Twelve Thousand at Shoprite, No Baseline Underneath

The largest retail AI programme on the African continent is running without a baseline. That is a choice, and the people underneath the deployment will live with it longer than the ones making it.

Last Friday the number carried by South Africa's most valuable food retailer crossed R199 billion of market capitalisation. Business Day's weekend read on 17 October 2026 carried the ledger figures the group had been pointing toward for a year: 3,700 stores, 174,000 employees, R21.7 billion paid in salaries and benefits for the 52 weeks ended 28 June 2026, R270.8 billion in group merchandise sales. Shoprite's own primary filing, the SENS results and dividend declaration released on 1 September 2026, gives the exact figures as 174,430 employees and R270.775 billion in continuing-operations sales. The business grew by 5,491 jobs on the year, according to CEO Pieter Engelbrecht in operational commentary carried by Sunday World on the same date. That is the floor the AI programme is running on.

the gateway, and what it reaches

The architecture is more interesting than any single headline figure. Group Chief Technology Officer Chris Shortt, interviewed by Duncan McLeod on TechCentral's Meet the CIO podcast and published on 8 September 2026, reported that approximately 12,000 Shoprite employees now use generative-AI tools through a platform the group built in-house, called Shoprite AI. The platform acts as a cost-conscious router. Each incoming request is scored on complexity and dispatched to the least expensive model that can handle it. The gateway sits in front of category management, pricing, promotions, replenishment, software development, and customer-facing applications. Shortt described the pattern plainly: core retail decisions use AI daily, and product teams pair one or two developers with several agents. Shortt also described modernising the group's SAP retail core as its largest technology project, which is the substrate the gateway writes against.

Twelve thousand of 174,430 is a share of six point nine percent. The remaining ninety-three percent are packing shelves, running tills, picking Sixty60 orders in Westrock and Montague Gardens and Krugersdorp, stocking the LiquorShops in Soweto, driving between distribution centres in the Western Cape. The gateway does not reach their shift. The question a procurement manager at a competing retailer, or a credit analyst writing on Shoprite paper, is going to be asked by somebody over the next twelve months is whether that reach is a floor the programme is climbing off, or a ceiling built into the architecture.

the clock Shortt did not print

The piece of the Shortt interview most worth sitting with is the paragraph that reads at first as straightforward engineering candour. Shortt said the software development lifecycle at Shoprite has "compressed." He attached no number to the word. He said AI outputs are "probabilistic models, not deterministic finalities," and that validating them remains the user's responsibility. He warned that without watching the economics, costs can "get out of hand very fast." Three honest sentences from the person who signs for the architecture. The figure missing in all three is the one a board reader expects: the percentage, the rand, or the hour saved per engineer per week, measured against a pre-deployment baseline the group logged before the gateway went live.

That absence is not an oversight. It is the current state of workforce-productivity reporting in large retail AI deployments, in South Africa as elsewhere. Vendor telemetry captures usage. Instrumented productivity against a logged baseline is captured almost nowhere. Shoprite's own 2026 full-year results presentation, as summarised on the group's filing page carrying the 9 October 2026 annual report, names ShopriteAI explicitly as the "workforce productivity" line item of its growth strategy, and names Pixie as the personalised shopping assistant on Sixty60. The strategy document does not publish a cycle-time, a defect-rate, or a per-engineer throughput figure against which either initiative can be audited.

pixie, measured in baskets

The customer-facing half of the programme has produced the headline figures the group has been willing to print. TechCentral's earlier coverage of Pixie's launch, developed in-house by ShopriteX, reports that 98 percent of Xtra Savings Plus members used the assistant within three months of its April 2026 launch, that customers added more than four million products to baskets through its swipe-to-add feature, that one customer put 730 products worth R36,236 into a basket, and that another completed a R1,500 shop of ten products in 15 seconds. Chief Strategy and Innovation Officer Neil Schreuder framed the assistant as a "24/7 personal shopping butler." The adoption figure is high. The basket-growth figure is specific. Treat the vendor-side framing with the pinch of salt any first-party number earns.

