AI in Supply Chain: Where It Works and What It Costs

Ask ten vendors what AI in supply chain means and you get ten answers, most describing software that existed before the label. Here is the plain one. AI in supply chain is the use of predictive, generative, and agentic models to make or support specific operational decisions: what customers will buy, how much stock to hold and where, which vehicle takes which route, which supplier is about to slip, and what a scanned document says. Everything else on the slide is one of those wearing a different hat.

We build logistics and supply chain systems for a living, including HaulBreeze, the supply chain management platform we run as our own product. That gives us a fairly unglamorous view. The model is almost never the hard part. The data underneath it is, and so is the plumbing into the ERP and WMS that already run the operation.

The short version

• AI in supply chain is seven things, not one: demand forecasting, inventory optimization, route and network planning, supplier risk, warehouse perception, asset health, and document extraction. Each attaches to a decision that already has an owner.
• Adoption is wide and shallow. BCG's January 2026 survey of over 180 logistics providers and shippers found around 40% deploying AI beyond pilots, but only 13% with AI embedded in core operations and 10% reporting measurable financial impact.
• The blocker is rarely the model. Gartner's survey of 140 senior supply chain leaders found 56% naming legacy integration as a major challenge and 50% naming a shortage of internal talent.
• Forecasting is the headline use case and the humbling one. In the M5 competition on Walmart data, only 7.5% of teams beat the best statistical benchmark, though those that did beat it won by more than 20%.
• Document processing is the quiet winner. Paper still moves freight: container carriers only committed to 100% electronic bills of lading by 2030, so extraction work has a decade of runway.
• The failure modes are boring: master data nobody cleaned, a forecast nobody trusts, and a pilot with no write-back path into the system of record.

What AI in supply chain actually means

AI in supply chain is the application of artificial intelligence, mostly statistical and machine learning models, to specific operational decisions inside planning, procurement, transportation, and warehousing. It is not a platform you install. It is a set of models, each attached to one decision, each fed by data that mostly already exists in your ERP. The useful way to sort them is by what they produce:

That last row is genuinely new. The first four have been in production since the 1990s, worth remembering when a demo claims novelty. What changed is the cost of compute, the quality of off-the-shelf libraries, and models that read unstructured text well enough to be trusted with a customs declaration.

The AI supply chain pitch, and the working version

The pitch says an AI supply chain is autonomous, self-healing, and end to end. The working version is narrower and more useful: a handful of models each improving one decision by a measurable margin, wired into software people already use.

The gap shows up clearly in survey data. The 2026 MHI Annual Industry Report, produced with Deloitte from 500 supply chain professionals, found 41% of companies currently using AI, up from 30% the year before. Real growth. But Gartner, surveying 140 senior supply chain leaders in November 2025, found only 17% pursuing an immediate transformational redesign of processes, with 83% applying AI to specific use cases or scaling it gradually. An earlier Gartner survey of leaders who had already deployed AI found just 23% had a formal supply chain AI strategy.

Most organizations, in other words, are running use-case-level AI without a strategy, and a small minority have made it stick. That matches McKinsey's 2025 State of AI survey, where 88% report regular AI use in at least one function but only around 6% qualify as high performers attributing more than 5% of EBIT to it. None of which argues against doing the work. It argues against doing it the way most people do.

Where AI in the supply chain genuinely works

Seven applications carry most of the value. The benefits are specific rather than general: less working capital sitting in stock, fewer stockouts, lower cost per delivery, less manual re-keying. They share a shape too. Each is a decision that repeats often, with outcomes to learn from and a recoverable cost of being wrong, and each sits at a specific point in the plan, source, make, deliver, return chain rather than floating above it.

Demand forecasting, the highest-value AI use case in supply chain

Demand forecasting predicts how much of each item will sell at each location in each future period. It is the highest-leverage application because everything downstream inherits it: purchase orders, production plans, safety stock, transport capacity.

