AI Demand Forecasting in North American Grocery: Seven Cases, None Independently Audited

Scope read as: AI/machine-learning demand forecasting and automated replenishment in North American grocers (US and Canada), selected for MBA operations-strategy teaching. Only Walmart's engineering team published an accuracy result with a stated method (a 52-week backtest), and it measured that result itself. Every other outcome figure comes from the retailer or its vendor. I found no named North American grocer that publicly abandoned a forecasting tool.

Case table

Organization	Sector/Region	Year(s)	What they did	Vendors/tools	Reported outcomes (metric, reporter)	Lessons	Source links	Evidence strength
Walmart U.S. (Walmart Global Tech), with Canada rollout	Mass grocer/general merchant, US and Canada	2018 to 2020	Replaced legacy exponential-smoothing forecasts with an in-house ML platform. Launched in Meat and Produce in June 2018, then all key US departments by July 2020.	In-house; gradient boosting, state-space models, random forests	Backtest over 52 weeks: "~300 bps of improvement in the accuracy"; international markets "~500 bps" (Walmart engineers, Oct 2020). The post gives no waste or in-stock figures.	Success was defined as waste dollars and in-stock %, with accuracy as a proxy. Planners could still edit forecasts, and the team tracked those edits ("touches that degrade forecasts").	Walmart Global Tech, 8 Oct 2020; Walmart Global Tech, 16 Jul 2019	self-reported
Cub (United Natural Foods banner), about 80 stores	Supermarket, Minnesota	c. 2021 to 2022 (inferred)	Rolled out Afresh ordering for produce. A tablet prompts the produce manager for inventory counts, then merges them with a demand forecaster.	Afresh	Produce shrink −18%, produce inventory held about −7%, produce sales +2.5% (Cub CIO, Tech Brew, 3 Oct 2022).	Chosen for small IT change and estimated savings. Afresh's own page describes a 3-month, 100-store pilot, so the store count differs by source.	Tech Brew, 3 Oct 2022; Afresh Cub page	self-reported
Kroger (with 84.51° data science unit)	Supermarket, US	2025	Applied AI in parts of the organization, with executives pointing to shrink.	Internal 84.51°; no forecasting vendor named in sources I opened	Interim CEO Sargent said AI is producing "shrink improvements" (Q2 2025 call, reported Sep 2025). No figure given.	Qualitative executive claim only. Treat as a stated direction, not a measured result.	Grocery Dive, 17 Sep 2025	self-reported
Albertsons Companies (Safeway, Vons, Jewel-Osco, Shaw's, ACME)	Supermarket, US, 2,257 stores	2022 to 2025	Began working with Afresh in 2022 and rolled it out to produce, meat, seafood, bakery and deli, plus daily DC demand forecasts for fresh buyers.	Afresh	No quantified outcome. Albertsons' chief merchandising officer calls it a "game-changer" for keeping favourites in stock (vendor press release, 23 Oct 2025).	Value is in department-by-department rollout and mapping subprimals to cuts for meat. Scale is documented; results are not.	Afresh/PR Newswire, 23 Oct 2025	vendor
Heinen's	Regional grocer, Cleveland, 24 stores	not dated	Chainwide AI fresh ordering with "no changes to existing systems".	Afresh	"$1M annualized savings", 1.4% sales increase, 70 bps shrink reduction (Afresh page; no baseline or window).	Plug-in deployment lowers adoption risk. Figures lack method.	Afresh Heinen's case study	vendor
Loblaw	Food and pharmacy, Canada	not dated	After "10+ years" of a mature forecasting process, adopted ML forecasting to break an accuracy and availability plateau.	Blue Yonder	Outcome unknown (no figures on the page).	Interesting framing: ML as a way past a plateau, not a first step.	Blue Yonder Loblaw session	vendor
Giant Eagle	Supermarket, US	not dated	Used Upshop forecasting for fresh items that previously needed heavy manual intervention.	Upshop	"98% forecasting accuracy" and "3-5% increase in sales in key departments" (vendor case page). Accuracy is not defined.	Headline accuracy without a definition cannot be compared with Walmart's bps gain.	CaseStudies.com, Upshop/Giant Eagle	vendor

