The chain knows and the store guesses
A retail group holds a loyalty file, a stock ledger, a price history and a promotions calendar, and it plans 140 units of a promoted line for a store in a week. The store manager, looking at the last four weeks of the same promotion in that store, has a reasonable view that the number is optimistic. They reason from different data, and the manager's data is closer to the floor.
Retail has a feature that makes this harder than it looks elsewhere. The store is both an execution point and a source of truth, and it is where somebody with local knowledge acts autonomously under pressure. A central function that issues a better plan but removes the store's ability to adapt is moving the decision to a place with less information, which is the opposite of the intent.
Where store-level information is usually lost
- Aggregation. Chain-level forecasting is built on store-level history, but that history is often kept only rolled up, so the reasoning behind a number cannot be re-examined.
- Late and partial capture. A stock count taken once a day, or a recommendation arriving after the truck has loaded, is of little use to a store deciding tonight.
- Definition drift. The same measure is computed differently by head office, region and store, and the disagreement is settled by whoever argues hardest.
- Discarded local knowledge. A manager's view about a catchment is recorded nowhere, so every new planner re-derives it.
Availability is a promise the store cannot keep
Availability is the number retail organisations argue about most and measure least carefully, and it is where a better system can do more harm than good. A store is not out of stock; it has a theoretical position that assumes the unit is sellable, in the back and not promised to an online order. A number ignoring those conditions tells a customer something is available when it is not.
The corrective work is definitional rather than technical. What counts as available, at what point in time, net of what is committed elsewhere, is a decision a group should write down once, and in retail it has to hold across thousands of sites before any of it is worth automating.
Promotion execution fails where the plan is thinnest
A national promotion is executed by a store given a range, a price and a display, which then decides locally how to staff it and when to reorder. Head office sees compliance as a percentage of stores that did as asked. The store experiences it as a trade-off against everything else it must do that week, and resolves it without recording why.
The useful intervention is to capture the trade-off, not tighten the instruction. If a store's deviation from a planned allocation is recorded with a reason, a region can see that the planned figures were unachievable there, and the next plan can be better rather than more insistent.
Scheduling is a local decision with a network consequence
Retail labour is the most visible controllable cost, and the one most often decided on incomplete information. A manager allocates hours across a week dominated by two peaks, a delivery day, a stock-take and a promotion. Head office supplies an average labour percentage and a compliance range, because it cannot see the local week either.
This is a good candidate for decision support: the inputs are structured, the constraint is known, and the output is a proposal rather than a commitment. A system that reads the delivery schedule, the promotion calendar and local trading history, and proposes a week for the manager to adjust, gives the manager something they cannot easily produce while doing their day job. The trap is proposing the final roster, which removes the judgement the manager is paid for.
What should not be automated here at all is equally worth stating. Defining availability and settling the labour compliance range are commercial and employment decisions belonging to the regional operations director, and a group that outsources them to a model has not removed a bottleneck but concealed who made the trade-off.
Reporting should end in an exception list
Retail is the one industry in this collection where the store is the source of the data and the decision-maker at once, so capture has to serve the person making the call rather than the person reporting on it. A useful exception list requires a definition of divergence, a reason code people actually use, and a threshold set by somebody accountable. The regional operations director who owns store performance should set that threshold, because a threshold set centrally is one the regions will work around. The shift from periodic reporting to continuous operational intelligence is described in from monthly reports to continuous operational intelligence.
What to start with on Monday
- Choose one decision currently made from a national average, and identify every store making it.
- Record the baseline: the rate at which stores deviate from plan, the size of the deviation, and the sales result in the weeks after.
- Write down the definitions actually in use, including what counts as available stock.
- Build the exception list for that one decision before improving any forecast, with a named owner per region.
- Test on one region against a comparable control region, and review at ninety days against the baseline.
The measures a store manager would accept are the share of decisions made with a stated reason, the rate of deviation from plan, and labour hours not converted into selling time. A chain that can show those three has moved from monitoring to operating on information.
Retail's constraint is that the decision-maker is close to the customer and the data is not, and that changes the design of any system put in front of them. The organisational work that has to accompany such a system is described in business and organizational transformation, and keeping such a system running rather than abandoning it is in how to make enterprise AI maintainable after the pilot.