The Future of Warehouse Intelligence

For most of the past two decades, warehouse management systems have been transactional systems. They record movements, enforce rules, and generate reports. The intelligence — deciding where to put something, how to sequence a wave, or whether a supplier shipment is going to cause a problem — lived in the heads of experienced supervisors.

That model is breaking down. Labour costs have risen significantly across all major markets. The UK, US, and EU have all seen warehouse labour cost increases of 25–40% since 2020, driven by a combination of minimum wage legislation, post-pandemic competition for workers, and a structural shift in the composition of the workforce. At the same time, customer expectations — shaped by the Amazon effect — have compressed acceptable delivery windows from days to hours.

The operations that are coping best are not those with the largest headcount. They are those that have started using the data their WMS already collects in more deliberate ways. AI does not replace the WMS — it sits on top of it, turning historical transaction data into forward-looking decisions.

Slotting — the process of deciding which product lives where in a warehouse — is one of the oldest problems in distribution. Most facilities slot products once during the initial racking setup, then adjust only when something goes badly wrong. The result is that as product velocities change, the slot configuration drifts further from optimal. Fast movers end up at the back of a rack. Slow movers take up prime real-estate near the dock.

AI-driven dynamic slotting continuously re-evaluates product velocity, pick density, and ergonomic constraints. In practice, this means running overnight batch jobs that generate slotting recommendations based on the last 30 to 90 days of actual pick data, weighted by seasonal trends and upcoming promotional activity. The recommendations are not arbitrary — they are constrained by product compatibility rules, weight limits, and the practical reality that moving product costs labour.

Across the Warewiser customer base, facilities that have enabled dynamic slotting see a mean reduction in picker travel distance of 31%. In one food distribution centre operating three shifts across 45,000 sq ft, the system moved 214 SKUs over a single weekend and reduced per-order pick time by 22 seconds — saving roughly 40 hours of labour per week at full throughput.

Traditional replenishment models work on fixed reorder points and economic order quantities. They are straightforward to configure and easy to understand, but they do not adapt to the real behaviour of demand. A product that sells 50 units a week on average might sell 10 units in quiet weeks and 120 units during a promotion. A fixed reorder point calibrated to the average leaves you either overstocked most of the time or short when it matters.

Machine learning forecasting models — specifically gradient boosting and LSTM neural networks, depending on the SKU count and data history available — learn the shape of demand for each product individually. They incorporate day-of-week patterns, known promotions, supplier lead time variability, and macro signals like weather or school term dates where relevant.

The practical outcome is a reduction in both safety stock and stockouts. This sounds contradictory, but it is achievable when forecasts are accurate enough that you do not need to carry extra stock as insurance. In a 2024 analysis of Warewiser deployments in the grocery sector, facilities using AI forecasting held 18% less safety stock while improving in-stock rates by 4.3 percentage points compared to their pre-deployment baseline.

Most warehouse operators underestimate how much of their supervisors' time goes on managing exceptions. A supplier delivers short. A batch fails a quality check. A pick location runs empty mid-wave. A customer order comes in after the wave cut-off with a VIP flag. Each of these is a small problem individually, but in a busy facility handling 2,000 to 5,000 orders a day, they happen dozens of times every shift.

Rule-based exception handling — the kind most WMS systems offer — can automate the simple cases. If a pick location is empty, find the nearest reserve. If an order misses cut-off, flag it for manual review. But more complex exceptions require judgment: how much disruption is it worth accepting to accommodate a late high-value order? Should you short-ship or hold the order? Which backorder should get priority when stock replenishment arrives?

AI-based exception handling uses a scoring model trained on historical decisions to suggest the appropriate resolution. Critically, it presents the reasoning — so supervisors can override it when the context warrants — rather than just executing silently. In early deployments, 73% of exceptions are resolved without human intervention, reducing the cognitive load on shift supervisors and allowing them to focus on genuinely unusual situations.

The honest picture is that AI in the warehouse delivers the most value when the underlying data is clean. Poorly maintained master data — inaccurate bin locations, wrong product weights, inconsistent unit-of-measure configurations — limits what any model can do. Before any AI rollout, it is worth conducting a data quality audit. In our experience, this typically takes two to four weeks and surfaces issues that would have caused problems regardless of AI.

The other honest caveat is that the results described in this report reflect facilities that have been running Warewiser for at least six months, with clean data and trained operators. The first 60 to 90 days after go-live are a calibration period. Models need sufficient transaction history to produce useful recommendations, and operators need time to build trust in the system's suggestions.

Facilities that approach the implementation with realistic expectations — measurable improvement over six to twelve months, not overnight transformation — consistently report higher satisfaction and better outcomes than those expecting immediate results.