Roughly 82% of retailers now use AI somewhere in their operation, but the honest version of that statistic is this: most of it is running in marketing and back-office operations rather than anywhere a shopper would notice. The sales floor is the last place AI in retail stores actually arrives. Demand forecasting, inventory allocation, and pricing systems came first because they touch a spreadsheet, not a customer, and a bad forecast costs money quietly while a bad in-store robot costs money loudly.
That gap between adoption numbers and visible change is the single most misunderstood thing about retail AI right now. So it is worth walking through where the spend is actually going.
Where the money is actually going
Break the retail AI market down by application and the ranking is not glamorous:
- Personalised recommendations — about 33% of the market
- Inventory management and demand forecasting — 22.81%
- Customer relationship management — 21.5%
- Supply chain optimisation — 13%
- Dynamic pricing — 9.69%
By underlying technology, machine learning takes just over half the market at 50.2%, natural language processing sits around 21%, and computer vision, the thing everyone pictures when they hear “AI in stores”, accounts for roughly 14%.
Computer vision being the smallest slice explains a lot. Shelf cameras, autonomous checkout, and heat mapping are the applications that get written about, and they are the applications getting the least money. There is a reason for that, and it is not that the technology does not work.
Physical stores are not the legacy channel
A detail that gets lost in most retail tech coverage: about 85.1% of US retail sales still go through physical stores. Not 40%. Not half. Eighty-five percent. And 71% of retailers report they are expanding their physical footprint rather than shrinking it, with forecasts pointing to more than 5,000 net new store openings.
This matters for how you read every AI deployment story. When a chain spends on AI, the store is not a cost centre being managed toward closure. It is the primary revenue channel, which means the tolerance for anything that makes the in-store experience worse is close to zero. A forecasting model that is wrong 8% of the time is an acceptable improvement over a planner who is wrong 15% of the time. A checkout system that fails 8% of the time is a disaster.
That risk asymmetry is the whole explanation for why back-office AI shipped first.
The inventory problem is the real business case
IHL Group puts the global cost of empty shelves and overstocks at about $1.77 trillion a year, with roughly $415 billion of that in North America. Empty shelves alone account for around $690.9 billion of the global figure.
Those are the numbers that justify retail AI budgets, and they are the reason demand forecasting keeps winning the internal funding fight. If a model improves on-shelf availability by even a percentage point across a few thousand stores, the arithmetic works immediately. Nobody has to argue about customer sentiment or brand perception.
Compare that to the shrink conversation, where the data is genuinely weaker than people assume. The National Retail Federation’s last comprehensive shrink figure is from 2022, at about $112 billion. Third-party estimates since then range from roughly $90 billion to $130 billion, which is a spread wide enough to tell you nobody actually knows. If a vendor pitches you loss prevention AI with a confident current shrink number, ask where it came from. I have yet to see one that traces back to a primary source newer than 2022.
What the performance data shows, and what it hides
Retailers described as early AI adopters report about 79% higher sales growth and 34% higher buy-online-pickup-in-store rates than late adopters. Macy’s has reported that shoppers who use its AI assistant spend roughly 400% more than those who do not.
Read that second number carefully. Shoppers who choose to engage with a store assistant were already higher-intent shoppers before they touched it. The 400% figure is measuring who uses the tool at least as much as what the tool does. Same problem with the early-adopter growth gap: companies with the operational maturity to deploy AI well in 2023 were probably already outgrowing their peers.
None of this means the tools do not work. It means the published uplift figures are almost all correlational, and the honest internal number is usually smaller. If you are building a business case, the holdout test is the only version of this data worth trusting, and very few retailers publish theirs.
Brand perception is quietly the bigger driver
Here is the finding I did not expect: 62% of retailers say they are investing in AI for brand perception rather than conversion.
That reframes a lot of deployments. The tablet-based style assistant that nobody uses, the app feature that got a press release and then went quiet, the “AI-powered” label on a loyalty programme that runs on rules written in 2019. Some of that is not failed conversion optimisation. It was never meant to convert. It was meant to signal that the brand is current, to investors and to a certain kind of customer.
Worth keeping in mind when you evaluate a competitor’s shiny in-store AI feature and wonder why your own numbers do not justify copying it. They may not justify it for the competitor either.
Six deployments that are actually working
- Replenishment forecasting at SKU-store level. The most boring and most reliable win. Models that predict demand per item per location, rather than per region, cut both stockouts and markdowns at the same time.
- Shelf-gap detection from existing camera feeds. Not new hardware. Retailers are running computer vision over CCTV they already installed for loss prevention.
- Labour scheduling against forecast traffic. Matching staff hours to predicted footfall by hour, which sounds trivial and is where a surprising amount of margin hides.
- Markdown timing. Deciding when to discount seasonal stock, which historically was a merchandiser’s instinct and is now a model output.
- Returns triage. Routing a returned item to restock, refurbish, or liquidation automatically instead of by hand.
- Search and recommendation on the retailer’s own site and app. Technically not in-store, but it is where the 33% of market spend sits and it feeds store pickup volume.
Notice that five of those six are invisible to the customer. That is the pattern. The systems doing real work in retail AI are the ones nobody photographs. For the broader operational picture, our guide to AI operations automation covers how these forecasting and scheduling layers connect across a business, and our overview of AI data analytics goes deeper on the data quality problem underneath all of it.
What breaks these projects
The failure mode is almost never the model. It is the data underneath it.
Perpetual inventory accuracy is the classic example. A forecasting system assumes the system of record knows what is on the shelf. In practice, inventory records in apparel and grocery are frequently wrong at item level, because of theft, mis-scans, damaged stock written off late, and deliveries received in bulk. Feed a good model bad stock counts and it will confidently order the wrong things.
The second killer is the store-level override. Head office rolls out an AI-driven planogram or replenishment suggestion, store managers disagree with it based on local knowledge that is often correct, and they override it. Within a quarter the system’s recommendations are being ignored at 40% of locations and the reported ROI collapses. Nobody writes a case study about that.
How to evaluate a retail AI vendor
A few questions that tend to separate real products from repackaged dashboards:
- Does it need new hardware, or can it run on cameras and POS data you already have? New hardware roughly triples the timeline.
- What happens when the model is wrong? Ask for the false-positive rate on shelf-gap detection specifically. Vendors quote accuracy; you want the error breakdown.
- Can they name a customer at your scale, in your category, with a holdout comparison? Not a logo wall. A number with a control group.
- Who owns the model output if you leave? Forecasting configurations built over two years are switching costs.
- How does it handle a new store with no sales history? Cold start is where most forecasting tools quietly fall back to a category average.
If you are earlier in the stack and selling online as well, our roundup of AI tools for ecommerce sellers is a more practical starting point than an enterprise retail platform. And if discoverability is part of the problem you are solving, the shift described in GEO vs SEO is changing how retail product pages get found at all.
The realistic outlook
Over the next two years, expect the invisible layer to keep getting better while the visible layer stays experimental. Forecasting, allocation, pricing, and scheduling will absorb most of the budget because they have measurable payback. Autonomous checkout will keep appearing and disappearing in pilots. In-store assistants will keep shipping for brand reasons.
The retailers who get the most out of this are not the ones with the most impressive store technology. They are the ones who fixed their inventory data first and then pointed a decent model at it. That is unglamorous, it does not generate press coverage, and it is where the $1.77 trillion actually lives.
