What problem does it solve?
A supermarket decides every day how much of each product to send to each store. Order too little and the shelf is empty, the sale is lost and the customer may go elsewhere. Order too much and fresh food expires, ties up cash and ends up as markdowns or waste. With tens of thousands of products and hundreds of stores, that is millions of small decisions a day.
Traditional replenishment ran on averages, rules of thumb and store staff counting shelves and keying in orders. It copes badly with the things that move demand from one day to the next: promotions, price changes, weather, holidays, new products and one product cannibalizing another. The cost shows up twice, in empty shelves and in waste, and grocers have both commercial and public commitments to reduce food waste.
How does it work?
- Assemble the demand signal. Sales history per item and location, prices, promotions, planned events, weather forecasts, holidays and stock positions are loaded daily.
- Forecast at the level decisions are made. Machine learning models forecast demand per item, store and day, days or weeks ahead, and learn effects such as weather on ice or cannibalization between promoted soft drinks.
- Turn forecasts into orders. A replenishment engine converts the forecast into orders, taking into account stock on hand, shelf life, pack sizes, delivery schedules and the service level chosen for each product.
- Flag exceptions. Unusual forecasts, sudden sales changes and data gaps are flagged for a planner instead of being ordered blindly.
- Clear what is left. Some retailers add dynamic markdowns, raising the discount during the day for products close to their date; Albert Heijn does this with electronic shelf labels.
- Learn every day. Actual sales, waste and stockouts feed back into the next forecast.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Mainstream
- Channels
- API and system to system, Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Lower cost to serve, Revenue growth, Risk and loss reduction, Employee productivity.
Indicative value
A grocery chain with EUR 2 billion in annual sales
EUR 1.1 million to EUR 6 million
Cost of stock no longer written off per year
How this is calculated
Formula: sales * wasteShare * costRatio * wasteReduction. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Annual sales sales, EUR per year | 2,000,000,000 | 2,000,000,000 | The reference retailer. |
| Stock written off as waste or deep markdown, as a share of sales wasteShare, fraction of sales | 0.02 | 0.04 | Editorial assumption for a grocer with a large fresh range. Replace with your own shrink and waste figures. |
| Cost of goods as a share of the sales value costRatio, fraction | 0.7 | 0.75 | Editorial assumption, so that waste is valued at cost rather than at selling price. |
| Reduction in waste from better forecasts wasteReduction, fraction of waste | 0.04 | 0.1 | Editorial assumption, replace with your own. No source on this page measures the waste cut by machine learning forecasting itself. For orientation only: RELEX reports a 4% reduction in fresh spoilage value (fresh products only, a vendor claim) from One Stop's 2019 forecasting and replenishment rollout, before machine learning was added; after adding machine learning, One Stop reports higher ultra fresh availability with no corresponding rise in spoilage. Albert Heijn hopes to cut its total food waste by more than 10% with its forecasting, a target rather than a result; its separate Dynamic Markdown initiative saves 250,000 kilos of food a year, a markdown effect rather than a forecasting one. |
What it leaves out: Counts only waste avoided at cost. It leaves out the sales won back from fewer empty shelves, the working capital released by lower stock, the store hours saved on manual ordering and the cost of the forecasting platform and the data work.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Walmart
United States · Retail and ecommerce · 2025
Walmart Global Tech describes how Walmart's supply chain systems use AI and forecasting models, drawing on signals such as historical sales, seasonality, local demand and weather, to decide which products are needed and where to position them across stores and fulfillment centers before customers order. Walmart says it combines AI foresight with human expertise to refine its demand forecasting, and once an order is placed the Walmart Fulfillment Engine takes over. Walmart gives no forecasting accuracy or inventory figures.
No outcome disclosed.
Albert Heijn
Netherlands · Retail and ecommerce · 2024
Albert Heijn, the leading supermarket chain in the Netherlands with more than 1,200 stores, forecasts daily sales for more than 15 million store and item combinations more than five weeks ahead, which its VP of Product Operations describes as almost one billion predictions a day, with the aim of bringing just enough stock to each store and cutting food waste. The retailer says it hopes forecasting will cut its total food waste by more than 10%, a target rather than a result. A related Dynamic Markdown initiative raises discounts on electronic shelf labels during the day for products close to their date, and the story says it now saves 250,000 kilos of wasted food per year.
No outcome disclosed.
One Stop
United Kingdom · Retail and ecommerce · 2022
One Stop, the Tesco owned convenience chain with more than 900 company and franchise stores in Great Britain, moved store and distribution center forecasting and replenishment to RELEX in 2019; RELEX reports that this first rollout raised store availability by 1.9 percentage points and cut fresh spoilage value by 4%. One Stop then added machine learning forecasting to handle short shelf life lines, weather driven demand such as ice and cannibalization between promoted products. RELEX reports that within four months forecast accuracy rose by 3.17 percentage points at product and week level and 1.82 points at product, store and week level, and One Stop's Head of Supply Chain says availability of ultra fresh products with under three days of shelf life rose 8.5% with no corresponding rise in spoilage.
No outcome disclosed.
