At six in the evening the shelf is empty, the back room is not, and the assistant on the shopper's phone has already promised the item. As of November 12, 2026, none of the retail AI toolkits released this quarter closes that gap on its own, and the reason is not model quality. So what does a toolkit actually change? Ster Software
A stockout that no model can see
Retail AI adoption sits at 45 percent on BCG's count, and the chains that treat it as a core operating discipline report three times the revenue and EBITDAR growth of the rest. Read the two numbers together and a third one appears: the advantage is not distributed evenly across the 45 percent, it clusters. The clustering follows authority rather than spend. Retail News Asia
That is why a stockout is the honest test of a toolkit. An empty fixture is a decision that was never taken, made visible to a shopper at the moment she was ready to pay. No amount of forecasting accuracy survives a store where replenishment before opening belongs to nobody in particular.
What the 45 percent is really measuring
BCG splits a retail AI transformation 10-20-70: about a tenth is the algorithm, a fifth is data and systems, and the remaining 70 percent is people, process and institutional structure. On that reading the adoption figure describes a management question, and the toolkits now shipping are answering the tenth. Retail News Asia
The failure mode repeats across formats. A buyer overrides the recommendation because it collides with a supplier agreement signed in March. A shift lead does not reset the fixture because the reset is not on the task list. The dashboard records a recommendation delivered, the like-for-like number stays flat, and the next budget round files the category under unproven.
Assortment is the lever with a visible output
Assortment is the clearest case in the research because the result is binary. Getting the range right yields up to a 5 percent increase in like-for-like sales, which against grocery operating margins is a large number, and either the item sells off the fixture or it does not. There is no partial credit and nowhere for a wrong decision to hide. Retail News Asia
Promotion depth and price architecture behave the same way in principle, but they leak. A price that is wrong online can be corrected at the till by a cashier who knows the customer, so the error is absorbed rather than recorded. A shelf gap cannot be absorbed; the shopper simply leaves.
Availability has to be counted at the fixture
Most chains still report availability against the distribution centre, which is the number that stays comfortable while the shelf is bare. On-shelf availability, counted by SKU and by hour at the fixture itself, is the only version that can be handed to a named person as a task. Retail News Asia
| Signal | Counted at | Who can act on it |
|---|---|---|
| Warehouse availability | Distribution centre, daily | Supply planning, next cycle |
| On-shelf availability | Fixture, by SKU and hour | Shift lead, same day |
| Substitution rate | Basket, at checkout | Category manager, next review |
The middle row changes behaviour because it can be scheduled. A shift lead who sees a 12 percent gap on a fixture at six in the evening has an action to take before closing; a manager looking at 96 percent warehouse availability has an explanation instead. The difference between an action and an explanation is where store-level AI budgets either pay back or quietly stop.
The agent layer makes inventory accuracy customer-facing
Agentic commerce adds a second reader to every catalogue. Software agents build baskets against household rules and spending limits by reading attributes, ingredient registries and live inventory across competing domains, which means availability stops being an internal reporting line and becomes an input to a promise made on the shopper's behalf. Troiana
Purchase options were merged into the main agentic commerce branch on November 10, 2026, and checkout inside AI Mode opened to eligible Canadian retailers on November 4, 2026. Neither date changes shopping behaviour overnight, but together they turn machine-readable stock into a distribution requirement rather than an internal nicety. Universal Commerce Protocol
The record an assistant actually reads
An assistant does not walk the aisle. It reads a stock status field, an identifier and a price, and if the status field says available while the fixture is empty, the failure lands on the store that has to explain it at the counter. Vendor kits that ship inventory forecasting as a module do not change that arithmetic, because the module predicts and the store still has to reset. Ster Software
The practical order of work is unglamorous: assign the reset to a shift, count availability at the fixture, and only then decide which toolkit earns the next tranche of budget. Chains that reverse the order buy forecasting they cannot execute, at a moment when 55 percent of the market has not scaled at all. Retail News Asia
FAQ
Why does a stockout persist after a forecasting tool is installed?
Because the forecast has no owner. If nobody is assigned the shelf reset before opening, the prediction is delivered and the fixture stays empty.
Which availability number should a store be judged on?
On-shelf availability, counted at the fixture by SKU and hour. It can be scheduled against a shift, unlike warehouse availability.
Do vendor toolkits replace an operating model?
No. They lower the cost of starting and cover customer service and forecasting, but authority, task assignment and incentives still decide whether anything reaches the shelf.
What changes once shopping agents read the catalogue?
Inventory accuracy becomes customer-facing. A wrong stock field produces a promise the store cannot keep, and the complaint arrives at the counter.
