AI in Supply Chain and Logistics: Forecasting, Inventory, Routing, and Sustainability

Supply chain is where enterprise AI has the longest track record and the least glamour. The four use cases that carry real money — demand forecasting, inventory optimization, transportation routing, and emissions accounting — are all optimization problems over messy operational data, not chatbot problems. That makes the buying decision different from most AI categories: the model is rarely the constraint, the data plumbing always is, and the vendor landscape has just been reshaped by hyperscalers retiring the turnkey services they sold five years ago. The decision in front of a CIO or supply chain VP is not "should we use AI" — it is which of three buy surfaces (planning suite, general ML platform, narrow solver API) each lane deserves, and in what order.

By the numbers

Four lanes, one data problem

The four lanes look like separate procurements, and vendors sell them that way. Structurally they are one system: the forecast sets safety stock and reorder points, inventory positions determine what moves, routing determines how it moves and what it emits, and the emissions ledger increasingly feeds back into carrier and supplier selection. Every one of them consumes the same upstream inputs — clean item masters, honest lead-time history, granular transactional data — which is why the strongest predictor of success in any lane is whether the previous lane's data foundation exists. The table below is the map this guide follows.

Demand forecasting: the anchor use case

Forecasting is the lane to fix first, because every downstream lane inherits its error. It is also the lane where the research evidence is most honest. The M5 competition — the largest public benchmark of retail demand forecasting, built on Walmart's hierarchical unit-sales data — was designed to evaluate the accuracy and uncertainty of forecasting methods on exactly this problem, and follow-up work comparing the M5 data against other grocery retailers found "only small discrepancies between the examined data sets, supporting the representativeness of the M5 data."^[3] The practical reading for a buyer: the machine-learning methods that win public benchmarks on Walmart-shaped data will probably transfer to your Walmart-shaped data — and the gap between a tuned gradient-boosting ensemble and a naive seasonal baseline is real but bounded. Vendors quoting dramatic accuracy lifts are usually comparing against a weak baseline, not against a competent statistical forecast.

The vendor landscape has consolidated hard, and the clearest signal is negative. Amazon Forecast — AWS's standalone ML forecasting service, launched as the packaged version of Amazon's own demand-planning technology — is now closed: AWS's documentation states plainly that "Amazon Forecast is no longer available to new customers," with existing customers able to continue and a published migration path to SageMaker Canvas.^[4] AWS's supply chain application layer has meanwhile been folded into a new product: the AWS Supply Chain page now presents Amazon Connect Decisions as "the next generation of AWS Supply Chain," with demand forecasting built on Amazon's internal forecasting technology and its supply planning capability still labeled preview.^[5] The prior generation's documentation — where Demand Planning "generates forecasts using proprietary machine learning algorithms based on historical forecasting experience" — already lives under a legacy path.^[6] Two product generations in four years is a fact a procurement team should price in.

Microsoft's approach is instructive by contrast, because it is legible. Demand planning in Dynamics 365 Supply Chain Management ships four named forecasting algorithms — auto-ARIMA, ETS, Prophet, and XGBoost — plus a best-fit model that automatically selects the best algorithm for each product and dimension combination by mean absolute percentage error, and an escape hatch for custom Azure Machine Learning models.^[7] That menu is roughly the technical frontier for tabular demand forecasting, and it tells you what any competent build would use too. The build-vs-buy question is therefore not about secret algorithms; it is about whether you want to own feature pipelines, retraining cadence, and planner-facing workflow. If you build, the forecast is a production ML system like any other — it needs the pipeline discipline covered in /guides/production-ml-pipelines-guide and the drift monitoring covered in /guides/model-monitoring-production-guide, because demand models degrade every time assortments, promotions, or channel mix shift.

