family; a passed security review with least-privilege data access disclosed at install; a committed support response time; and a documented data-handling posture covering residency, retention, tenant isolation, and no cross-tenant training without opt-in, consistent with GDPR and CCPA expectations. Certification is how a taxonomy becomes a standard rather than a label.
The deeper reason this is a taxonomy and not a catalog is the data model underneath. All six FMCG Cloud product categories sit on one shared data layer. The shelf can talk to the order, the order can talk to the route, and the deduction can be traced back to the claim, because they describe the same world in the same terms. Agents only compose if they share that world. A feature list does not need a common data model; a system of cooperating agents cannot exist without one. The taxonomy is the human-readable map of that system, and FMCG Cloud Intelligence is the layer that lets agents reason and act across it.
We do not yet publish customer outcomes, and we are not going to dress up industry averages as our own. What we will say is structural and, we believe, durable: organizing an FMCG AI ecosystem by agent type, on one data model, under one certification, gives buyers a vocabulary for what they are adopting and vendors a clear contract for how to participate. A feature list goes stale the week after you ship it. A well-governed taxonomy is the thing the whole ecosystem can grow on.