AI Solutions for Retail

A stock recommendation is useful only if a store can execute it before the selling opportunity disappears. Retail AI should connect what the system believes is available with what staff can find, what customers can buy and what the supplier or distribution centre can actually deliver. The buyer is the retail operations or merchandising manager who owns availability across branches.

A useful first scope within AI solutions is one category, several comparable stores and one daily decision: which stock exceptions deserve action today? This is deliberately narrower than replacing the point-of-sale system or launching personalised pricing. It produces a checked store task, an approved movement or a corrected stock record.

Compare availability with promotion execution and store assistance

A retailer should distinguish three possible purchases. Availability support addresses what can be sold and where. Promotion execution addresses whether an approved offer is represented consistently in prices, stock, labels and staff instructions. Store assistance gives employees a verified answer about a product or policy while serving a customer. Each can be useful, but they require different evidence and should not share a single acceptance score.

For promotion execution, prepare an offer record containing eligible products, branches, dates, exclusions, price basis and required approvals. The assistant can compare approved terms with channel records and collect exceptions for review. A promotion that exists in headquarters' plan may not have reached a shelf label or till rule. Confirm execution in the relevant systems and stores before treating the offer as live.

Judge this alternative through incorrect-price incidents, unresolved branch exceptions, margin after the actual discount and the effort required to correct execution. A sales increase during the promotion does not establish incremental contribution: customers may have brought purchases forward or switched from a higher-margin item. The first project can concentrate on reliable execution without making unsupported claims about causal sales uplift.

For store assistance, begin with a product family that generates repeated questions about compatibility, dimensions, warranty conditions or use. The answer should cite the approved product record and identify when the staff member needs specialist advice. A generated compatibility claim can create a costly return even when it sounds helpful. Evaluate complete product questions and the consequences of wrong answers, not merely whether the assistant responds politely in Arabic.

The choice depends on the branch's constraint. A replenishment system will not fix unreliable price rules. A product assistant will not make an unavailable item saleable. Sample a week of store exceptions, classify the reason and estimate the staff time and commercial exposure for each. Select the problem that has a feasible response and an owner willing to maintain its source records.

Use the same discipline when expanding. Pilot a new category with its own pack sizes, returns and product attributes before assuming the original model transfers. For a new branch, confirm receiving calendars, authority limits and local stock accuracy. Preserve the original trial results separately from later cohorts so a successful early branch does not conceal weaker rollout performance.

Distinguish missing stock from missing demand

Low sales can mean weak demand, a product unavailable on the shelf, stock left in a receiving area or an incorrect product mapping. Training a forecast directly on those observations can reproduce the problem. Capture sales, inventory movements, returns, reservations, waste, promotions and availability evidence at the appropriate store and product level.

Start with a reason-coded queue. A stock discrepancy goes to a count; a shelf gap with stock in the building goes to replenishment; a genuine shortage goes to purchasing or transfer review. A markdown opportunity needs a different commercial decision. Demand forecasting and inventory optimisation explain their respective mechanics; the sector solution must connect them to the daily work of the branch.

Identifiers need to survive pack-size and supplier changes. GS1's EPCIS standard provides a vocabulary for sharing product event information. Whether or not the retailer adopts that standard, an event needs a product, location, time and business meaning. A transfer dispatched is different from a transfer received and saleable.

Make the store action card the deliverable

Example action card for a branch stock exception

Card field	Required content
Item and location	Product variant, pack conversion, branch and shelf area
Evidence	Last sale, system stock, latest count and open movements
Proposed action	Count, shelf refill, transfer or order; quantity and reason
Feasibility	Receiving window, case size, shelf space and usable life
Authority	Store confirmation and any purchasing approval
Closure	Actual quantity, time, outcome and unresolved discrepancy

The card should expose the information available when the recommendation was made. If yesterday's receipt arrives after the recommendation, retain both the original suggestion and the later correction. Otherwise the evaluation will credit the model with information it did not possess. Supply-chain operations covers the coordination needed when purchasing, warehouses and stores share the same exception.

Worked example: a transfer with two competing costs

Illustrative example: branch A needs 18 units before its next replenishment. Branch B has 30 units above its protected requirement, and a case holds six. A proposed transfer of 18 units is feasible only after B confirms its physical stock and both branches confirm timing. Moving 24 units merely because a vehicle has room would change the decision and its economics.

Assume 14 of the transferred units sell in the relevant window and contribute SAR 12 each before transfer cost. If incremental transport and handling cost SAR 80, the observed contribution after that cost is 14 × 12 − 80 = SAR 88. This is not automatically incremental profit: some sales might have occurred at B or later at A. Record those displaced sales, residual stock and any expiry cost before claiming a benefit.

Compare transfer candidates with doing nothing and with the next normal delivery. The cheapest movement can still be unnecessary. For short-life products, saleable life on arrival is a hard constraint; a large forecast cannot make unusable stock valuable.

Evaluate by category and operating conditions

Use historical periods that include promotions, ordinary weeks, launches and supplier interruptions. Reconstruct the data known before each ordering cut-off. Compare against the existing rule and a simple baseline. Store-level errors matter because a good chain-wide average can hide repeated shortages in small branches.

During a live trial, track stock availability, lost selling opportunities where measurable, waste, transfer cost, task completion and staff minutes. Record rejected suggestions with reasons: count wrong, promotion changed, vehicle unavailable, or quantity commercially unsuitable. A rejection is useful evidence, not automatically user resistance.

Control workload as well as recommendation quality. If fifty low-value tasks bury three important ones, the queue has failed. Set a daily capacity per role and sort by the expected value of a feasible action, while preserving urgent safety or quality exceptions outside a financial ranking.

Reconcile the shelf before changing the replenishment policy

A useful store trial includes a deliberately inconsistent stock record. The system shows twelve units, but staff find six on the shelf and four in the stockroom. Two of those ten physical units are reserved for collection, leaving eight available to sell and a two-unit difference from the ledger. The assistant should distinguish physical quantity from quantity available to sell. Moving stock to the shelf may solve a presentation problem without creating additional stock or justifying another order.

Record the count time, reservation state and receiving activity before comparing the figures. A delivery accepted after the count can explain an apparent discrepancy. If the data remains unresolved, assign a count task instead of silently correcting the inventory ledger. The buyer can then decide whether the constraint is inaccurate records, slow shelf replenishment or genuinely insufficient supply. Those require different interventions and different success measures.

Keep commercial authority visible

Product descriptions and Arabic shelf messages must preserve pack size, price conditions and exclusions. Marketing operations owns promotion approval; a stock assistant should not invent a new offer to clear inventory. Quality assurance helps structure checks for wrong quantities, stale inputs and failed writes.

Specify who can alter stock, approve transfers and place orders. A recommendation should become a draft action first, with duplicate prevention and a receipt from the stock system. If a connection fails after approval, the system must establish whether the action occurred before trying again. The manual queue remains available during outages.

A proposal should include the category definition, branch cohort, event reconciliation, action card, approval limits, evaluation period and responsibility for maintaining product mappings. Trusted Data's Saudi retail guide provides related commercial reading across retail decisions. It does not establish performance for this project.

Scope a store-availability pilot using masked product movement samples, current ordering rules and examples of suggestions that store staff cannot execute.