Inventory optimization: where forecasts become dollars

A forecast has no P&L impact until it changes an inventory parameter. The classical machinery — safety stock sized from demand variability and lead-time variability, reorder points set at expected lead-time demand plus that buffer — assumed both distributions were stable and normal. They are neither, and this is where ML earns its keep: probabilistic models forecast the full distribution of demand and lead time jointly, per SKU and location, so safety stock becomes a service-level decision ("hold enough to hit 98% fill on A items, 90% on C items") instead of a one-size multiplier. The same machinery re-estimates continuously, so a supplier whose lead times are quietly stretching gets a rising buffer before a stockout, not after.

Two implementation truths decide whether this works. First, lead-time data is usually worse than demand data — purchase-order timestamps polluted by receiving delays, backdated confirmations, and suppliers quoting fiction — and a probabilistic model trained on fictional lead times will confidently misplace millions of dollars of buffer. Audit that data before modeling it. Second, planner trust is the binding constraint. A system that moves reorder points without showing its reasoning gets overridden into irrelevance; the overrides then contaminate the history the model retrains on. The pattern that survives contact with a planning organization is explainable recommendations with a tracked override rate — pilot it on a low-risk segment (C items, one distribution center), measure realized service level and inventory turns against a holdout, and expand only on evidence. Treat a falling override rate as a success metric in its own right.

Routing and carrier selection: solvers, not oracles

Transportation is the most mathematically mature lane. Vehicle routing is a constrained-optimization problem — time windows, capacities, driver hours, pickup-delivery pairing — and the state of the art is a solver fed by ML-estimated inputs (travel times, service durations, demand), not an end-to-end learned model. This is also the lane where a hyperscaler offering survived by being narrow: Google Cloud launched Cloud Fleet Routing in April 2022, positioned to solve small routing problems in seconds, batch-process large ones, and let operators reoptimize an existing plan up to 20 times a day without additional cost; Google cited Brazilian retailer Magalu applying it across more than 2.1 million shipments per month.^[8] The capability lives on as the Route Optimization API, which "assigns tasks and routes to a vehicle fleet, optimizing against the objectives and constraints that you supply for your transportation goals."^[9] The durable architecture lesson generalizes beyond Google: buy routing as a stateless solver API behind your TMS, keep your constraint model in your own code, and you can swap solvers without re-platforming.

The intraday reoptimization cadence matters more than the headline optimality gap. A plan that is 3% cheaper at 6 a.m. and frozen all day loses to a plan that re-solves when a truck breaks down, a dock backs up, or a same-day order lands. Evaluate solvers on re-solve latency against your live network size, on how faithfully your real constraints can be expressed, and on whether dispatchers can see why a route changed — the same explainability constraint that governs planner-facing forecasting. Carrier selection is the softer half of the lane: scoring carriers on landed cost, tender-acceptance behavior, on-time performance, and claims history is a well-behaved supervised-learning problem on data most shippers already have in their TMS. The failure mode is organizational, not technical — a procurement team that negotiated annual contracts will not let an algorithm tender freight off-contract, so the model must optimize within the contracted routing guide and quantify the cost of the guide itself at renewal time.

Warehouse automation: robotics, slotting, and picking

Inside the four walls, AI shows up in three tiers of increasing capital intensity. Slotting optimization — deciding where each SKU lives — is pure software: models trained on order history, product affinities, and seasonality re-slot continuously instead of the traditional quarterly manual exercise, shortening pick paths without touching hardware. Picking optimization is the middle tier: sequencing pick paths, batching orders, and guiding human pickers through wearables or smart carts. The top tier is robotics — autonomous mobile robots ferrying goods to stations, goods-to-person systems, robotic picking arms — where AI handles task allocation, path planning, and computer-vision grasping, and where the spend shifts from opex to multi-year capex with deep WMS integration and real human-robot safety validation requirements.