Stuff's write-up of the Smart Chat upgrade on 13 August 2026 adds voice, text and image inputs to Pixie, rolled out in phases to Xtra Savings Plus subscribers. The gap in that update is the one a plant engineer would spot inside ten minutes. The group has not disclosed Pixie's effect on basket size, order frequency or conversion. Sixty60's growth is unambiguous. Pixie's contribution to it remains unmeasured. Both facts carry at once.

the slowdown finding, read on the merits

Any claim that enterprise AI tools raise workforce productivity has to pass through the piece of literature that measured, carefully, the opposite. Joel Becker, Nate Rush, Tom Cunningham, David Rein and Khalid Mahamud of METR revisited their 2025 slowdown finding on 24 February 2026 with a second round of 57 experienced open-source developers, more than 800 tasks across 143 repositories. The raw numbers in the second run show a speed-up: returning developers minus 18 percent, new developers minus 4 percent. The authors do not report those as findings. Their own written conclusion is that the data is now too compromised to serve as a reliable signal of the current effect of AI tools on developer productivity. Developers refused to work without AI, biasing the sample. Pay was lower in the second experiment. Task types moved between conditions. Time-logging broke when developers ran agents concurrently.

Shoprite's deployment is a different context from METR's open-source repositories. The code base is proprietary. The task mix is retail-industry back-office. The engineers run against a Shoprite-built gateway rather than a public tool. What the METR paper hands a procurement officer at a Shoprite peer is the shape of the measurement problem read without flattery: the second-round numbers that look like a speed-up are the ones the researchers themselves reject as unreliable, because the counterfactual (an experienced developer working without the tools she now relies on) has become harder to construct every quarter. If METR cannot measure productivity cleanly inside a randomised trial, a retailer running 12,000 users on an unaudited gateway cannot measure it cleanly either, no matter what the Shoprite AI dashboard shows.

two peers, two lineages

A reader checking whether Shoprite's gateway shape is the regional pattern should look at the two retailers closest to it on the Johannesburg Stock Exchange. African Business magazine's long read of 22 July 2026 on how AI is rewiring South Africa reports Pick n Pay has launched a conversational grocery assistant it calls Penny, built on Google's Gemini models, integrated into its own app, with Enrico Ferigolli leading the rollout. The piece notes Shoprite has expanded its AI-powered personalised shopping and replenishment tools, but gives no comparative figures for either retailer. The information gap between the two camps is itself the finding: both companies run live AI in production, neither has published a workforce-productivity figure attached to it, and the magazine's long-form review of the sector could not source one. The analysts on the two shares are presently valuing an unmeasured delta.

a Thursday shift in Montague Gardens

If you work a Sixty60 pick station on a Thursday evening in Montague Gardens, your software is a scanner, a trolley-weight sensor, and an order list routed by an inventory layer running on top of the SAP core. The gateway in Cape Town's head office is not on your shift. The Pixie butler is on the customer's phone. The thirty-six percent growth in Sixty60 is in your weekly order count. The fifteen-second basket is the shape of what arrives in your queue. The twelve thousand Shoprite AI users are at a desk somewhere else. Your workday is faster because the orders arrive faster. The group has not published whether its hour-per-order figure moved up, down, or stayed level alongside the headline growth. Your pay for the ordinary hour you worked that shift was R28.79, under the national minimum wage schedule gazetted for 1 March 2026. That number is published. The one above it is not.

the question, left open

At what point does a 174,000-person employer running the largest retail AI gateway on the African continent publish the hour-per-task, the error-rate, and the pre-deployment baseline against which both can be audited? Six point nine percent of the workforce is on the gateway. The other ninety-three percent are on a shift the gateway has reshaped from the top down without being asked to clock from the bottom up.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.