Use-case guide / Retail
A useful retail prediction has somewhere to go.
A forecast cannot refill a shelf. An inventory alert cannot locate a missing item, and an order summary cannot hand a bag to a customer. Choose retail AI work by following the suggestion into the store: who checks it, what they can change, and which record proves the work happened.
Start at the point of work
Name the action that follows the signal.
Choose one item family, store group and operating decision. Preserve the observation, proposed action, staff response and resulting transaction separately. If nobody can inspect or act on the suggestion, improving the model alone will not resolve the task.
- Sales are an observed result
- Low sales may reflect weak demand, missing stock, a closed store or incomplete data. Investigate those conditions before interpreting a quiet period as a demand signal.
- Stock needs a location and state
- A quantity without its unit, location, timestamp and reservation or hold state cannot support a customer promise. A discrepancy should prompt a check, not an invented adjustment.
- Execution needs its own evidence
- A suggested transfer is not received stock. A pickup summary is not collection. Record what staff accepted, changed, completed or could not do.
Six retail trials
Choose a task the store can finish.
These examples connect analytical assistance to an accountable operating step. Use the current rules or manual process as the comparison, and include the work needed to correct the suggestion.
Review replenishment exceptions
01Store action: Which item-location shortage deserves a planner's attention?
- Starting situation
- Planners inspect many item-store combinations but can investigate only a limited number.
- Required data
- Sales by item, location and period; availability history; current stock states; receipts, lead times, promotions and ordering constraints.
- AI assistance
- Estimate demand or flag unusual consumption, then present the shortage hypothesis with input dates and uncertainty. Keep the forecast distinct from the order calculation.
- Retained authority
- Planners validate assumptions and approve quantities under purchasing, capacity and stock policies. An estimate cannot place an order by itself.
- Completion evidence
- Forecast version, reviewed exception, approved order or transfer, receipt and remaining shortage.
- Failure to test
- A stockout is interpreted as low demand, a promotion is smoothed away, or the recommended quantity ignores case size and receiving capacity.
- Trial measure
- Compare forecast error by horizon and item group, missed shortages, excess stock and planner effort through receipt.
Investigate stock and shelf discrepancies
02Store action: Which item and location should staff physically check?
- Starting situation
- Recorded stock disagrees with shelf observations or repeated picking failures.
- Required data
- Exact SKU and pack, inventory state, recent sales and movements, count history and timestamped shelf observations collected for an approved purpose.
- AI assistance
- Group conflicting signals and prepare a prioritized inspection list. Show the conflicting records instead of converting a visual estimate into a stock fact.
- Retained authority
- Staff identify and count the item; authorized inventory processes approve adjustments. A discrepancy does not establish theft or employee fault.
- Completion evidence
- Original alert, inspection time, verified count, reason and authorized correction or unresolved case.
- Failure to test
- A nearby variant is mistaken for the item, an old image drives a new adjustment, or alerts repeatedly send staff to the wrong location.
- Trial measure
- Measure confirmed discrepancies, missed problems, unnecessary checks and staff effort per resolved discrepancy.
Prepare pickup and store-fulfillment exceptions
03Store action: What needs attention before the customer promise fails?
- Starting situation
- Store teams switch between order lines, picking work and customer messages to understand incomplete orders.
- Required data
- Order and line identity, quantities, reservations, local fulfillment states, promised time, staff work status and permitted contact details.
- AI assistance
- Summarize unresolved lines and suggest the next investigation. Map each source status explicitly, including partial quantities; do not infer collection from a convenient label.
- Retained authority
- Store staff confirm physical work. Authorized systems and people control allocation, substitutions, cancellation, payment and customer messages.
- Completion evidence
- Line-level pick or shortage evidence, accepted resolution, communication receipt and actual collection or shipment record.
- Failure to test
- Partially picked becomes complete, two workers pursue the same line, or a packed status is described as customer delivery.
- Trial measure
- Compare missed promises, failed picks, duplicate work, customer contacts and effort through resolved fulfillment.
Prepare local assortment reviews
04Store action: Which assortment question should the merchant investigate?
- Starting situation
- Merchants compare stores whose sales mix differs, but availability and local constraints make simple rankings misleading.
- Required data
- Item-store sales and returns, availability periods, assortment history, shelf capacity, season and authorized commercial constraints.
- AI assistance
- Find comparable patterns and prepare candidate additions, removals or tests with the comparison group visible. Explain which observations are missing.
- Retained authority
- Merchants decide assortment and supplier commitments. A pattern cannot establish customer preference or justify sensitive customer profiling.
- Completion evidence
- Review cohort, proposal, merchant decision, actual assortment change and later availability and sales observations.
- Failure to test
- An item looks unpopular because it was rarely available, or store differences are hidden by an aggregate ranking.
- Trial measure
- Evaluate comparable store-item periods, availability, sell-through, returns and total change effort; record reasons a proposal was rejected.
Compare markdown scenarios
05Store action: Which price scenario merits commercial review?
- Starting situation
- Merchants need to assess ageing inventory without treating a sales forecast as a price instruction.
- Required data
- Exact item and location, sellable stock, inventory age, costs, approved price history, promotion calendar and applicable pricing constraints.
- AI assistance
- Estimate possible demand under disclosed assumptions and compare scenarios. Keep estimated revenue and margin separate from observed transactions.
- Retained authority
- Authorized pricing owners approve offers and customer-visible terms. Product safety and withdrawal decisions remain separate from markdown planning.
- Completion evidence
- Scenario assumptions, price approval, exact effective offer and subsequent sales, returns and remaining stock.