The macro numbers counsel humility about the top tier. US warehousing employed 1,834.6 thousand people in July 2026, and the industry's labor productivity — output per hour — grew just 0.1% in 2024.^[2] Whatever individual sites achieve, automation has not yet moved the industry-level productivity needle, which means the median deployment is roughly offsetting its own overhead. The disciplined sequence is therefore bottom-up: take the software-only slotting gains first (they also generate the clean movement data robotics vendors need), instrument picking with clear KPIs — labor cost per order, order accuracy, throughput per hour — and let robotics proposals compete against that improved baseline, not against the pre-optimization status quo. Robot-vendor selection then hinges on interoperability: insist on documented WMS integration and multi-vendor fleet-orchestration support, because a warehouse locked to one robot fleet has recreated the point-service dependency problem with forklifts attached. The adjacent playbook for plant-floor AI — vision inspection, predictive maintenance — is covered in /use-cases/ai-manufacturing-guide, and its central finding applies here too: the failure points are physics, data, and integration, rarely the model.

Sustainability: emissions accounting grows up

Emissions tracking is the newest lane and the one with the most regulatory whiplash. The structure is set by the standard greenhouse-gas accounting hierarchy: Scope 1 covers direct emissions from what you own and burn, Scope 2 covers purchased energy, and Scope 3 covers everything upstream and downstream in the value chain — which for a manufacturer or retailer is where most of the footprint sits, spread across suppliers who report inconsistently or not at all. That data shape is exactly what the AI in this lane does: normalize heterogeneous supplier and logistics data, estimate gaps with emission-factor models, and flag anomalies for human review. Microsoft Sustainability Manager is representative of the platform pattern — automated data connections into business systems, calculation models across Scope 1, 2, and 3, dashboards per scope, and a preview Copilot for querying the data in natural language.^[10]

The US regulatory picture inverted mid-flight. The SEC adopted climate-related disclosure rules on March 6, 2024, requiring large accelerated and accelerated filers to disclose material Scope 1 and Scope 2 emissions with assurance — notably stopping short of a Scope 3 mandate.^[11] Then, on May 29, 2026, the Commission proposed rescinding those rules in their entirety, on the stated ground that they "exceed the scope of the agency's statutory authority," with a 60-day comment period following Federal Register publication.^[12] The tempting read is that emissions AI can wait. The strategic read is narrower: the US federal disclosure mandate is in retreat, but the data demand did not originate there and does not die there — large customers cascade Scope 3 questionnaires down their supply base, other jurisdictions maintain reporting regimes, and the routing lane already prices fuel burn. Build the emissions ledger as a byproduct of the logistics data you are cleaning anyway, and regulatory volatility becomes a timing question instead of a rebuild.

Honest objections

The steelman against aggressive supply chain AI spend deserves a fair hearing, because parts of it are correct. First, the productivity evidence at the sector level is thin — warehousing's 0.1% labor-productivity growth in 2024 is not the curve a decade of automation marketing promised.^[2] Second, attribution is genuinely hard: inventory and service-level outcomes move with demand shocks, supplier behavior, and planner overrides, so a forecasting vendor's before/after chart rarely isolates the model's contribution — insist on holdout comparisons against your incumbent method, not against a strawman. Third, adoption base rates are modest: even among large firms, 37% AI use as of May 2026 means a majority of your peer set is still running on spreadsheets and heuristics — and competing.^[1] Fourth, the vendor landscape is demonstrably unstable; the retired and rebranded services documented above were all sold as strategic platforms. None of this argues for doing nothing. It argues for sequencing investments so that each one leaves a durable asset — clean data, instrumented baselines, portable constraint models — even if the vendor attached to it does not survive.

The read

Treat the five lanes as one program with three buy surfaces, and match deliberately. Suite modules (the Dynamics 365 and Connect Decisions pattern) win when your planning organization lives in that suite and you value workflow over algorithmic control.^[7][5] General ML platforms win for forecasting and inventory when you have the engineering bench to own pipelines and monitoring — that is where the retired point services are herding their customers anyway.^[4] Narrow solver APIs win for routing, where the problem is crisply specified and statelessness makes vendors swappable.^[9] Sequence by data dependency, not by ROI slideware: forecasting first, inventory parameters second, routing in parallel, warehouse software-tier before robotics, and the emissions ledger built on the data exhaust of all of it. The portable takeaway: in supply chain AI, the durable asset is never the model — it is the cleaned data, the instrumented baseline, and the constraint model you can carry to the next vendor.