Leaders Insights
Leaders Insights

Rester au meilleur niveau, un peu chaque jour.

DomainesMarketingDataFinanceIA
RessourcesApprendreTestOutilsBlogGlossaire
© 2026 Leaders Insights — Tous droits réservés.
Formations/AI in retail/AI in retail/AI in store operations and fulfillment
4/4+150 XP

AI in retail

1Demand forecasting and replenishment with AI+1502Personalization and recommendation engines+1503AI-driven pricing and markdown optimization+1504AI in store operations and fulfillment+150

AI in store operations and fulfillment

# AI in store operations and fulfillment

A customer walks into a store on a Saturday afternoon. The cereal shelf is half-empty, a promotional endcap has the wrong price tag, three staff are clustered at the registers while the fitting room line grows, and an online order sitting in the system will not get picked for another hour. Nothing here is broken exactly. Everything here is slightly wrong. That slight wrongness, multiplied across thousands of square feet and hundreds of stores, is where retail margin quietly leaks.

This lesson is about closing those gaps with AI applied to three concrete problems: knowing what is on the shelf, putting the right people in the right place, and turning stores into fulfillment engines.

The store as a data-poor environment

Most retailers know their sales to the penny. What happens on the physical floor is far murkier. A stockout (an item a customer wants but cannot find) often goes unrecorded because the sale simply never happens. Industry estimates commonly put on-shelf availability gaps in the high single digits to low teens as a percentage, though the real number varies wildly by category and store.

The core problem: the shelf is the source of truth, and until recently nobody was watching it continuously.

AI changes the economics of watching.

Computer vision for shelf monitoring

Computer vision is software that interprets images the way a trained employee would, but at scale and without fatigue. Point cameras at shelves (fixed cameras, employee phone photos, or cameras mounted on autonomous floor robots) and a model can answer three questions:

  • Is this product in stock on the shelf?
  • Is it in the right place (planogram compliance, meaning the shelf matches the intended layout)?
  • Is the price label correct?
  • What it actually detects

    A well-tuned system flags an empty facing (the front-row slot for a product) before a human notices, cross-references it against the stockroom inventory, and generates a task: "restock aisle 7, bay 3, item still in back."

    The high-value insight is the split between two very different failures:

    • Phantom stockout: the item is in the backroom but not on the shelf. This is a labor and process failure, and it is fixable in minutes.
    • True stockout: there is no inventory at all. This is a supply and replenishment failure.

    Traditional systems cannot tell these apart. Vision plus inventory data can. That distinction alone changes what you tell the store to do.

    The honest limitations

    Computer vision struggles with reflective packaging, deep shelves, obstruction by customers, and products that look nearly identical (think flavor variants of the same drink). Accuracy also degrades fast if the model is not retrained as packaging changes. Treat vendor accuracy claims as best-case lab numbers, not store-floor guarantees. Always run a pilot in a handful of real stores before committing.

    For a plain-language primer on how this technology works, the Google Cloud overview of computer vision is a solid free starting point.

    🎬 [VIDEO: "How Computer Vision Is Transforming Retail" — youtube.com — a short accessible walkthrough of shelf monitoring and in-store vision use cases]

    AI labor scheduling

    Now to the staff clustered at the registers while the fitting room backs up. This is a scheduling and allocation problem, and it is where AI delivers fast, unglamorous wins.

    Forecasting demand at the right grain

    Legacy scheduling relies on rough weekly patterns and a manager's gut. AI-driven workforce management forecasts labor demand in fine time slices (often 15 or 30 minute intervals) using inputs like historical foot traffic, local weather, promotions, paydays, school calendars, and events.

    The output is not "how many people this week." It is "how many people, doing what, at 2:15pm on Saturday."

    From forecast to a real schedule

    The harder half is constraint solving. A schedule has to satisfy:

    • Employee availability and skills (only some staff can run the pharmacy counter)
    • Labor laws and predictive scheduling rules (some cities require advance notice of shifts; changing a posted schedule can trigger penalty pay)
    • Contract and overtime limits
    • Fairness and employee preferences, which drive retention

    AI can generate schedules that respect all of these at once, something a spreadsheet cannot do well. Here is the shape of the problem in simplified terms:

    minimize:  labor_cost + understaffing_penalty + overstaffing_penalty
    subject to:
      staff_on_shift[t] >= forecast_demand[t]        for each 15-min slot t
      hours_per_employee <= contracted_max
      required_skill_coverage[station] == True        for each station
      advance_notice_rule == satisfied
      employee_availability == respected

    The point is not the math. It is that AI balances competing costs simultaneously: understaffing loses sales and frustrates customers, overstaffing burns margin. The optimum sits between them, and it moves hour by hour.

