The honest answer to how AI helps a supply chain is narrower than the marketing suggests. Machine learning models shorten the time between a signal arriving (a demand shift, a supplier incident, a storm forecast) and someone acting on it. This article focuses on three areas: demand forecasting, inventory control and disruption monitoring. In inventory management, one study found the best results came from models working alongside operations research methods and people, not in place of them. Human judgement stays in the loop because most machine learning models perform badly once their input data moves too far from the data they were trained on.
What is the biggest impact AI has on supply chain management?
Our answer is speed. AI shortens the time it takes to turn a signal into a decision. A 2024 Kinaxis global survey of 1,800 supply chain leaders found that most human-led teams typically take five days on average to respond to a disruption. An agentic disruption-monitoring framework published on arXiv completed the full analysis, from detecting a disruption to recommending mitigation, in a mean of 3.83 minutes per test scenario, on manually supplied articles.
Speed and accuracy are different things, though. A faster forecast that is wrong just reaches a bad order sooner. That is why the rest of this article looks at where the models are right, where they are wrong, and who catches the difference.
AI in supply chain planning and demand forecasting
In one study, linking prediction to the decision it feeds improved both forecast accuracy and operating efficiency. That study, of textile and PPE supply chains, built a framework called HAF-DS. It couples an LSTM demand forecaster with a mixed-integer linear programming optimisation layer, so the replenishment decision is optimised together with the forecast. On the combined dataset it cut mean absolute error from 15.04 to 12.83 (14.7%). Inventory cost fell by 5.4%, stockouts by 27.5%, and service level rose from 95.5% to 97.8%. Against a classical EOQ baseline the gaps were larger. Inventory cost fell by about 15.6%, and the stockout rate dropped from 5.7% to 2.9%.
The authors explain why the coupling matters. Forecasting models are usually trained to minimise prediction error without regard to what happens downstream, so a better forecast does not always produce a cheaper or more reliable operation. In HAF-DS, cost and service-level losses flow back into the forecaster during training.
A second study, MCDFN, combined CNN, LSTM and GRU channels. It was trained on five years of sales from a Bangladeshi retailer and outperformed seven other deep learning forecasters across its metrics. Its test MAPE was still about 20%. That figure is a useful reminder that “best of the benchmarked models” and “accurate enough to automate” are separate claims. The MCDFN authors cite earlier work finding that better forecasts can save 5% to 10% on transportation, storage and production planning.
Large language models add a different form of reasoning. They can use product descriptions, calendar information and qualitative market signals that a formal model has no place for. On real H&M sales data in the InventoryBench study, every method involving an LLM beat the operations research baseline. The authors attribute this to world knowledge such as seasonality improving the demand predictions.
Read these as academic results on specific datasets. They show what the method can do under the authors’ conditions. They do not predict the gain on your SKUs.
Inventory control works best when AI sits on top of operations research
For inventory decisions, the key InventoryBench finding is that hybrids win. InventoryBench compared a classic capped base-stock heuristic from operations research (OR), an LLM, and both combined. Both hybrid orderings beat either method alone. Running OR first and the LLM second scored best overall (0.538), a 21% improvement over OR alone.
The breakdown matters more than the headline. OR alone was the best method when demand really was stationary (0.802), because there the LLM may overfit to patterns that are not real. The LLM-based methods did best where demand changed in a directional, describable way, such as a mean shift or trend. The heuristic itself is not designed for stochastic lead times. It assumes a known, fixed lead time and has no way to detect lost orders.
One example in the paper makes this concrete. The OR heuristic recommended ordering zero units because it counted 936 units as pipeline inventory. The hybrid recognised that several shipments were lost, recalculated effective inventory as 390 units and ordered 210 instead.
The design pattern we would take from this is a division of labour. OR does the base-stock arithmetic it is precise at, and the language model reads context and flags regime changes and lost shipments for a person to confirm. The authors’ own framing supports it: OR provides precision for base-stock calculations, LLMs contribute contextual reasoning, and humans step in when the LLM misreads the situation. A related article covers combining AI with operations research in data analytics.
How AI helps supply chains see disruptions beyond Tier-1
The visibility gap sits deeper in the network than most monitoring reaches. Berger et al. (2023) showed that over one-third of supply chain disruptions emerge beyond Tier-1 suppliers. The monitoring platforms the paper reviews cover only direct supplier relationships. The same paper later puts the deeper-tier share at over 50%, while its own literature review says “over one-third, and in some cases up to half”. We treat the one-third figure as the defensible floor.
