Logistics and supply chain run on data — orders, inventory, shipments, lead times, costs — and the constraint has rarely been a shortage of data but the time to make sense of it. AI helps in two distinct ways that are easy to confuse: predictive/optimisation models (forecasting, routing) that are mathematical and trained on your data, and language models that summarise, explain, and analyse documents. Knowing which tool fits which job is the whole game. This guide walks the main applications and where each kind of AI actually belongs.
Demand forecasting
Forecasting is a statistical machine-learning problem. Time-series models — trained on your historical sales, seasonality, promotions, and external signals — predict future demand far better than any chatbot. The LLM’s role is around the forecast: explaining which drivers moved the numbers, drafting a planning brief, or generating scenario narratives (“what if lead times rise 20%?”) that a planner can react to.
A useful prompt once you have the numbers: “Given this forecast and these assumptions, summarise the key risks and the inventory implications in five bullet points for an operations review.” The model turns a spreadsheet into a decision-ready brief — but it does not produce the forecast itself.
Anomaly detection in shipment data
Across thousands of shipment records, the problems hide in the outliers — a lane whose cost suddenly spiked, a transit time three times the norm, a supplier whose defect rate is creeping up. Statistical anomaly detection flags these automatically, and an LLM can then draft a plain-language explanation and a suggested next step for each flag.
The discipline here is that an anomaly is a question, not an answer. A flagged outlier might be a data error, a one-off, or a genuine problem. AI surfaces it and proposes a hypothesis; a human confirms the cause before anything changes.
Supplier contract analysis
This is where language models shine directly. Feed a supplier agreement to an LLM (using a tool cleared for confidential documents) and ask it to extract key terms — pricing, SLAs, penalty clauses, renewal dates, liability caps — into a structured summary, and to flag anything unusual versus your standard template.
This turns a stack of contracts you would never have time to read closely into a searchable, comparable set. Treat extracted terms as a draft for legal review, not as legal advice, and verify any clause before relying on it.
Last-mile and route optimisation
Route optimisation is an operations-research problem — variants of the vehicle routing problem — solved by specialised solvers, not by asking a chatbot to plan deliveries. The genuinely AI-flavoured part is feeding live signals (traffic, delivery windows, vehicle capacity) into those solvers and using a language model to explain the resulting plan to drivers and dispatchers.
The throughline across every workflow is the same: use predictive and optimisation models for the math, use language models for the explanation and the documents, and keep a planner in the loop for every decision. Teams that draw that line clearly get the productivity gain without betting the supply chain on a hallucination.