Logistics has had AI longer than most industries — operations research, routing algorithms, forecasting models. These are decades-mature. Where LLMs add new value is in the unstructured-text layer around the optimization core: exception handling, customer communications, supplier intelligence, operations copilots. The mistake is trying to let LLMs replace the OR core. The win is letting them handle the edges OR has always struggled with.
Where LLMs genuinely add value
Exception handling
A driver reports a delivery exception: customer not home, address invalid, damage to shipment. Historically a free-text field reviewed by a dispatcher. LLM reads the narrative, classifies the exception, suggests action (reschedule, redirect, claim). Dispatcher approves or overrides. Routine exceptions become 2-minute handoffs instead of 15-minute investigations.
Customer communications
ETA changes, delivery delays, proactive notifications. The numbers come from the OR system; the narrative explanation comes from the LLM, customized to the customer's preferences and channel. 'Your package is delayed because of storms in the Dallas hub and should arrive Thursday instead of Wednesday' beats a generic 'delay' alert every time.
Supplier and partner intelligence
Carrier performance reports, customs news, port congestion updates, geopolitical disruption signals. LLMs read the news, classify relevance to your specific lanes, summarize for the ops team. The value compounds: a planner who reads 20 news items instead of 200 can actually act on what they read.
Operator copilots for planning
Ops planners work across many systems. An AI copilot that answers 'how many trucks are available in Dallas Tuesday' or 'which carriers are under-utilized this week' saves minutes on every decision. When ops decisions happen hundreds of times a day, the savings compound to real value.
Where LLMs do not belong
Routing itself. Vehicle routing (VRP) is a mature OR problem with specialized solvers that guarantee optimality or bounded approximation. LLMs cannot match this and never will — the problem structure rewards algorithms, not language models. Use or-tools, Gurobi, OptaPlanner, or specialized vendor solvers. LLMs can configure the constraint parameters conversationally; they cannot solve the optimization.
Time-series demand forecasting. Statistical and ML methods (ARIMA, Prophet, gradient boosting, causal models) work well here with interpretable uncertainty. LLMs add overlay (incorporating news signals, weather events) but should not replace the statistical base.
Load-planning and container-loading optimization. Another classical OR problem. Again, LLMs can help with input interpretation and output explanation; the core solve is not their job.
The unglamorous reporting layer
A disproportionate share of logistics AI ROI comes from reporting automation. Weekly performance reports, anomaly explanations, trend narratives. Ops teams spend hours hand-writing these. An LLM reading dashboard data and writing the narrative saves hours per week per analyst. Not a flashy feature — and one of the highest-returning ones.
Variance analysis is similar. 'Why is on-time percentage down 3% this week?' Historically takes an analyst an afternoon. AI that cross-references schedule changes, weather, carrier performance, and exception data can generate a credible explanation in seconds, with the analyst spot-checking and adding judgment.
Change management
Logistics is an industry with tight margins and conservative buyers. Anyone proposing 'AI-powered everything' gets filed as a bluff. Anyone proposing specific productivity wins in narrow workflows gets trials. Sequence accordingly: start with one workflow, earn trust, expand.