Morrisons
United Kingdom · Retail and ecommerce · 2018
The UK grocer Morrisons replaced its manual model of stock replenishment with Blue Yonder's AI powered demand forecasting and replenishment solution, built on Microsoft Azure, which predicts customer demand and orders the right level of stock for its stores. Blue Yonder's customer page says it helped Morrisons increase shelf availability of more than 29,000 products in 130 categories across its 500 stores and headlines a 30% on shelf availability improvement. Technology Record, a publication produced with Microsoft's support, reported in January 2018 that the solution cut shelf gaps in Morrisons stores by 30% and stockholding in store by two to three days.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Several years of sales history per item and location, including promotions and prices
- Accurate stock on hand, deliveries and waste recorded per store
- Promotion, price and range change calendars known in advance
- Supply constraints such as lead times, pack sizes and delivery schedules
Systems to integrate
- Point of sale and ecommerce order data
- Inventory and store stock systems
- Ordering, warehouse and supplier systems
- Promotion, pricing and range planning tools
- Weather and event data feeds
Complexity: High
Forecasting models are mature and available from several vendors. The hard parts are clean, timely data from every store (sales, stock, deliveries, waste), promotion and price calendars the model can trust, and changing the operating model so store staff and planners stop overriding the system out of habit.
- 1
Pick the categories where error costs most
Start with fresh, short shelf life and weather driven categories, where both waste and empty shelves are expensive and the gain over rules of thumb is largest.
- 2
Fix stock and waste data first
A forecast is only as good as the stock figure it starts from. Audit stock accuracy and make recording of waste and markdowns part of the store routine before trusting automatic orders.
- 3
Run in parallel against a holdout
Compare model forecasts and proposed orders with current orders for several weeks, on forecast error, availability and waste per category, and agree the targets before switching.
- 4
Automate with limits and exceptions
Let the system place orders within bounds (maximum change against last week, maximum stock cover) and send exceptions to planners with the reason the forecast changed.
- 5
Bring stores along
Explain why an order looks the way it does and track overrides. Stores that keep overriding need either better data or better explanations, not permission to ignore the system.
- 6
Add markdowns and allocation later
Once forecasts are stable, link them to dynamic markdowns for products close to their date and to allocation across stores and fulfillment centers.
Guardrails
- Order limits per item and store, with larger changes held for a planner
- Minimum service levels set per product so essentials are never cut to meet a waste target
- Human review of forecasts for new products, major promotions and unusual events
- Monitoring of data freshness, so a missing sales feed stops automatic ordering instead of producing zeros
- Override tracking with reasons, reviewed weekly
KPIs to instrument
- Forecast error per category at the level orders are placed (item, store, day)
- Shelf availability and lost sales estimates per category
- Waste and markdowns as a share of sales, per category and store
- Days of stock cover in stores and distribution centers
- Share of orders placed without manual change, and override reasons
Human in the loop
Planners set the service levels, order limits and promotion inputs, review exceptions and approve forecasts for new lines and big events. Store managers can override orders, but every override is logged with a reason and reviewed, because overrides are the main way a good forecast loses its value.
Common failure modes
- Garbage stock data
- Phantom stock (the system thinks it is there, the shelf is empty) stops reordering. Audit stock accuracy and let stores flag empty shelves quickly.
- Promotions the model did not know about
- A promotion or price change missing from the calendar produces a stockout or a mountain of waste. Make the promotion calendar a governed input with deadlines.
- Override culture
- Staff who distrust the system reorder by hand and the gain disappears. Explain orders, measure overrides and fix the data behind them.
- Optimizing waste at the expense of availability
- A tight waste target empties shelves. Set service levels per product and measure both waste and availability.
What are the risks and rules?
EU AI Act
Minimal risk
Forecasting product demand and ordering stock is not listed in Annex III. It would become high risk under Annex III point 4(b) only if the same system allocated tasks to employees based on their individual behavior or personal traits, or monitored and evaluated their performance, for example scheduling store staff by individual productivity. GDPR applies only when loyalty or customer level data feeds the forecasts; item and store aggregates on their own are not personal data.
Controls to put in place
- Documented model ownership, validation per category and monitoring of forecast error
- Change control for model updates, order limits and service levels
- Data quality monitoring on sales, stock and promotion feeds
- Audit trail of automatic orders and manual overrides
Frequently asked questions
- How accurate is AI demand forecasting in retail?
- It depends on the category and the level of detail. RELEX reports that machine learning raised One Stop's forecast accuracy by 3.17 percentage points at product and week level within four months, and 1.82 points at product, store and week level. Accuracy tends to be lower at finer levels of detail, so measure it at the level where orders are placed.
- Can store orders really be automated?
- Yes, within limits. Morrisons replaced its manual model of stock replenishment with Blue Yonder's AI powered demand forecasting and replenishment in its stores, and One Stop manages forecasting and replenishment in one RELEX system that draws store level forecasts automatically into replenishment planning. Keep planners in charge of the rules and the exceptions.
- Does it reduce food waste?
- It can, but published figures are few, mostly vendor claims, and none isolates machine learning. RELEX reports a 4% cut in One Stop's fresh spoilage value from its 2019 forecasting and replenishment rollout, before machine learning was added; after adding machine learning, One Stop reports 8.5% higher availability of ultra fresh products with no corresponding rise in spoilage. Albert Heijn hopes forecasting will cut its total food waste by more than 10%, a target rather than a result. Microsoft's customer story reports that Albert Heijn's Dynamic Markdown initiative, which discounts products close to their date, now saves 250,000 kilos of food a year. Measure waste and availability together, because a tight waste target can empty shelves.
How to cite this page
Blits.ai AI Use Case Library, "AI demand forecasting and automated replenishment for retail", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/retail-demand-forecasting-and-replenishment. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published