Ocado’s Waste Rate Rose Last Year: Reading Demand-Planning Claims Honestly

How CodoraTech is funded: CodoraTech is supported by advertising and, in some articles, by affiliate links. Where an article contains affiliate links, we say so at the top of that article.

Ocado Group publishes more operational detail about machine-learning demand forecasting than almost any comparable retailer, which makes it the right place to test what these systems are claimed to deliver. It also publishes a food waste series in its sustainability reporting, and that series moved the wrong way last year. Both facts come from the company. Holding them together produces a more accurate picture of AI demand planning than either the vendor material or the sceptics allow, and it explains why the strongest available claims in this field remain unaudited.

The published numbers

• Ocado Group says its forecasting engine generates over 70 million supply chain forecasts daily
• Ocado Retail food waste: 0.59 percent of food handled in 2022, 0.43 percent in 2023, 0.49 percent in 2024
• That is a 17 percent reduction against the 2022 baseline, and an increase year on year
• Ocado’s claims of up to 40 percent greater accuracy and up to 98 percent automated ordering carry no published methodology
• Walmart described six named supply-chain AI systems in October 2025 and published no performance figures for any of them

Seventy million forecasts a day, on the company’s own account

An Ocado Group newsroom piece dated 8 October 2025 describes a deep-learning demand forecasting engine organised around a shared control centre trained on billions of sales data points spanning multiple retailers and regions. The company puts throughput at more than 70 million supply chain forecasts every day. The architecture described is encoder-decoder neural networks processing three input classes: static features such as product type, size and storage requirements; past variables including sales history and stockouts; and future variables covering promotions, holidays and delivery slots.

The same page carries two performance claims: the system is “Up to 40% more accurate than traditional retailer systems, designed for bricks and mortar operations” and delivers “Up to 98% of stock ordered automatically with minimal manual intervention.” Both are Ocado’s own marketing claims on Ocado’s own newsroom. Both are hedged with “up to,” and neither comes with a methodology, a defined baseline or a named individual standing behind it. The identical claim appears in a January 2025 Ocado piece on food waste, again unsourced. Repetition across two company pages is not corroboration.

Trade coverage adds description rather than measurement. Supply Chain Digital, on 18 September 2025, quoted Toni Radzikhovska, vice president of product for supply chain at Ocado Group, saying: “At Ocado, we build technology that responds to the demands of a real-time, high-velocity ecommerce environment.” The piece was published ahead of a London conference on 23 and 24 September 2025 and contains no independent figures. It reports the company’s account of itself.

From 0.43 percent to 0.49 percent

Ocado Retail’s sustainability reporting contains the numbers that can actually be tracked over time. Food waste ran at 0.49 percent of food handled in 2024, against 0.43 percent in 2023 and a 2022 baseline of 0.59 percent. Measured against that baseline the reduction is 17 percent, which is the framing the company leads with and which is accurate.

Measured year on year, waste rose. From 0.43 percent to 0.49 percent is movement in the wrong direction, and the company’s reporting does not explain what caused it. Any article asserting that AI forecasting monotonically reduces waste is contradicted by the published series of the retailer with the most sophisticated system in the sector. That is not an argument against the technology. It is an argument against treating a single year’s improvement as a trend, and it is the sort of detail that disappears from vendor decks.

One further detail on that page repays attention. Ocado Retail attributes its waste performance to “accurate forecasting,” to “unique technology” that “optimises stock management,” and to “automated replenishment.” It does not use the words artificial intelligence or machine learning anywhere in that explanation, even while the group newsroom describes the underlying system in neural network terms. The related operational figure the company does give is that over 80 percent of orders are sold before they arrive.

Walmart names six systems and publishes no outcomes

Supply Chain Dive reported on 7 October 2025 on Walmart’s supply-chain AI portfolio, and the level of architectural detail is unusual.

Indira Uppuluri, senior vice president of supply chain technology at Walmart, told the outlet that “End to end, every segment of what we do is driven by some form of intelligence,” that the digital twin “allows us to evaluate tradeoffs…derive valuable insights,” and that the systems “automatically detect, diagnose, and correct issues in real time,” letting the company flex its supply chain around weather disruption.

The important thing about that article is what is absent from it. There are no percentages and no published performance figures for any of the six systems. One of the most heavily AI-invested retailers on earth will describe its architecture in granular detail and publish no outcome metric at all. That silence is the honest finding, and it is a better thing to report than a number invented to fill the space.

• An internally built multi-horizon recurrent neural network for demand forecasting across planning horizons
• A digital twin providing a performance baseline and a modelling test environment
• Agentic AI for unified inventory visibility across stores, fulfillment centres and facilities
• Computer vision for inbound inventory quality control
• A generative AI routing system directing associates to warehouse disruptions using task-management and skill-profile data
• Adaptive large neighbourhood search models for driver routing

2022, when the best-resourced forecasts missed at the same time

The clearest limit on demand planning is historical. CNBC reported on 7 June 2022 that Target’s inventory stood at nearly 15.1 billion dollars as of 30 April that year, 43 percent higher than the same point twelve months earlier. The company cut its second-quarter operating margin guidance to around 2 percent, from roughly 5.3 percent in the first quarter, in order to clear stock through markdowns.

Chief executive Brian Cornell attributed the position to “a rotation away from items like TVs, small kitchen appliances and bicycles,” compounded by fuel and freight costs and the discounting required to move the goods. He framed the response as deliberate: “We thought it was prudent for us to be decisive, act quickly, get out in front of this, address and optimize our inventory.”

The wider point is that models trained on historical patterns failed at the same moment across the largest and best-resourced retailers in the market, because consumer behaviour changed faster than any training set could reflect. Whatever a forecasting engine does well, it does not eliminate that risk, and a claim that it does is refuted by what happened in 2022.

Four names this research could not stand up

Unilever, Procter and Gamble, Maersk and Zara’s parent Inditex appear constantly in writing about AI demand planning. Research for this article located no published outcome figure for AI demand planning at any of the four, in annual reports, earnings calls or named news outlets reachable in this work. What searches returned instead were vendor blogs and content-farm pages recycling each other without a primary source.

The correct response to that is to leave the gap open. An absent public figure is not evidence that these companies’ systems do not work, and all four are large enough that sophisticated forecasting is close to certain. It is evidence that there is nothing to cite, and an article that manufactures a percentage to fill the space has done its reader more harm than the omission would. The recurrence of the same four names across marketing material with no primary sourcing behind them is itself worth knowing when assessing a vendor’s customer list.

Putting a forecasting claim through five checks

1. Identify who published the number. A supplier newsroom, a customer’s audited sustainability report and an independent trade outlet are three different grades of evidence.
2. Preserve any “up to” in the original wording. It marks a ceiling, and removing it converts a best case into an average.
3. Ask what baseline and what period the claim is measured against, and whether the baseline is defined anywhere.
4. Look for an operational series rather than a single figure, and check whether it moves in one direction. Ocado’s waste rate improved against 2022 and worsened against 2023.
5. Check whether the outcome page and the technology page describe the same thing. Where a retailer credits automated replenishment for a result and only its technology team calls that system AI, the attribution chain is worth following before signing anything.

Sources: Ocado Group · Ocado Retail sustainability reporting · Supply Chain Dive · CNBC · Supply Chain Digital