Patterns

Evidence is thin where it matters. Only Walmart states a method (52-week backtest) and its own metric. Cub, Heinen's and Giant Eagle report outcomes without baselines or controls.
The AI works through ordering, not forecasting alone. Cub, Albertsons and Heinen's all use Afresh, which combines a forecast with in-store inventory counts. Tech Brew describes the tablet-based count step, which suggests data quality, not model choice, is the constraint (inference).
Human discretion is the recurring friction. Walmart tracked manual forecast edits. A study of an upmarket supermarket chain, as summarized in an abstract, found managers deviated downward from automated proposals after a stockout, which raised the chance of another stockout (Exa abstract of Özdemir and Tenhiälä, MSOM, 2026).
Success metrics differ and are not comparable. Accuracy in bps (Walmart), shrink % (Cub), a self-defined 98% (Giant Eagle) and margin gains (see near misses) measure different things.
Transfer. Walmart and Cub transfer best to a classroom because they state what was done and give at least one number.

Teaching angles

Walmart: build versus buy, and measuring a proxy (accuracy) when the business goal is waste and in-stock. Q: Why track manual edits to forecasts, and what does a planner gain by overriding?
Cub: the CIO's 18% shrink figure does not appear on the vendor page, which shows different pilot-period figures. Q: What would a credible control group look like for an 18% shrink claim? The 3-month pilot versus full-rollout framing is a useful prompt.
Albertsons: scale without published results. Q: What internal metrics would you demand before extending a vendor from produce to deli?
Loblaw: diminishing returns on a mature process. Q: When does a mature process justify a platform change, and what gain would you require?
Cross-case: Q: Design a pilot that separates the model's effect from the store staff's behaviour, given the censorship-bias finding.

Do next

Open the Pacific Coast Collaborative PDF (below) to find which retailers and vendors it covers; it is the closest to a third-party evaluation.
Pull the full MSOM paper from a library; the abstract gives no magnitudes or retailer name.
Check Kroger's and Albertsons' 10-K and earnings transcripts for shrink or in-stock metrics.
Ask Afresh and Shelf Engine for the evaluation method behind their published case figures.

Near misses and open items

Pacific Coast Food Waste Commitment AI study (Dec 2022): the program's headline is a 14.8% reduction in food waste per store across 1,300 stores and two major retailers (program's own figure; I did not read the study itself), a philanthropy-supported effort; retailers and vendor were not named on the pages I read, so it fails the named-organization test. Vendors named in a secondary search summary were Shelf Engine and Afresh (unconfirmed).
Shelf Engine regional grocer: an anonymous grocer reports +31% gross margin growth over a year in the deli (vendor). Anonymized, so excluded. The 31% is relative growth with an unstated baseline.
Afresh meat pilot: "21% greater reduction in shrink" versus non-pilot stores at an unnamed pilot store group. Anonymized.
Save-On-Foods: Scale AI funded project page lists $2.8M of Scale AI funding and $7.2M total investment; no outcomes reported.
Hy-Vee (Manhattan Associates): Progressive Grocer, 11 Nov 2024 quotes an executive with no figures.
Walmart Eden and Ocado/Sobeys: the Queen's Smith article cites Eden (produce quality screening, not forecasting) and Ocado's prediction platform; secondary student-written summary.
Kroger and Ocado (failure, adjacent): a counter-example for strategy, not forecasting. Kroger wrote off about $2.5 billion and paid Ocado $350 million when it unwound automated fulfillment centers (trade newsletter, 28 Jul 2026), citing density problems, not forecast failure. Useful for class on automation economics only.
Failures in forecasting itself: none found in North America. A UK example (suppliers told the adjudicator AI forecasts were "hugely variable") is out of region and supplier-sourced (The Grocer, 10 Mar 2026, not opened by me).
Fresh Thyme Afresh page: returned 404.

Coverage

Not reachable: Forbes Afresh profile (403), INFORMS MSOM full text (403), the Pacific Coast PDF (over 5 MB), and the Northwestern Meijer slides (image-only). The Shelf Engine PDF text could not be read locally, so its 31% figure is not independently confirmed here; recheck it before use in course material.