The Living Shelf: AI in Retail Demand Sensing and Inventory Intelligence

Retail demand sensing estimates what may be purchased, where, and when. Replenishment decides whether to order, transfer, allocate, substitute, or mark down under operational constraints. Those are different systems with different errors and authority. A forecast can be statistically strong while the resulting order is poor because lead time, case pack, shelf capacity, cash, spoilage, supplier reliability, or a buyer’s policy was wrong.

AI should make that chain inspectable. It should not turn a probability into an automatic purchasing commitment, nor use protected or sensitive customer profiling simply because granular data might improve a metric.

Define the decision grain and service objective

Start with the decision, not the model. Specify:

SKU, variant, pack, location, channel, and fulfillment node;
forecast horizon and order cadence;
service target and acceptable stockout risk;
shelf, back-room, warehouse, and working-capital limits;
minimum order, case pack, lead time, review period, and delivery calendar;
substitution, allocation, markdown, and return policy;
perishability, seasonality, launch, and end-of-life behavior;
buyer, planner, store, and supplier authority.

A store-SKU forecast for tomorrow should not silently drive a twelve-week purchase order. Different horizons need different features and uncertainty. A unit forecast, revenue forecast, traffic forecast, and probability of at least one sale are not interchangeable.

Separate commercial goals. Availability, margin, freshness, working capital, waste, and customer choice can conflict. The optimization layer needs explicit priorities and constraints rather than a single opaque “inventory health” score.

Build an inventory and demand ledger

Point-of-sale data is not demand when customers could not buy what they wanted. Build a time-aware ledger connecting:

transaction, return, cancellation, and substitution;
on-hand, on-order, in-transit, reserved, damaged, held, and display stock;
shelf capacity and planogram;
price, promotion, coupon, placement, and media;
stockout interval and lost-sales estimate;
supplier order, confirmation, shipment, receipt, and short;
transfer, adjustment, cycle count, and shrink event;
product, store, channel, calendar, and location master data.

Preserve event time, recording time, source system, unit, currency, and correction history. Inventory snapshots alone cannot explain whether a sale occurred before or after replenishment. Negative stock and large adjustments should enter an exception queue, not be treated as ordinary demand signals.

AI data quality and observability is central because a stale on-hand balance can make an accurate forecast operationally useless. Monitor latency, duplicate transactions, master-data changes, missing store-days, unit conversion, and reconciliation with finance and warehouse systems.

Correct for censored demand and changing conditions

Observed sales are censored by availability. If an item was absent for six hours, zero sales during that period do not imply zero demand. Estimate lost sales with explicit methods and uncertainty using comparable in-stock periods, nearby stores, substitutions, traffic, or customer requests. Do not quietly train on imputed values as fact.

Distinguish:

regular demand from promotion lift;
true preference from forced substitution;
return timing from original purchase timing;
one-time event from recurring seasonality;
product launch from mature-item behavior;
clearance from baseline willingness to buy;
data outage from sudden demand collapse.

External signals such as weather, events, search trends, or mobility should have provenance, geographic fit, availability at prediction time, and a lawful usage basis. Backtests must use the vintage that would actually have been known, not revised or future data.

The U.S. Census Monthly Retail Trade Survey provides estimates of U.S. retail sales and end-of-month inventories, including seasonally adjusted and unadjusted statistics. It is useful macro context, not a store-SKU demand feed. The page also documents methods, release timing, revisions, and a 2025 scope change, reminding teams to version external series.

Produce distributions, not one magic number

For each item-location-horizon, store a predictive distribution or quantiles, not only a point forecast. Preserve model version, data cutoff, features available, cold-start method, and flags for promotion, stockout, or structural change.

Use forecast hierarchies carefully. Store-SKU forecasts should reconcile with category, region, channel, and company totals without erasing local patterns. New products may borrow information from attributes and analogs, but the analog choice should be reviewable.

Evaluate with scale-aware and business-relevant metrics:

bias and absolute error by horizon;
quantile loss and interval coverage;
weighted error by units, revenue, or strategic importance;
error during promotion, stockout, launch, and holidays;
calibration of stockout or sell-through probabilities;
stability after source or model change;
performance by item velocity, store type, region, and supplier.

No single metric is sufficient. Percentage error behaves badly near zero; aggregate accuracy can hide severe misses on low-volume critical items.

Turn forecasts into governed replenishment proposals

The replenishment engine should consume the forecast distribution plus current inventory state, lead-time distribution, order constraints, service policy, spoilage, and cost. It should produce an explainable proposal:

proposed quantity and arrival date;
expected demand range over protection period;
current usable stock and confirmed supply;
safety-stock or service-policy contribution;
binding constraints and rounding;
projected stockout, overstock, waste, and cash exposure;
alternatives such as transfer, substitute, markdown, or no order;
approval requirement and latest decision time.

A forecast model should not write directly to the purchasing system. Apply spend, quantity, novelty, confidence, supplier, and category thresholds. New or irreversible commitments require buyer approval. Reversible store transfers may receive more automation after validation.

AI supply-chain optimization can coordinate network inventory, but local recommendations must respect allocation policy and downstream constraints. Optimizing one store can create shortages elsewhere.

Model lead time and supply uncertainty honestly

Planned lead time is not actual lead time. Build distributions from order, confirmation, shipment, customs, appointment, receipt, and put-away events. Segment by supplier, lane, product, season, order size, and service type.