The technique that wins in practice is unfashionable. Gradient-boosted trees over engineered features (calendar effects, price, promotions, lags, rolling statistics) beat deep learning on most retail and distribution datasets, and they train in minutes on a laptop. That was the lesson of the M5 competition, run on 42,840 Walmart series and published in the International Journal of Forecasting. The top fifty submissions beat the best statistical benchmark by more than 14%, and the winner by 22.4%.

The same paper carries the warning worth pinning above the desk: only 7.5% of teams beat that benchmark at all. Machine learning applied carelessly to demand data usually loses to well-tuned exponential smoothing, which is free. If a vendor cannot show their forecast beating a naive baseline on your data, they have shown you nothing. We go further into the modeling side in predictive analytics in the supply chain.

Two structural traps come up constantly:

• Sales history is not demand history. A stockout records as a week of low sales. Train on that and the model learns to under-forecast exactly the items you keep selling out of. You need censored-demand handling, or at minimum an out-of-stock flag on every historical row.
• Most SKUs barely move. In a distributor catalog the long tail sells a handful of units a year in unpredictable lumps. Intermittent demand needs its own methods, and a candid conversation about which items deserve a forecast at all rather than a min/max rule.

Inventory optimization, which is where the forecast earns money

A forecast on its own changes nothing. Inventory optimization converts it into a stocking decision, and the conversion is where the money is.

Safety stock is set by forecast error, not by the forecast itself. A model that is slightly less accurate but honestly reports its uncertainty produces better stocking decisions than a confident one wrong in unknown ways. This is why probabilistic forecasts matter more than point estimates, and why the M5 competition scored uncertainty separately.

Multi-echelon inventory optimization extends this across a network, deciding whether a buffer sits at the DC or the store given lead times and correlation between locations. McKinsey has reported inventory reductions of 20% to 30% alongside logistics cost reductions of 5% to 20% for distribution businesses that get this working. Treat those as a ceiling for a mature deployment rather than a promise.

Route and network planning

Routing is where people most often misattribute the AI. The vehicle routing problem is combinatorial optimization, solved by heuristics and metaheuristics that have nothing to do with machine learning. What machine learning contributes is the inputs: predicted travel time by time of day and road segment, predicted service duration at each stop, predicted dwell at a dock.

That division of labor matters when you scope a build. The solver is mostly a solved problem with good open libraries. The predictions are yours to build, because they depend on your lanes, drivers, and customers. We go deeper in our piece on AI route optimization.

Network design (where to put a DC, which lanes to contract) sits at the other end: a once-every-few-years decision with no training data, belonging to simulation rather than a learned model.

Supplier risk and disruption monitoring

Supplier risk models watch external signals (news, filings, port congestion, weather) and match them against your supplier and part master to answer one question: what is about to break, and what does it feed?

McKinsey Global Institute's analysis of 325 companies found that disruptions lasting a month or longer occur every 3.7 years on average, and companies can expect to lose more than 40% of a year's profits every decade as a result.

The hard part is not the model. It is entity resolution. Your vendor master holds the same company as "Acme Corp", "ACME Corporation" and "Acme Corp. GmbH", each with its own payment terms, none tied to the legal entity named in the news article. Every supplier risk project spends its first month there, and the ones that skip it produce an alert feed nobody reads.

Warehouse automation and computer vision

Warehouse AI splits into two things that get bundled. Robotics is mechanical and control engineering with a perception layer; computer vision is that layer on its own, and it earns its keep without a single robot.

The robotics side is growing measurably. The IFR World Robotics 2025 report on service robots recorded 102,900 units sold in 2024 for transportation and logistics, up 14%, making it 51% of all professional service robots sold. Around 81,800 were mobile robots for intralogistics.

The vision side is cheaper to start and often pays back faster: dimensioning cartons on a conveyor, verifying pallet build, catching damage at receiving before it becomes a claim you cannot prove. Each is narrow, high-frequency, and measurable, the profile of a use case that survives.

Asset health and predictive maintenance

Predictive maintenance uses sensor data (vibration, temperature, current draw, cycle counts) to estimate how much life a component has left, so a conveyor or a tractor unit gets serviced before it fails rather than on a calendar.

