Why would a 5 percent assortment gain fail to show up on the shelf? Boston Consulting Group splits an AI programme into 10 percent algorithm, 20 percent data and technology, and 70 percent people, process and organisation, which inside a store means the facing gets reset or it does not. Only 45 percent of retailers are scaling the technology at all, even as the ones that do earn three times the revenue and EBITDAR growth, as at November 12, 2026.
The 5 percent sits behind a shelf reset
The payback claim is narrow and specific. BCG's Charlie Reid put the assortment gain at upwards of 5 percent like-for-like when AI combines customer preference, product benefit, store demographics and store constraints down to a single store, with further margin from more own-brand sales and less waste. That number is a merchandising number, not a technology number, and it arrives only after someone moves the product on the shelf and updates the price file behind it.
The scaling picture explains why so few retailers collect it. Only 45 percent are scaling AI across operations, while the operators treating it as a core discipline post three times the revenue and EBITDAR growth. Reid's Asia findings put about 40 percent of respondents citing cross-functional collaboration and speed of execution as the drag, and fewer than 20 percent citing the technology platform. The constraint is inside the building, not inside the model.
| Signal | Figure | Where the availability loss hides |
|---|---|---|
| Retailers scaling AI | 45 percent | Pilots that never reach a planogram change |
| Leader growth premium | Three times revenue and EBITDAR growth | Over 80 percent of budget on two use-case areas, against about half for laggards |
| Assortment uplift | Up to 5 percent like-for-like | Only realised if the facing is reset and the price file agrees |
| Asia execution blockers | About 40 percent cite collaboration and speed, under 20 percent the platform | Merchandiser overrides and unclear decision rights |
| APAC capital allocation | Nearly a quarter of an extra 10 percent to store networks | Space is added while range decisions stay centralised |
| AI spend intensity | Nine in ten invest under 1 percent of revenue or do not track it | No budget line for the reset labour the tools assume |
Where the money actually returns
Merchandising before enterprise
Capital is flowing to the store and the supply chain rather than to back-office functions, because merchandising returns cash fastest. Adjusting assortment, pricing, promotion schedules and vendor negotiations produces money that funds the next rollout, while human resources, procurement and finance tools do not. Reid's advice is blunt: do not start with enterprise AI, because it does not move the profit and loss account much in the first year.
The agentic numbers are the ones to treat carefully. Walmart reports baskets about 50 percent larger through agent-assisted shopping and Kroger customers reaching a final basket five times faster, both early-stage figures from a channel that is still tiny. They matter for availability because an agent that hits an out-of-stock line does not browse an alternative aisle; it drops the item or the whole basket, and the loss never appears as a stockout in the store's own reporting.
Execution is the missing 70 percent
Decision rights and the reset calendar
Reid's example of failure is specific: the best tool fails if merchants do not trust its recommendations, if decision rights are unclear, or if store operations are out of step with the calendar of changes. In availability terms that is a familiar sequence. The recommendation arrives on a Monday, the reset labour is scheduled for the following month, the promotional calendar changes twice in between, and the shelf ends the quarter carrying a range that no longer matches either the plan or the price file.
Two conditions have to be in place before any of this works: clean store-level data and an orchestration layer that lets the merchandising, replenishment and pricing stacks talk to each other inside governance guardrails. Neither is a purchase decision. Both are the reason nine in ten retailers still invest under 1 percent of revenue in AI, or do not track the spend at all, while nearly 70 percent of leaders expect the technology to improve EBITDA within three years.
What to measure before the next pilot
Before another pilot is funded, the useful test is a single cluster run end to end: one assortment recommendation, executed in full, measured on sales per square metre and on the three stockout classes. If the reset does not happen, the pilot has measured the model and not the store, and the 5 percent stays a slide. APAC retailers say they would put nearly a quarter of an extra 10 percent of capital into store networks, so the space will exist; whether the range inside it is decided per store is the open question.
There is a second measurement worth adding, because the channel is moving under the shelf. Online is forecast to pass 10 percent of sales for 61 percent of APAC retailers within three years, against 37 percent today, which means the same facing has to serve a walk-in shopper and a picker. A stockout that costs one sale today will cost a route slot tomorrow, and the cost of that is not in any merchandising business case written this year.
FAQ
Is the 5 percent assortment gain real or a vendor claim?
It comes from BCG's own client work rather than from a software vendor, and it is described as upwards of 5 percent like-for-like, with extra margin from own-brand mix and lower waste. The condition attached to it is execution: the recommendation has to reach the shelf inside the planogram window.
Why do 45 percent of retailers scale AI when the returns look so large?
Because the returns sit in processes that are not owned by the AI team. Merchandising, store operations and the promotional calendar have to change together, and in BCG's survey about 40 percent of respondents name collaboration and speed of execution as the blocker, against fewer than 20 percent naming the platform.
What does agentic commerce change for on-shelf availability?
It removes the shopper's ability to substitute. A person facing an empty slot picks something else, while an agent drops the line or the basket. Walmart's early figure of baskets about 50 percent larger makes that failure more expensive per incident, not less.