    The people caveat

    Algorithmic scheduling has a bad reputation where it has been used to squeeze workers with erratic "clopening" shifts and zero predictability. That backfires: turnover is expensive, and in retail it is brutal. Use these tools to improve fit and give workers more input and stability, not less. In several jurisdictions, fair workweek laws now make the abusive version legally risky anyway.

    Store-fulfillment routing

    The online order sitting unpicked is the third gap, and it is now often the most strategic one. Stores are no longer just places to shop. They are miniature warehouses.

    Two dominant models:

    • BOPIS (buy online, pick up in store): customer orders online, staff picks it, customer collects.
    • Ship-from-store: the store fulfills an e-commerce order and ships it, competing with the central distribution center.

    The routing decision

    When an online order comes in, an AI system decides which location should fill it. This is an optimization across:

    • Inventory accuracy at each candidate store (do not promise what the shelf cannot deliver)
    • Distance and shipping cost to the customer
    • Store labor capacity right now (do not dump 40 picks on a store already slammed)
    • Markdown risk (fulfilling from a store overstocked on winter coats in March clears slow inventory)

    A good router might split one order across two stores, or route it to a slightly farther store because the nearest one is out of labor hours. The goal is lowest total cost to serve while keeping the delivery promise.

    The pick path

    Once assigned, AI sequences the pick: the order in which a worker collects items so they walk the shortest route through the store. In a store not designed as a warehouse, walking is the dominant cost. Shaving seconds per item compounds across thousands of orders.

    The whole system depends on trustworthy inventory data, which loops back to computer vision. This is the key architectural insight of the lesson: these three capabilities reinforce each other. Vision feeds accurate availability into the fulfillment router. Fulfillment demand feeds labor scheduling. Better staffing keeps shelves accurate. Deploy them in isolation and each underperforms.

    Vérification des acquis

    1. Why does the lesson describe the physical store as a 'data-poor environment' despite retailers knowing their sales precisely?

    2. In the context of shelf monitoring, what does 'planogram compliance' primarily verify?

    3. The lesson frames retail margin as leaking through 'slight wrongness' rather than outright failures. What is the key implication of this framing for applying AI?

    CHOIX MULTIPLES

    4. Select ALL correct answers. According to the lesson, what can a well-tuned computer vision shelf-monitoring system do?

    Sélectionnez toutes les réponses correctes.

    CHOIX MULTIPLES

    5. Select ALL correct answers. Why is computer vision described as changing 'the economics of watching' the shelf?

    Sélectionnez toutes les réponses correctes.

    Putting it together: a day on the floor

    Return to that Saturday store, now with the stack running.

    Overnight, vision-equipped scans flagged the half-empty cereal shelf as a phantom stockout. A restock task was queued for opening. The labor model, seeing forecasted afternoon traffic plus a local event, scheduled an extra floor associate from 1pm to 5pm and pulled one cashier over to self-checkout support. The mispriced endcap was caught by label-detection and corrected before the promotion drove traffic.

    When the online order arrived, the router checked that the item was genuinely on the shelf (not just in the system), confirmed the store had labor headroom, and released it with an optimized pick path. The customer collected it in fifteen minutes.

    None of these are moonshots. They are the removal of slight wrongness, systematically.

    How to sequence adoption

    For a retailer starting out, a sensible order:

    1. Fix inventory visibility first. Everything downstream depends on it. Start with vision in a few high-traffic, high-stockout categories.

    2. Layer in labor forecasting where sales are most traffic-sensitive.

    3. Add fulfillment routing once inventory accuracy is trustworthy enough to promise against.

    Skipping step one is the most common and most expensive mistake. A fulfillment router built on bad inventory data cancels orders, and canceled orders lose customers faster than slow ones.

    Key Takeaways

    • The shelf is the source of truth, and it was historically unwatched. Computer vision makes continuous shelf monitoring economically viable, and its highest-value output is separating phantom stockouts (a labor fix) from true stockouts (a supply fix).
    • AI labor scheduling wins by matching staff to demand in fine time slices while respecting skills, laws, and fair-workweek rules. Use it to improve stability and retention, not to squeeze workers.
    • Stores are now fulfillment nodes. Routing decisions balance inventory accuracy, shipping cost, labor capacity, and markdown clearance to minimize total cost to serve.
    • These three capabilities compound. Accurate inventory feeds fulfillment; fulfillment demand feeds staffing; good staffing keeps shelves accurate. Deploy them together.
    • Fix inventory visibility before promising against it. A fulfillment engine built on unreliable stock data destroys trust faster than slow service does.

    Précédent

    AI-driven pricing and markdown optimization