It belongs here because unplanned downtime propagates. A sorter that stops for four hours during peak does not cost four hours of throughput; it costs a missed cutoff, a re-planned route, and a service-level miss three days later. The trap is well known: you need recorded failures to learn from, and healthy assets produce very few. Most first deployments do anomaly detection against a normal-operation baseline rather than remaining-useful-life prediction, a smaller and more defensible claim. Mechanics in predictive maintenance software.

Document processing, the unglamorous winner

If you want one place to start where the payoff is nearly certain, it is documents. Bills of lading, packing lists, commercial invoices, rate confirmations, proofs of delivery, customs paperwork. Freight still runs on them, and a person still retypes most of them.

Large language models changed the economics here. Template-based OCR needed a configuration per counterparty and broke whenever a carrier redesigned a form. Language models read layouts they have never seen and return structured fields, which makes the long tail of small carriers viable for the first time. The runway is long: container carriers under the Digital Container Shipping Association have committed to 50% electronic bills of lading by 2027 and 100% by 2030, from roughly 1% adoption at the start of this decade.

One design rule decides whether these projects work: measure straight-through processing rate, not accuracy. A model that is 95% accurate and knows which 5% it is unsure about, routing those to review, beats a 97% model that is confidently wrong in silence.

Where AI in supply chain does not work

Being specific about the failures is the fastest way to save a budget. These are the cases where we advise against a model.

Decisions that happen once every few years. Network redesign, plant siting, a make-or-buy call on a new category. No training data exists, the stakes are enormous, and the right tool is a simulation with explicit assumptions a human can argue with. The same applies wherever ground truth arrives too late: if you learn in year four whether a five-year supplier contract was wise, no feedback loop closes fast enough to train anything.

Regime changes. A model trained on three years of history has no opinion about a tariff that landed last month, a canal that closed, or a competitor that exited your category. It will confidently extrapolate the world it was shown. Every forecasting deployment needs an override path and a planner expected to use it.

Anything you have to defend line by line. Customs classification, duty calculation, and regulated safety records need a deterministic, auditable rule. A gradient-boosting model cannot testify. Use AI to propose a classification and route it for confirmation, but keep the record of decision a rule. Governance stops being a slide here too. The EU AI Act classifies systems by risk and attaches documentation, logging, and human-oversight duties to the higher tiers. Most supply chain models sit in the low-risk band; anything touching worker safety or employment decisions does not, and finding that out after the build is expensive. Keep a decision log from day one either way.

Places a rule already works. Hours-of-service compliance, reorder points on stable fast movers, allocation of a scarce item by contractual priority. If a deterministic rule produces the right answer, a model is a liability with a running cost. The same goes for real-time control with hard safety constraints, where a wrong output damages equipment or hurts someone: that belongs to control engineering, with AI confined to the advisory layer.

The data AI in supply chain needs first

This is the section most articles skip and most projects die in. Four things have to be true before a model is worth building.

Master data that agrees with itself. One item master where a SKU is a SKU, not three records created by three buyers. Consistent units of measure, because a model that cannot tell an each from a case from a pallet produces forecasts wrong by a factor of forty. A location master where the site code in the WMS and in the ERP mean the same site. The work is unglamorous, usually takes weeks, and cannot be skipped.

Transactional history with timestamps that mean something. Orders, shipments, receipts, inventory movements, adjustments. The timestamp records when the event happened, not when someone got around to keying it. If receiving transactions all post at 06:00 because that is when the nightly batch runs, you have destroyed the signal you would need to predict dock congestion.

History that is point-in-time correct. The one nobody warns you about. If history tables are updated in place, so a cancelled order disappears and a revised forecast overwrites the old one, you cannot reconstruct what the business knew on a past date. Every backtest will be contaminated by information that did not exist at the time, the model will look excellent in evaluation, and it will underperform the day it goes live. The fix is snapshotting, and that is a decision you have to make before you need it.

Demand, not just shipments. Covered above, and worth repeating because it is the most common silent defect in supply chain training data.

Gartner's finding that 56% of supply chain leaders name legacy integration as a major roadblock is really a finding about this layer. The friction is not the algorithm. It is getting trustworthy history out of systems designed to answer "what is true now" rather than "what did we believe last March".

