Every time a field rep opens a customer's order screen, the Suggested Order engine produces a ranked list of products with recommended quantities. Behind that simple list is a machine learning model that blends seven distinct signals into a single optimized recommendation. Understanding how each signal contributes helps you trust the suggestions and configure the system to match your business strategy.
Signal one is purchase history. This is the foundation. The model analyzes what each customer has ordered in the past, how frequently, in what quantities, and at what intervals. It identifies regular replenishment patterns, one-time purchases, and gradually increasing or decreasing trends. A store that orders 10 cases of a beverage every two weeks will see that product suggested at the right time with the right quantity. The model also detects when a customer stops ordering a previously regular item, flagging potential churn at the SKU level.
Signal two is seasonality. FMCG demand is heavily seasonal. Soft drink orders spike in summer, chocolate sales peak before holidays, cleaning product demand increases during Ramadan. The model captures these patterns at multiple levels: product-level seasonality, category-level trends, and market-wide cycles. It automatically adjusts recommended quantities based on where you are in the seasonal curve, without manual configuration. A store in a tourist area might see a different seasonal profile than one in a residential neighborhood, and the model captures this granularity.
Signal three is cross-sell affinity. Products are not purchased independently. A store that buys cola likely buys chips. One that orders laundry detergent often buys fabric softener. The model learns these affinities from transaction data and uses them to suggest complementary products that the customer might not have considered. This is especially powerful for new product introductions, where the model can identify which existing products are most associated with the new SKU and target those customers.
Signal four is promotion eligibility. Active promotions are a first-class signal, not an afterthought. The model checks which products have active promotions, whether the customer is eligible based on tier, region, or purchase history, and how the promotional pricing affects the optimal order composition. Promoted products receive a boost in the ranking, but the model balances promotional coverage against the customer's actual demand patterns to avoid over-ordering promoted items that will sit unsold.
Signal five is credit limit. Every recommendation must respect financial reality. The model checks the customer's available credit balance and caps the total order value accordingly. When the credit limit is binding, the model prioritizes high-margin products and essential replenishment items over nice-to-have additions. This prevents the common problem of reps building orders that exceed the customer's credit capacity, only to have the order rejected at fulfillment.
Signal six is stock health. If the customer still has significant inventory of a product from the last order, suggesting more of the same is counterproductive. The model estimates current stock levels based on historical consumption velocity and time since last order. Products with healthy estimated stock are deprioritized, while products approaching stockout are boosted. This signal directly reduces over-ordering and the waste associated with expired or damaged excess inventory.
Signal seven is return rate. This is the signal that most order suggestion systems ignore, and it is one of the most impactful. If a product has a high return rate for a specific customer or store type, the model penalizes it in the recommendation. Returns are expensive for everyone in the supply chain. By factoring in historical return patterns, the model avoids suggesting products that are likely to come back, protecting margins for both the distributor and the retailer.
The blending mechanism combines these seven signals using a weighted ensemble. Each signal produces a score for every candidate product, and the ensemble combines them into a final recommendation score. The weights are configurable at the tenant level, allowing you to emphasize different signals based on your business priorities. A company focused on reducing returns might increase the return rate signal weight, while one launching new products might boost the cross-sell affinity signal.
Cold start handling is a critical capability. When a new customer has no purchase history, signals one, six, and seven have no data. The model falls back to a hierarchical approach: first it uses data from similar customers in the same channel and region, then category-level patterns across the tenant, and finally population-level bestsellers. As the customer builds order history, the model gradually transitions from these proxies to personalized signals. Most customers see fully personalized suggestions within three to five order cycles.
A concrete example illustrates how the signals interact. Consider a convenience store that regularly orders beverages and snacks. It is mid-July (seasonality boosts cold beverages), there is an active promotion on a new energy drink (promotion signal), the store's last cola order was eight days ago and they typically reorder every seven days (stock health suggests they are running low), and they have a 15% return rate on a particular juice brand (return signal penalizes it). The model will suggest a replenishment quantity of cola, boost the new energy drink recommendation, and reduce or exclude the problematic juice brand, all while keeping the total order within the store's credit limit.
Explainability is built into every recommendation. The API response includes per-signal contribution scores for each suggested product, so the rep can see exactly why a product is recommended and why it is at a specific quantity. This transparency builds trust with the field team and enables informed conversations with store owners about why certain products are being suggested.