For thirty years, the unit of business software has been the app: a database wrapped in screens, forms, and buttons. You bought a field-sales app, an order-management app, a route-planning app. Each one gave your people a place to enter data and a dashboard to read it back. The software held information and presented choices. The actual work of judging, deciding, and acting stayed with humans clicking through the interface. That model is now ending, and FMCG is one of the industries where the change will be felt fastest.
The reason is simple. Consumer goods is a business of relentless, repetitive operational decisions made at enormous scale. A single distributor running hundreds of routes touches tens of thousands of order lines, shelf checks, and delivery sequences every week. The classic app made each of those decisions slightly easier to record. It did not make them for you. The promise of an agents-first approach is different: software that does a job rather than software that helps a person do a job. Audit the shelf. Optimise the promotion. Sequence the route. Flag the order that should have been bigger. The deliverable is an outcome, not a screen.
This is what we mean when we say the unit of software is moving from apps to agents. An app is a tool you operate. An agent is a worker you supervise. The difference is not cosmetic. When the unit of work changes, so does everything around it: how you buy software, how you measure it, how you trust it, and how the pieces fit together.
Consider what an agents-first product actually looks like in a route-to-market context. A Shelf Agent does not just store a photo of a gondola; it inspects the image, identifies the gap or the planogram violation, and raises the issue. An Order Agent does not just present a blank order form; it proposes the order line by line based on what the outlet sells, then surfaces the exceptions worth a human's attention. A Route Agent does not just draw a map; it sequences the stops against time windows, vehicle constraints, and service commitments. In each case the human moves up a level, from doing the task to reviewing the agent's work and handling the edge cases that genuinely need judgment. That is the practical meaning of agents-first: the routine decision is made by the agent, and the person's time is reserved for the decisions that are actually hard.
None of this works if every agent lives in its own silo, which is precisely where the old app stack broke down. When the shelf system, the order system, and the route system are separate applications with separate databases, the agents inside them are blind to each other. An agent is only as good as the context it can see. This is why an agents-first platform has to be built on one data model underneath. At FMCG Cloud we call that shared layer ConnectX, and it is the reason the categories we serve are integrated rather than bolted together. Field Sales, Retail Execution, B2B Ordering, Route and Delivery, Revenue Growth AI, and Shelf Intelligence are not six products that happen to exchange files at midnight. They are six views onto the same data, which means the shelf can talk to the order and the order can talk to the route. A gap detected during retail execution can become a reordering signal without a human re-keying anything. That cross-category awareness is what turns a collection of agents into a working system.
There is a discipline question hiding inside all of this. If agents are going to act on your behalf, you need to know which agents are trustworthy, what they are certified to do, and how they behave under observation. A marketplace of agents without standards is just a longer menu of risk. That is why we organise every solution, first-party and partner alike, under a single classification: the FMCG Cloud Agent Taxonomy, sixteen agent types grouped into five families spanning commercial, operations, execution, data, and finance work. Anything that lists in our marketplace must classify under that taxonomy and earn the FMCG Verified certification, which covers interoperability, an observability event schema, latency and support commitments, a security review, and a clear data-handling posture. The taxonomy is not marketing scaffolding; it is how a buyer can compare two agents that claim to do the same job and know what they are actually getting.
It is worth being plain about what this is and is not. The intelligence layer that makes our agents act rather than merely report is FMCG Cloud Intelligence, and it is early. We are building this in the open, against a clear thesis, rather than pointing to a wall of outcomes we have not yet earned. The industry context is well understood: route-to-market in consumer goods is fragmented, decision-dense, and starved of the labour to do every task by hand. Agents are the natural response to that shape of problem.
For leaders running distribution and brand route-to-market, the shift to agents-first is less a product decision than an operating one. The question is no longer which app your teams will log into. It is which jobs you are willing to hand to a supervised agent, what you will measure it on, and how you will keep a human in the loop where judgment still matters. Teams that frame the question that way will spend the next few years compounding. Teams still buying screens will spend it clicking. The unit of software is changing, and in FMCG the work is well suited to it.