1. Extraction. Nightly batch or change data capture out of ERP, WMS, TMS, and carrier feeds.
2. Modeling store. Historical fact tables, snapshotted so any past date is reconstructable.
3. Model service. Training, evaluation against a naive baseline, and scheduled inference.
4. Write-back. The path that turns a prediction into a planned order, a min/max change, or a task in the system people already use.

Build or buy your AI supply chain capability

Almost nobody writing on this topic answers it honestly, because they are all selling one side. Here is our version, and we sell the build side, so weigh it accordingly.

Licensed planning suites (SAP IBP, Kinaxis, Blue Yonder, o9, ToolsGroup and the rest) encode a great deal of hard-won domain modeling you would otherwise fund from scratch. If your network looks roughly like the industry norm the vendor modeled, buying is usually correct and the argument ends there.

Building earns its place in three situations. First, when the decision being optimized is the thing you compete on: a 3PL whose margin comes from how cleverly it consolidates freight will not differentiate on a configuration screen everyone else has. Second, when your network shape falls outside what packaged tools represent, such as multi-leg cross-dock flows, consignment inventory at customer sites, or any case where the unit of planning is not a SKU at a location. Third, the one nobody likes to say out loud: the data work is identical either way. A licensed suite still needs the clean item master, the point-in-time history, and the integration layer. If you are funding that regardless, owning the model costs less at the margin than the licence quote implies.

Most mid-market operators, roughly the $50M to $500M distributors and 3PLs we work with, land on a hybrid: a licensed system of record for the standard flow, custom models for the two or three decisions that are theirs alone. That is less satisfying than either vendor's answer, and it survives the year.

How AI supply chain systems get built and integrated

The architecture that works is the four layers above, and the sequence matters more than the tooling. Extraction and the modeling store come first, because everything else reads from them. The model is third. The write-back path is last, and it decides whether any of it survives.

Integration surfaces you will meet. SAP through IDoc, OData, or RFC depending on the vintage. Oracle, NetSuite, and Dynamics through their respective APIs. Warehouse systems from Manhattan, Blue Yonder, or Körber, each with its own idea of an interface. Carrier feeds speaking EDI X12 or EDIFACT, where a 214 status message and an 856 advance ship notice carry data your model wants. None of it is hard. All of it takes time, and the count of these surfaces predicts your timeline better than the model's sophistication does.

Write-back is where pilots die. A model that produces a dashboard is a report. A model that creates a planned purchase order in the ERP, adjusts a reorder point in the WMS, or files an exception task for a named person is a system. The difference is a few weeks of integration nobody scoped, and it is why so many well-built models sit in a notebook a year later. Keep an approval step at first, but build the path.

Shadow mode before live. Run the model in parallel with the existing process for a full seasonal cycle and log both answers. At the end you have a defensible comparison against the incumbent method, and the planners have watched it for months rather than being handed it on a Monday. It is the most effective thing you can do about trust, and it costs nothing beyond patience.

How to know the model is beating the incumbent

Most arguments about whether an AI supply chain project worked are really arguments about measurement, so settle it up front.

For forecasting, drop plain MAPE. It divides by the actual, so it explodes on low-volume items and is undefined on zeros, which describes most of a distribution catalog. Weighted MAPE or MASE survives the long tail. Track bias separately from error: a model 12% off in both directions is a different animal from one 12% low every week, quietly draining your service level. Then run forecast value added, comparing the model against a naive baseline and against the planner's current number. A step in that chain that adds no accuracy is a step to remove, and occasionally that step is the model. Backtest with a rolling origin rather than a single holdout, using only data that existed at the time of each forecast, which is what the point-in-time requirement was for.

None of those are the business metric. Fill rate, working capital, expedite spend, and planner hours are. Agree which one this project moves before it starts, because agreeing afterwards never happens.

If you are earlier than that, AI readiness covers the self-assessment and AI implementation the pilot-to-production gap; our AI development practice does this scoping with clients. For the planning stack it plugs into, see supply chain planning software, and the event layer feeding it in supply chain visibility software.