The agentic framework mentioned above shows how detection works mechanically. LLM-powered agents read unstructured news, extract the affected countries, industries and companies, and map them onto a multi-tier supplier network. A knowledge graph in Neo4j holds that network down to Tier-4. The risk score itself comes from deterministic calculations over network metrics, not from the language model. The resulting action plan goes to a human reviewer for approval, revision or override.
Across its core tasks the system scored F1 between 0.962 and 0.991. In the paper’s worked Mercedes-Benz scenario, it traced a Tier-2 dependency on Russian palladium through Johnson Matthey and Norilsk Nickel.
The limits are stated just as plainly. The test set was 30 synthesised scenarios on networks centred on three automotive manufacturers, and articles were ingested manually rather than from live news feeds. The knowledge graph was a static snapshot, and the system was not stress-tested under production load.
A separate prediction study shows more modest numbers. A temporal graph attention network trained on six months of maritime logistics data reached a test AUC of 0.761 and a recall of 0.504. Its authors treat recall as the primary deployment metric, because missing a real escalation costs more than a false alarm. A recall of 0.504 means the model missed about half of real risk-escalation events.
Walmart describes its own use of AI to anticipate disruptions. It uses predictive models on historical weather and real-time feeds, plus digital twins of its logistics network, to simulate how goods move under stress.
Real examples of AI in the supply chain, as the companies describe them
The examples below are company statements or vendor reports, not independent measurements. They are not benchmarks you can transfer to your own operation.
| Company | Use | Stated result | Who says it |
|---|---|---|---|
| Amazon | Demand forecasting foundation model adding weather and holiday data to sales history | 10% better forecasts for deal events, 20% for several popular products | Damco Group blog |
| Walmart | Holiday inventory placement trained on sales, searches, page views, macro-weather and local demographics | No figure given; engines “forget” one-off anomalies such as a rare Florida snowstorm | Walmart Global Tech |
| Walmart | Severe-weather response with predictive models and logistics digital twins | No figure given; strengthens rather than replaces planners’ judgement | Walmart Global Tech |
| Unilever | Ice cream production planning from weather data | 10% forecast accuracy improvement in Sweden | Damco Group blog |
| Siemens | Senseye Predictive Maintenance at a milk processing plant | Unplanned stops avoided; no figure given | Damco Group blog |
Walmart says its engines “forget” one-off anomalies such as a rare Florida snowstorm; it gives no performance figure. Deciding which events a model should learn from, and which it should discard, is a design decision. It is not something the model works out for itself.
Predictive maintenance belongs on the list because equipment failures in manufacturing interrupt the workflow too. Models trained on sensor data and maintenance history forecast failures before they happen. Our article on AI in logistics operations covers that side in more depth.
How does AI affect logistics costs?
In these studies, better forecasts lower logistics cost by cutting holding costs, stockouts and expedited shipping. The HAF-DS study measured a 5.4% fall in inventory cost alongside a 27.5% fall in stockouts. Literature cited by the MCDFN authors puts savings from more accurate forecasts at 5% to 10% on transportation, storage and production planning, with backorders down by up to 30%.
Disruption analysis is another cost line. The agentic disruption framework reported a mean cost of $0.0836 per disruption analysed. That compares with what the authors describe as multi-day work by risk, supplier-relationship and procurement staff. The system also needs a multi-tier supplier network, bought from data providers or built from internal records.
Walmart says its systems optimise delivery routes, but gives no cost figure in that statement.
Where AI falls short in supply chain operations
Forecasting models fall short on rare, sudden events, and language models can overreact to noise. In the HAF-DS study, failures occurred primarily during sudden demand shocks. The forecaster under-predicted peaks because rare events were underrepresented in training, and when supplier lead times exceeded learned bounds, the optimiser inflated orders.
The pandemic showed the same failure in production. MIT Technology Review reported that some models trained on normal behaviour stopped working as they should once behaviour changed. One condiment supplier’s inventory system broke when bulk orders arrived that it had never been trained on.
Language models showed the opposite failure in InventoryBench. One model read a single high but ordinary demand draw as a regime shift and kept pattern-seeking after demand returned to its mean. LLMs also struggled with tracking pipeline inventory and were less calibrated than OR to the cost trade-off between understocking and overstocking.
