Ask any distribution leader what makes a reorder suggestion useful, and you tend to get the same answer twice. The first time, they say accuracy. The second time, after a pause, they say trust. Those are not the same thing, and the gap between them is where most ordering tools quietly fail. A number that is right but unexplained still gets overridden by the rep, ignored by the outlet, or padded "just in case" by the distributor. A trustworthy suggested order is one the person on the receiving end is willing to act on without second-guessing it, and that only happens when the suggestion is built from signals they recognise as real.
The most important of those signals is sell-out history, not sell-in. It is tempting to base a reorder on what an outlet last bought, but what an outlet buys and what it actually sells through to shoppers can diverge sharply, especially when a previous order overshot. A credible suggestion leans on a meaningful window of demand at the point of sale, enough weeks to separate a genuine trend from a single noisy fortnight, while still being recent enough to catch a store that is accelerating or fading. The aim is to model the rate at which product leaves the shelf, then size the next order to cover the cycle until the next visit or delivery, with a sensible buffer rather than a guess.
Seasonality is the second signal, and it is the one humans are worst at holding in their heads. Demand for most consumer-goods categories breathes on a calendar: weather, paydays, holidays, school terms, regional events. A flat run-rate average will under-order going into a peak and over-order coming out of one, and the cost shows up as either lost sales or expiry. A suggestion that understands the shape of the year proposes a different basket in the week before a known surge than it does in the lull after it. Critically, this has to be expressed in a way the user can sanity-check against their own intuition about the season, not buried inside a black box.
The third signal is the assortment gap. Sell-out history can only tell you about products the outlet already stocks; it is silent on the lines it should stock but does not. This is where a single shared data model earns its place. When B2B Ordering sits on the same ConnectX data layer as Field Sales, Retail Execution and Shelf Intelligence, a suggested basket can reason about what comparable outlets in the same channel and cluster sell successfully, and surface the obvious void: the format competitors carry, the flavour the neighbouring store moves well, the entry-price pack missing from a premium-skewed range. An assortment recommendation is only as honest as the comparison behind it, so the basis for "stores like yours" should be visible, not asserted.
Promotions are the fourth signal, and the easiest to get wrong. A live deal changes the right answer in two directions at once: it should lift the suggested quantity on the promoted line, and it should account for the pull-forward and cannibalisation that ripple through the rest of the basket. A suggestion that ignores active trade terms feels broken to anyone holding a promo sheet; one that blindly maximises on every deal teaches outlets to distrust it. The discipline is to fold the promotion in as a transparent adjustment, clearly attributed, so the user sees both the base recommendation and what the deal added.
There are quieter inputs that matter too: the rep's brief from the last visit, the most recent shelf scan flagging an out-of-stock, credit and payment status that should gate a basket without silently deleting lines, minimum order quantities, lead times and shelf-life constraints. The point of naming all of these is not to suggest more inputs make a better number. It is the opposite. The more signals feed a recommendation, the more essential it becomes that the system can say, in plain language, why it landed where it did.
This is why explainability is not a feature bolted on for compliance; it is the mechanism by which a suggestion becomes an order. In FMCG Cloud, the Order Agent is built to show its reasoning: this quantity because sell-out is running at this rate, this extra case because the season is turning, this new line because similar outlets stock it, this uplift because the deal is live. A rep can defend that to a sceptical owner. An outlet placing an order at midnight, long after rep hours, can accept it without a phone call. A distributor can audit it. The explanation is also what makes the system improvable, because an override stops being noise and becomes feedback against a stated assumption.
Every solution in the FMCG Cloud marketplace classifies under the FMCG Cloud Agent Taxonomy and has to earn FMCG Verified certification, and explainability is a large part of what that verification is for. The bar we hold ourselves and our partners to is not "the model is confident." It is "a category manager, a delivery planner and a shop owner would each look at this basket, understand how it was reached, and agree it is reasonable." A good order recommendation is not the one with the cleverest math. It is the one a human is willing to own.