Most supply chain AI projects chase headline-grabbing use cases while two of the largest cost leaks — perishable waste and freight rate timing — sit in plain sight, solvable with data that already exists.
We keep meeting operations leaders who have an AI strategy document and a proof-of-concept backlog — but their biggest cost leaks are not on either list.
Two problems came up in back-to-back conversations this month, in completely different industries. Both had the same shape: a well-understood cost, volatile inputs that humans struggle to track, and data that was already being collected but not used predictively.
Perishable waste in grocery retail
A fresh grocery operation writing off 8–12% of perishable stock every week is not unusual. It is also not inevitable.
The root cause is forecast lag. Traditional demand planning uses historical sales and seasonal curves, but a heatwave, a local festival, or a trending recipe can shift buying patterns within hours. By the time last week's data informs next week's order, the moment has passed.
AI demand sensing changes the timing. By ingesting real-time signals — weather forecasts, local event calendars, even social trend data — alongside POS streams, it adjusts inventory predictions fast enough to matter for products with a 48-hour shelf life.
The engineering is not exotic. A streaming ingestion layer, feature engineering at the store-SKU-day level, gradient-boosted forecasting models, and a feedback loop that retrains weekly. The hard part is data quality: POS gaps from scanner errors, inconsistent markdown tracking, and stock-count mismatches all need cleaning before models can learn.
Transformations like this typically target a 30–40% reduction in perishable waste. For a chain doing £500M in fresh sales, that is serious recovered margin.
Freight rate volatility in logistics
A freight forwarder managing 50,000+ annual shipments faces a different version of the same problem: timing.
Freight rates move with fuel prices, port congestion, seasonal demand, and geopolitical disruptions. Procurement teams negotiate contracts based on recent averages, then absorb spot-market premiums — typically 20–40% above contracted rates — when reality diverges from the average.
The data to predict these movements now exists in accessible, API-ready form: digital freight exchange indices, AIS vessel tracking for port congestion, fuel futures, and macroeconomic indicators. Training trade-lane-specific models at multiple time horizons (1–2 weeks for spot decisions, 1–3 months for contract negotiations) turns this data into actionable procurement signals.
The practical payoff is a 10–18% reduction in spot-market exposure and 5–8% improvement in overall freight procurement costs. For an operator spending £80M on carriers, that is £4–6M.
The pattern worth noticing
Both problems share three traits:
- The cost is well-known — everyone in the business can point to the waste line or the spot-market premium.
- The volatility is driven by external signals — weather, fuel, events, congestion — that are now digitally accessible.
- The prediction window is short enough for ML to outperform human judgment — hours to weeks, not quarters.
These are not moonshot AI projects. They are engineering problems with clear data inputs, measurable outcomes, and payback periods measured in months.
Where to start
If either of these sounds familiar, the first step is unglamorous: audit your data. Can you get clean, daily POS data by store and SKU? Do you have historical freight bookings tagged by trade lane with actual rates paid? The model sophistication matters far less than the data foundation.
Our teams at Skillikz have been building these kinds of data-to-decision pipelines across retail and logistics. If you want to explore what is feasible with the data you already have, we would welcome the conversation.