Errors also cascade through multi-agent pipelines. In the agentic framework, a failed entity extraction at the first agent left every downstream agent with nothing to work on.
| Condition | What tends to work | Basis |
|---|---|---|
| Stable, stationary demand | Classical OR heuristic alone | InventoryBench |
| Directional shifts and trends | LLM-informed forecast feeding OR ordering | InventoryBench |
| Stochastic lead times, lost orders | Methods where the LLM makes the final ordering decision | InventoryBench |
| Forecast must drive cost and service | Forecaster trained jointly with the optimiser | HAF-DS |
| Sudden shocks outside training data | Human monitoring; models under-predict | HAF-DS, MIT Technology Review |
| Deep-tier supplier events in the news | LLM extraction with deterministic risk scoring and human approval, tested on 30 scenarios | Agentic framework |
What data do you need before AI is worth deploying?
Integration requires clean, standardised data pipelines across procurement, production and logistics databases. If they are missing, a model cannot fix the problem. A checklist we would run before any build:
- Enough history to cover the cycles you care about. MCDFN found weekly and monthly cyclical features to be its most influential predictors, and HAF-DS lost accuracy with input windows shorter than 15 days.
- Shock periods in the backtest. Rare events underrepresented in training were the main source of failure in HAF-DS.
- A multi-tier supplier network if you want disruption monitoring. The agentic framework cannot work without supplier relationships to around Tier-4, and most ERP systems hold only direct suppliers.
- Live external feeds. Real-time detection needs news and advisory ingestion pipelines that most organisations do not yet have.
- Actual, not promised, lead times. The OR baseline in InventoryBench fails partly because it uses promised lead times and cannot see lost orders.
- A named owner for model drift. One source quoted by MIT Technology Review said a data science team is needed to connect what is happening in the world to what the algorithms are doing.
Is AI in the supply chain a security risk?
AI in supply chains carries real operational risk. LLM deployments carry well-documented risks of hallucination, prompt sensitivity and limited grounding. The agentic framework’s authors answer these with retrieval-augmented generation, deterministic tools for critical calculations and human oversight for high-stakes decisions. A separate early-warning system constrained its LLM to explain only the structured evidence it was given, without drawing on outside domain knowledge.
Recommendations such as replacing a supplier need executive and legal review, escalation protocols and audit trails. In regulated industries a supplier change can also need regulatory approval.
In the US Census Bureau’s Business Trends and Outlook Survey working paper, privacy and security concerns were cited by 6.6% of firms that did not expect to use AI. Our own view is that a multi-tier supplier map is commercially sensitive. Where it goes, including to any hosted model API, should be settled before deployment, not after.
Can AI replace supply chain planners?
One controlled experiment found planners adding value beyond what automated methods achieved alone. In InventoryBench’s controlled experiment, human-AI teams beat the better of human-only and AI-only baselines by about 10.2%. Humans did best when they kept final decision authority with access to both the OR recommendation and the LLM’s reasoning. Where the LLM sat in the pipeline mattered: it worked best as decision support a person could follow, adjust or override, not as the main controller.
That result is not guaranteed, and the paper says so. It cites a meta-analysis finding that human-AI teams frequently fail to beat the better of their two components. It also cites Fildes et al. (2009), who found that small judgemental adjustments to algorithmic forecasts at four supply-chain companies tended to reduce accuracy, while large ones improved it. Our reading is that complementarity depends on the design. It comes from giving planners the right information and clear override points, not from simply putting a person next to a model.
Operator statements point the same way. Walmart states that its inventory systems recommend while the associate is ultimately in charge. The agentic framework routes its action plans to a human reviewer.
Firm survey data shows little sign of AI cutting jobs so far. The Census Bureau working paper found that a substantial fraction of AI-using firms use AI to substitute for worker tasks, yet few report cutting employment. An increase in employment from AI use was more common than a decrease. The ILO’s 2025 Working Paper 140 found that few jobs consist of tasks fully automatable with current generative AI, and that the bundle of tasks in an occupation acts as a buffer. The World Economic Forum’s Future of Jobs Report 2025 expects AI and data processing to create 11 million roles and replace 9 million by 2030. These are economy-wide figures.
The question to settle before choosing a tool
The InventoryBench breakdown suggests a more useful question than “should we use AI”. Ask how much of the time your demand behaves like a stable distribution, how often it shifts in ways a person could describe, and how often it is hit by shocks nobody has seen before. Stable demand favours the OR heuristic. Directional, describable shifts are where language models add value. Shocks are where the HAF-DS forecaster under-predicted peaks. A team that can answer that question from its own history knows where AI belongs in its supply chain. A team that cannot answer it has a data problem to fix before it has a model to choose.
