AI in warehousing: what's real, and what's marketing

The label changed before the technology did

A demand forecast built on historical sales data, seasonality and promotional lift is not a new idea. Neither is an optimisation model that decides which pallet position a SKU should sit in, or a route solver that sequences a delivery run. These are pattern-recognition and optimisation problems operations research has worked on for decades, running on the same underlying statistics that any data-science course covers. What changed recently is mostly the label on the box, not the mathematics inside it.

That distinction matters for a buying decision, not just for accuracy's sake. A tool doing genuinely new pattern-recognition work on unstructured data, such as reading a camera feed, is a different kind of claim to a spreadsheet-era forecasting model repackaged with a new name. Telling the two apart is most of what this page is for.

What is genuinely working today

Four categories of warehouse software are backed by real, working deployments rather than a slide describing a future state. Demand forecasting and slotting optimisation both earn their keep because they are pattern problems over data an operation already holds — sales history, pick history, seasonality — rather than requiring the system to interpret anything it has not effectively been shown thousands of times before.

Labour scheduling that matches rostered headcount to a forecast pick volume, tied closely to overall labour scheduling and productivity, and route optimisation that sequences a picker's path or a delivery run against real-time constraints, both sit in the same category: known optimisation problems fed by better and more current data than a planner working from memory or a static spreadsheet ever had. Computer vision used for cycle counting and for dimensioning stock as it moves through the building is the newest of the four, but it is deployed and working, not merely promised — a Gartner survey of supply-chain professionals, run in December 2023, found roughly a fifth of respondents had already adopted some form of AI-enabled vision system by that point.

Where the claims get thinner

Two categories deserve real scepticism. The first is anything sold as autonomous decision-making — a system that not only recommends an action but is trusted to take it without a person reviewing the recommendation first. Vendors describe this readily; independently verified examples of it running unsupervised in a live warehouse, at the level of consequence a stock or safety decision carries, are much harder to find.

The second, and the one this site cares about most directly, is AI cameras claimed to detect racking damage specifically. One peer-reviewed academic study demonstrates the underlying technique working under research conditions. Every commercial product claim found in the research behind this page is self-published by the vendor selling it, with no named customer and no independent audit — treat that gap as the honest state of the category, not as a reason to dismiss it outright.

The Gartner forecast, read for what it actually covers

Gartner's 2024 press release forecasting that half of companies with warehouse operations will use AI-enabled vision systems by 2027 gets cited constantly, and it is a genuine analyst forecast, not marketing copy. It is worth reading for what it actually covers, though: Gartner's own recommendation is to start with cycle counting and worker safety monitoring as the simplest, most tailor-made entry point. That is a statement about counting stock and watching for hazards. It is not a statement about structural racking assessment, and nothing in the underlying survey supports stretching it that far.

It is also worth being precise about what a forecast is. A prediction that adoption will reach a certain level by 2027 is not the same claim as a measured result already achieved — the two get run together constantly in trade coverage, and they should not be.

What to ask before buying any of it

The genuinely working categories share a trait: they improve a known, bounded problem with more and better data than a business had before, and their vendors can usually point to a named customer and a stated methodology. The thinner claims share the opposite trait: broad performance percentages with no named deployment behind them. Before buying into either, ask for the same three things a vendor should be able to hand over without difficulty — a named customer willing to be checked, a stated measurement methodology, and, ideally, verification by someone other than the vendor itself. A business weighing whether to model a bigger decision before committing capital, such as an ASRS investment, will find the same evidentiary gap in digital twin claims and should apply exactly the same test — whether the claim comes from a small vendor or from a headline figure quoted from the world's largest fulfilment operator.

None of this is a reason to avoid the categories that are genuinely working. It is a reason to keep the two lists separate, and to notice which list a specific claim actually belongs on before it shapes a budget.

Common questions

Is AI actually new in warehouse operations?

Not mostly. Demand forecasting, slotting optimisation and route planning have run on statistical and operations-research techniques for decades. The label "AI" is largely new; the underlying mathematics for these particular uses mostly is not.

What is the strongest evidence that AI genuinely works in a warehouse today?

Demand forecasting, slotting, labour scheduling and route optimisation all have real, working deployments built on pattern recognition over data an operation already holds. Computer vision for cycle counting and stock dimensioning is the newest of these but is also genuinely deployed, not merely promised.

Can AI cameras detect racking damage today?

One peer-reviewed academic study has demonstrated the underlying computer-vision technique under research conditions. Every commercial vendor claim found in the research behind this site is self-published, with no named customer or independent audit behind the figures.

What should I ask a vendor before buying an AI-branded warehouse product?

Ask for a named customer willing to be checked, a stated measurement methodology, and independent verification if it exists. A vendor confident in its own results will usually have all three ready.

Read next

• system-directed picking
• automated visual inspection of racking
• AI demand forecasting and slotting
• rack-supported building
• TR34 floor tolerance
• AGV vs AMR