AI in Logistics: From Demand Forecasting to Right-Turn-Only Routes

In this episode of Rozmowy Logistyków, Piotr Skobało and Adam Sobolewski discuss where artificial intelligence is already changing warehouses, customer service, and supply chain planning. They walk through concrete AI applications available not just to the largest companies, with examples from Polish and European vendors.

WHAT "AI" REALLY MEANS IN LOGISTICS

Artificial intelligence is broader than machine learning, which is broader than deep learning — these technologies have been at work in logistics for decades, just not always under that label. It's worth distinguishing today's applications from the Hollywood vision of general artificial intelligence (AGI) — what actually works are specific, narrow applications to specific problems, not a universal "Skynet" running everything at once.

DEMAND FORECASTING: DON'T CORRECT THE MODEL BY A LITTLE

Amazon has been limiting manual corrections to algorithmically generated forecasts since 2014 (the "Hands off the Wheel" initiative), because research (including Michael Gilliland's book Business Forecasting) on samples of thousands of forecasts showed that people correcting forecasts usually make them worse. A practical rule: if you correct a forecast, do it by a large percentage (not 5-10%, but clearly more) — a small correction usually means noise, not real knowledge the model failed to capture. Neural-network-based demand forecasting solutions are available today not just to giants like Amazon, but also to small and mid-sized companies (the Polish solution Dgca, formerly Slimstock, deployed them as part of an NCBR grant).

RISK MANAGEMENT: IMAGE ANALYSIS WHERE THE DATA IS "DIRTY"

While demand forecasting works on structured, tabular data, transport risk management requires analyzing camera footage — a different type of data that classic tools can't handle. The Polish-German company SNC uses image recognition to monitor the space between the driver's cab and the trailer — the system detects cuts in the tarpaulin or the unusual presence of a person and sends a real-time alert, combining visual analysis with the internet of things.

PREDICTING PEOPLE'S MOVEMENT IN THE WAREHOUSE, NOT JUST GOODS

Beyond classic slotting based on turnover (mostly simple regression models, used for years by WMS vendors), systems are emerging that predict in real time where a specific employee will be — based on their schedule, past productivity, and order types — in order to avoid collisions between workers in the aisles. This is a newer application: modeling the movement of people, not just the optimal placement of products.

ROUTE OPTIMIZATION: COUNTERINTUITIVE RESULTS ARE A STRENGTH, NOT A FLAW

UPS's ORION (On Road Integrated Optimization) system analyzes hundreds of possible delivery routes, delivering around a 2-3% improvement in fuel consumption — at UPS fleet scale, that's an enormous amount in absolute terms. A classic, older example: UPS programmed routes to avoid left turns (waiting at lights creates collision risk and wastes fuel), which delivered over a 10% efficiency improvement — even though it meant more physical right turns. Practical takeaway: when deploying AI, don't try to fit its results to your intuition — the strength of neural networks lies precisely in detecting relationships a human would miss, because they result from the interaction of too many variables at once.

CUSTOMER SERVICE: FROM SIMPLE BOTS TO FULL AUTOMATION

The first wave of customer-service bots relieved the front line of simple questions ("where's my package"), passing harder cases to a human. Today the boundary keeps shifting: bots powered by a company's own knowledge base can hold a substantive conversation about a product, and the next step is full integration with the order system — automatically rerouting a shipment in transit, canceling or modifying an order without human involvement, if the status allows it.

HOW TO START IF YOU RUN A SMALL OR MID-SIZED COMPANY

The first question isn't "which AI should we deploy," but "where does it hurt most in my organization" — if the warehouse is outsourced, look for solutions in customer service, not in automating the physical warehouse process. A warning from practice: one company advertised itself as an AI provider for image classification, while in reality it employed a team of people in India manually reviewing and describing images (the "protein algorithm"). It's worth asking a vendor about the specific tools actually used and where the data is really stored — most AI solutions run on a cloud (SaaS) model, which means sending data outside your own infrastructure.