Personalisation demos in ninety seconds. A replenishment policy change does not, so the carousel gets funded while the ordering rule still runs on a weeks-of-cover table set in 2019. This framework ranks retail AI by the margin it can move, and names the precondition each use case needs before it pays.
POS, ecommerce and ERP disagree in most estates, one product carries three identifiers and two hierarchies, and the model trains on a sales history that never happened. It is the most common root cause and the least interesting to fix, which is exactly why it survives so many programmes.
A shelf that holds nine facings. A delivery window with two people rostered. A backroom with no room for the case pack just ordered. Shelf capacity, case pack rounding and rostered labour hours belong in the optimisation, because an order nobody can receive is a complaint rather than an improvement.
A model fitted on last autumn degrades quietly through the current one until the post-season review finds it. Tie retraining to the trading calendar, run a champion-challenger test before each season, and monitor drift on inputs as well as error by category. Seasonal drift with no cadence is a slow, unattributed loss.
Forecasting and replenishment, markdown and price optimisation, cross-channel allocation, returns and shrink, then personalisation. Write the order down and hold it, because demo quality and margin impact rank retail use cases in almost opposite orders.
Re-derive safety stock, reorder points, review periods and service levels from the new demand distribution, in the same release, and give the category team authority to change the rule. A better number wired into an unchanged rule produces a more accurate dashboard.
The decision consumes an order quantity for one line at one location, so chain-level accuracy is a reporting artefact. Name one system as the inventory position of record, measure store-level accuracy against physical counts, and give the intermittent tail its own method.
Run a randomised holdout from launch on personalisation and pricing, long enough to cover a full purchase cycle, and report incremental margin against control. Track gross margin and availability together, because either target alone finds the gap within a quarter.
Your replenishment system is probably still running its original weeks-of-cover logic. Safety stock, reorder points, review periods and service levels all have
to be re-derived from the new demand distribution and shipped alongside the model. Until that happens, a better input feeds an unchanged decision and the margin line stays flat.
Before allocation, ship-from-store and any published availability figure, yes. Forecasting and markdown work can start on category-level history while
reconciliation runs in parallel. What you cannot do is make cross-channel allocation decisions on four systems that disagree at store level, because the error costs there are asymmetric.
Its measured incremental effect is usually a fraction of the attributed figure, because much of the recommended-product revenue is cannibalised from products the customer would have bought anyway. The real lift sits in discovery of unfamiliar lines. It is worth funding, fifth, and worth measuring against a withheld group.
Most of the assortment at SKU-store-week is intermittent: long runs of zeros broken by a sale of one or two units. Standard methods fit that badly, and
standard metrics reward forecasting zero. Use a Croston-family or count model that returns a distribution instead, and give the tail its own metric.
It can. A customer-level score thatrestricts refunds is a consequential decision about an identified person, and it will sometimes be wrong. Human review on every account-level restriction, a working appeal path, retained score history and fairness testing including postcode proxies keep it defensible. Build those in the first sprint.
Retail CTOs and COOs holding an AI roadmap that leadership expects to produce margin. It assumes you already run pilots, already have a personalisation programme, and now have to explain which use case moves gross margin return on inventory and which one moves a chart.
Drop your details and we'll send Retailers Fund AI In Demo Order. Margin Moves In The Opposite Order. straight to your inbox - no spam, unsubscribe anytime.
Bring your inventory data and your current AI roadmap. We will tell you which use case pays, and which precondition has to be funded before it does.
Book an inventory review