Generative AI and agents in the supply chain

Generative AI in supply chain has a smaller real footprint than the coverage suggests. Language models read unstructured text well and do arithmetic badly, so the rule is simple: let them read and route, and let solvers and statistical models decide numbers.

Where it earns its place today:

• Document extraction, covered above, and the strongest case by a distance.
• Unstructured supplier and carrier communication. Turning an email saying a shipment will be two days late into a structured exception with a shipment reference.
• Querying operational data in plain language, so a planner can ask which orders are at risk without filing a report request. We built an AI-assisted answering layer on live operational data into Conveya, the auto-transport TMS we rebuilt.

Agentic AI in supply chain goes a step further, putting the model in a loop with tools so it can gather context and act rather than answer once. The use case that works today is exception triage: an alert arrives, the agent pulls the shipment, the order, and the carrier's last status, resolves the routine case, and hands a summarized version of the ambiguous one to a human. Our explainer on how agentic systems actually work covers the mechanics and the failure modes.

Expectations here are running ahead of deployments. Gartner predicts 60% of supply chain disruptions will be resolved without human intervention by 2031, and separately found 55% of supply chain leaders expect agentic AI to reduce entry-level hiring needs. Both are forecasts about a category that mostly has not shipped. Treat them as direction, not as a plan.

What AI in supply chain costs in effort and time

Cost tracks integration surface, not model complexity. A forecasting model on data already sitting in a warehouse is a different project from the same model on data spread across two ERPs, a legacy WMS, and forty carrier feeds, even though the modeling code is identical. A realistic shape for a single-decision build:

Three to six months from start to production is the realistic range for a first use case, and the second is meaningfully cheaper because the first three layers already exist. That reuse is the argument for treating this as a platform investment rather than a series of projects, and it holds only if the first build anticipated the second.

Two lines teams forget to budget: model maintenance, because a demand model degrades as the business changes, and change management, because an override rate above a third means you have bought an expensive suggestion box.

For the broader engineering cost picture, see our breakdown of AI development cost, and for the platform side, what supply chain software costs.

How AI supply chain projects fail in practice

The failure modes are consistent enough to list, and none of them are about the algorithm.

Bad master data. Duplicate items, inconsistent units, site codes that changed in a 2019 migration and were never reconciled. The model trains happily on all of it and produces plausible nonsense. It is the most common cause and the least discussed, because "we need to clean the item master" does not make a good slide.

A forecast nobody trusts. No baseline comparison, no explanation of why a number moved, no shadow period. Planners override it into a spreadsheet, the spreadsheet becomes the real plan, and the model's output never reaches a purchase order.

A pilot that never leaves the pilot. No write-back, no owner in operations, no line in next year's budget. The BCG figures quantify how common this is: 13% with AI in core operations, 10% reporting measurable financial impact.

Buying the platform before defining the decision. The most expensive version of all of these. A capable system arrives, an implementation partner arrives with it, and eighteen months later there is a working installation no operational decision depends on.

Where to start with AI in supply chain

The sequence that works is deliberately unambitious.

Pick one decision that repeats weekly or daily, has a measurable outcome, and a named owner willing to be measured on it. Forecasting for one category and document extraction for one document type are both good first choices, for opposite reasons: forecasting has the highest ceiling, extraction the highest certainty.

Audit only the data that decision needs, not the enterprise data estate. Finding problems on one narrow slice is much cheaper than finding them everywhere at once.

Build the model and benchmark it against the dumbest reasonable alternative: last year's actuals, a moving average, the planner's current number. If you cannot beat that, you have learned something valuable for the price of a few weeks.

Run it in shadow for a full seasonal cycle, then wire the write-back with an approval step. Only then start the second use case, and expect it to move faster.

That is what AI in supply chain looks like when it works: narrow, measured, integrated, and boring in the way production systems are boring. The technology has improved, particularly for reading documents and for probabilistic forecasting. The constraint sits where it usually sits, in data quality and integration, and no model release will fix that for you.