- Failure to test
- The trial credits a promotion for demand that would have occurred anyway, overlooks costs, or publishes a candidate price without approval.
- Trial measure
- Compare realized contribution, remaining stock, returns, corrections and execution cost against a suitable unchanged process.
Prepare return investigation packets
06Store action: What information does the returns owner need next?
- Starting situation
- Staff repeatedly reconstruct product, order and customer-reported reasons across channels.
- Required data
- Order and item identity, return request, original terms, prior communication, physical inspection and permitted customer records.
- AI assistance
- Group reported reasons and highlight missing evidence or recurring item issues. Preserve the customer's words and distinguish a reported defect from an inspected condition.
- Retained authority
- Authorized staff apply return policy and applicable rights, inspect goods, decide disposition and authorize refunds. No customer fraud label or automatic rejection is proposed.
- Completion evidence
- Return case, inspection, decision, inventory disposition and refund transaction where applicable.
- Failure to test
- A reason code becomes a finding of abuse, an uninspected item returns to sellable stock, or a refund promise is mistaken for payment.
- Trial measure
- Measure missing evidence, repeated contacts, correction effort and time through both case resolution and any financial completion.
The missing evidence
Do not let the suggestion become the record.
The same item can be forecast, reserved, found, moved and sold at different times. Each action needs a record at the level where it happened.
| Trial | Model contribution | Still unproven | Decision owner | Required evidence |
|---|---|---|---|---|
| Replenishment | Demand estimate | Approved or received supply | Planner | Order decision and receipt |
| Stock discrepancy | Inspection priority | Physical count or cause | Inventory owner | Verified count and correction |
| Store fulfillment | Exception summary | Picked, collected or shipped | Store fulfillment owner | Exact line quantities and receipts |
| Assortment | Candidate change | Customer preference or benefit | Merchant | Approved test and comparable observations |
| Markdown | Scenario estimate | Authorized price or realized margin | Pricing owner | Effective offer and transactions |
| Returns | Investigation packet | Condition, eligibility or refund | Returns owner | Inspection, decision and financial record |
Design the comparison
Measure the model and the work separately.
An accurate estimate can arrive too late or create work the store cannot complete. The trial must expose both problems.
- 01
Bound the operating slice
Choose the stores, items, time horizon and decision. Name the person who can act and the ordinary process used for comparison.
- 02
Reconcile the records
Check item and pack identity, location, timestamps, missing periods, returns and stock states. Preserve original data and document every transformation.
- 03
Replay realistic decisions
Use only information available at the decision time. Include promotions, stockouts, partial orders, new items and late updates; compare with a simple baseline.
- 04
Observe staff execution
Run a bounded advisory trial. Record accepted, amended, rejected and unfinished suggestions, plus the physical or transaction evidence that follows.
- 05
Review total value
Compare service, inventory, realized financial results and customer and staff effort. Account for season, availability and other changes before attributing an improvement.
Operating controls
Make correction possible at the store.
These are design requirements for a trial, not evidence that a deployment already meets them.
- Keep source states explicit
- Map each system's fields and partial quantities. Never assume the same status word means the same physical event across systems.
- Control the action
- Give suggestions an expiry and an owner. Authorized workflows must approve inventory, price, order and refund changes and handle duplicate requests.
- Limit personal information
- Use only records needed for the task. Customer or employee profiling, surveillance and consequential scoring require separate decisions and review.
- Keep an ordinary fallback
- Staff need to inspect records, reject advice and complete service during an outage. Preserve corrections and unresolved work for the next shift.
Questions before a trial
Choose the technique after the task is clear.
Prediction, extraction, rules and physical process repair solve different parts of retail work.
- Does every retail use case need generative AI?
- No. Forecasting may use statistical or machine-learning methods; stock calculations and transaction rules should remain deterministic. Generative assistance can summarize evidence, but it should not manufacture quantities or completion states.
- Can sales history be used without checking stock?
- It can be analyzed, but the interpretation needs care. Missing stock, closures and missing data can all affect observed sales. Keep those conditions visible rather than silently treating every quiet period as weak demand.
- Can the assistant tell a customer an order is ready?
- Only through a separately authorized communication workflow using the actual readiness evidence. Microsoft documents distinct and sometimes differently mapped store and headquarters states, including partial quantities. Verify the contract in the retailer's own system.
- How is this different from ecommerce content assistance?
- This guide follows predictions and exceptions into physical store work, allocation and commercial review. Catalog drafting, online support and storefront merchandising have different release checks; the ecommerce guide covers those trials.
- What counts as a successful retail pilot?
- A fair comparison must show useful completed work with acceptable errors, customer impact and total effort. Better forecast accuracy or more alerts alone does not establish higher margin, fewer shortages or a better customer experience.
Source basis
Sources behind the control model.
- 01
Google Cloud
Forecasting inventory and placing it at relevant nodesA May 2023 reference architecture separates demand estimates, profitability, capacity constraints, allocation guidance and merchant overrides. Used as a design example, not evidence of retail results or current product availability.
- 02
Microsoft Learn
Store order fulfillmentDocuments Dynamics 365 Commerce line states, partial quantities, picking, packing and pickup. POS and headquarters labels can differ; these mappings are product-specific.
- 03
Microsoft Learn
Design forecast modelsDocuments input versions, missing-value handling and outlier processing. Signal features are marked preview and some strategies depend on version; this guide does not assume their availability or adopt zero filling as a universal rule.
Start with one real workflow
A Systems Audit is the usual starting point. If the opportunity is already clear, we can move directly into a focused build.
Show Us the WorkflowStart with the free automation readiness checklistOBSERVEQUANTIFYDECIDEBUILD