Separate supplier short, carrier delay, receiving congestion, and system lateness. Do not compensate for every data problem by adding safety stock. Root-cause action may be a supplier plan, schedule change, master-data correction, or alternate source.

NIST SP 800-161 Rev. 1 addresses cybersecurity supply-chain risk for systems and organizations. In retail planning, it informs due diligence for forecast vendors, optimization services, data providers, integrations, and the software supply chain; it does not provide a merchandise replenishment formula.

Protect purchase-order credentials and integrations with least privilege, signed requests where appropriate, idempotency, limits, and separation of proposal, approval, and release. Maintain a manual ordering path for outage and vendor exit.

Keep protected profiling out of routine demand sensing

Demand can often be forecast from product, location, time, price, promotion, and aggregate behavior without identifying individuals. Start with the least granular data that meets the operational need. Do not infer disability, health, ethnicity, religion, immigration status, pregnancy, economic distress, or other sensitive traits to decide store assortment, price, promotion access, or service.

The NIST Privacy Framework is a voluntary tool for identifying and managing privacy risk. Use it to map data processing, affected individuals, purpose, retention, access, and potential problems—not as a claim of legal compliance.

For customer-level features, establish purpose, lawful basis, notice, consent where required, opt-out or objection mechanisms, retention, deletion, access control, and restrictions on secondary use. Separate identity from forecasting keys. Aggregate or use on-device and privacy-preserving approaches where feasible.

On-device AI and privacy can reduce central collection for some personalization, but local processing does not by itself make profiling fair or lawful. Test outcomes across locations and customer groups without turning protected traits into operational targeting features.

Treat shelf sensing as a measured observation

Shelf cameras, robots, scales, RFID, and associate scans can estimate availability, facings, placement, labels, and damage. Each sensor has coverage, timing, and failure modes. Preserve image or event evidence, device identity, planogram version, model threshold, and reviewer correction.

A visual gap is not necessarily a stockout: product may be in a customer basket, misplaced, in the back room, or awaiting replenishment. Conversely, a full-looking shelf can contain the wrong variant. Connect sensing to item identity and inventory records, then route uncertain cases to a store task.

Measure precision, recall, time-to-detect, false tasks per aisle-hour, completion evidence, and on-shelf availability after action. Do not rank employees by raw alert counts; layout, traffic, delivery timing, and sensor coverage differ.

Simulate policies before live ordering

Historical replay must reconstruct what was known at each decision time: inventory, open orders, forecast, lead-time estimate, constraints, and approval. Avoid using final corrected inventory or future supplier performance.

Compare the proposed policy with real baselines such as seasonal naive forecasts, current min-max rules, buyer orders, and constrained variants. Simulate lost sales, holding, waste, markdown, transfer, expedite, and ordering costs. Stress demand spikes, supplier outage, delayed data, promotion changes, and model failure.

Then run a shadow pilot. Generate recommendations while buyers continue the existing process. Record acceptance, override, reason, and downstream outcome. An override is not automatically model error; the buyer may know about a display, contract, competitor, or local event absent from data.

The voluntary NIST AI Risk Management Framework, whose version 1.0 NIST says is being revised, can structure governance, context mapping, measurement, and management. It does not certify a retail model or transfer buying authority.

Monitor decision and business outcomes separately

Forecast metrics show predictive performance. Decision metrics show whether proposals were usable:

recommendation acceptance and reasoned override;
quantity changed at approval;
late or duplicate order prevention;
constraint and rounding errors;
time from recommendation to release;
percentage of automated actions reversed safely;
approval and integration failures.

Business outcomes include on-shelf availability, fill rate, stockout duration, lost-sales estimate, inventory turns, days of supply, aged stock, markdown, spoilage, working capital, gross margin, supplier service, and customer substitution. Segment by item, store, region, category, horizon, and intervention.

Use controlled pilots where possible. Comparing a new model after a favorable season with an old model during disruption proves little. Guardrails should prevent the system from improving availability by flooding stores with inventory.

Govern change and phase automation

Merchandising, planning, supply chain, store operations, finance, privacy, security, data, and procurement should share ownership. Maintain an inventory of models, features, data providers, item-location scope, policies, thresholds, integrations, approvers, evaluations, and fallback rules.

Version source data, calendar, product hierarchy, promotions, forecast, lead-time model, cost parameters, constraint logic, and replenishment policy. Re-evaluate after a major assortment change, channel change, supplier disruption, acquisition, price architecture change, or model update.

Roll out in phases:

reconcile inventory and transaction truth for one category;
benchmark forecasts without changing orders;
expose uncertainty and causes to planners;
run replenishment proposals in shadow mode;
pilot human-approved orders in selected locations;
automate only bounded, reversible decisions;
expand after stable service, waste, privacy, and workload results;
keep manual operation and rollback tested.

The goal is not automatic ordering everywhere. It is a controlled chain from observed demand to forecast, constrained proposal, authorized decision, executed movement, and measured outcome.

Source notes

Substantive review completed 2026-07-30. Census retail statistics are presented as U.S. macroeconomic estimates with documented revisions, not as store-level truth. NIST AI RMF and Privacy Framework are voluntary tools; AI RMF 1.0 is identified as under revision. NIST SP 800-161 is scoped to cybersecurity supply-chain risk, not retail replenishment mathematics. Forecasts are separated from inventory decisions and protected customer profiling.