AI in logistics is being sold on autonomy while most operators are still fixing their data. US companies spent $2.40 trillion on business logistics in 2025, transportation took 61.8% of it, and Gartner forecasts spend on supply chain software with agentic AI rising from under $2 billion in 2025 to $53 billion by 2030.
Yet McKinsey found under 10% of surveyed distributors had an AI roadmap with prioritized use cases. This paper maps which use cases pay back first, compares statistical, machine learning and foundation-model forecasting, and sets out where agents fit.
Its evidence comes from analyst forecasts, a US cost report, a public forecasting competition on Walmart data and model research papers. It is not a survey of logistics operators and does not measure any single company's results.
None of this requires a supply chain leader to pick a model. It requires asking three plain questions of any AI proposal, each answered by one document the team either has or does not. The answers show whether the work rests on a baseline, a policy and a shared identity for every shipment.
Ask which forecast the new model is measured against, and accept only an answer that names a naive baseline such as last period's actual. Forecast value added shows whether each step, including planner overrides and consensus meetings, improves accuracy or makes it worse. Without it, a better model and a worse one look the same in a slide. The artifact is a monthly forecast value added report broken out by segment.
Ask what an agent is allowed to spend, which systems it can call and who can stop it. Gartner advises keeping full automation to low-risk decisions for now. An agent with no spend cap or audit trail is a liability that has not yet cost anything. The artifact is a written agent policy listing the spend cap per action, the tool allow-list, the logging scope and the kill switch.
Ask how one shipment is identified as it moves from ERP to TMS to WMS to the carrier. Gartner names data completeness and accessibility as the main barriers to touchless forecasting, and broken handoffs are where that shows. If nobody can trace a load end to end, no model will. The artifact is a shared shipment ID definition plus a 90-day EDI 214 completeness report by carrier.
See which use cases rely on data that is already digital, such as document automation, ETA prediction and freight rate prediction, and which need cleaner master data and stronger governance first.
Review what the M5 competition showed about LightGBM and global models, where Chronos and TimesFM fit for cold starts and sparse series, and how hierarchical reconciliation keeps forecasts coherent.
Use six agent patterns, from track-and-trace to freight audit, with spend caps, tool allow-lists, full logging and a kill switch mapped to the NIST AI Risk Management Framework.
Rate data integration, EDI, forecasting process, master data, decision integration, governance and talent on a 1 to 4 scale, then follow a quarter-by-quarter roadmap matched to your score.
AI in logistics applies machine learning, optimization, language models and agents to freight, warehouse and fleet decisions. Common uses include demand forecasting, freight rate prediction, route optimization, ETA prediction, warehouse slotting, document extraction and exception management.
The value appears when predictions are written back into the TMS, WMS or planning system that people already use, so decisions change rather than reports.
Document automation, ETA prediction and freight rate or tender acceptance prediction usually pay back fastest. They use data that is already digital and do not require a new planning process. Demand forecasting follows close behind.
Network optimization and autonomous control towers take longer because they need clean master data, stable forecasts and stronger governance before teams will trust the outputs.
Often, but not always. In the M5 competition on 42,840 Walmart sales series, LightGBM-based global models beat the statistical benchmarks. The gains came from learning across many related series and using drivers such as price and events.
For stable, low-volume series, exponential smoothing can match machine learning, so test both in a backtest and keep whichever adds value over a naive forecast.
They are ready as benchmarks and for cold starts, sparse series and fast prototyping. Chronos and TimesFM produce zero-shot forecasts without per-dataset training, and Chronos-2 adds covariate support. They are not automatic replacements for tuned models on rich data.
Include them as candidates in a model tournament, measure forecast value added and review governance before any production use.
McKinsey estimates AI in distribution operations can cut inventory 20 to 30% and logistics costs 5 to 20%, and that AI tools can add 7 to 15% capacity to warehouse networks. Its earlier research found early adopters reported 15% better logistics costs and 35% better inventory levels than slower competitors.
Treat these as direction rather than targets, and measure against your own pre-project baseline.
AI agents are software that monitors events, reasons over options and takes actions through tools such as TMS APIs, email and EDI. Examples include rebooking rejected tenders, triaging late shipments and auditing freight invoices.
Gartner forecasts 60% of enterprises using supply chain management software will adopt agentic AI features by 2030, up from 5% in 2025, but advises keeping full automation to low-risk decisions for now.
It covers a use-case map, a forecasting comparison, agent patterns and guardrails, a data checklist, KPIs, a 12-month roadmap and a readiness scorecard. The figures come from Gartner, McKinsey, BCG, the 2026 State of Logistics report, BTS, the M5 competition and model research papers.
Analyst figures are forecasts and estimates, not measured outcomes, and the roadmap and scorecard are practical guidance rather than survey data.
Drop your details and we'll send AI in Logistics and Supply Chain Forecasting straight to your inbox - no spam, unsubscribe anytime.
Bring your current forecast process, your TMS and WMS feeds and the use case you are weighing. We will score it against the readiness scorecard with you and scope a first use case your data can support. SECTION 7 - FAQ - 5 to 8 questions
Book a readiness session