Today on The Merchant Desk: agentic commerce is shipping faster than it's working. Walmart's Sparky agent is driving real AOV lift while Starbucks just quietly killed its AI inventory tool — and underneath both, the payments networks are repricing the trust layer for a world where transactions look human even when they aren't.
Walmart CEO John Furner disclosed in Q1 FY2026 results that Sparky's weekly active users doubled YoY, customers using the agent spent 35% more per order than non-users, and units purchased through Sparky have quadrupled quarter-on-quarter. The agent now spans web, mobile and in-store, with personalized replenishment, meal planning and Spanish-language coverage live. Walmart simultaneously confirmed ~50% of US e-commerce fulfilment is automated, with 60% of stores receiving freight from automated DCs — the supply-chain signal feeding Sparky's routing decisions.
Why it matters
This is the cleanest unit-economics datapoint yet for agentic commerce at a tier-1 merchant. The 35% AOV lift and 4x unit growth cut against the consumer-trust skepticism in every survey (16% trust agents to spend autonomously in the PaySpace data). What's underneath it matters more: Sparky only works because Walmart's supply-chain AI is generating reliable inventory and fulfilment signals — discovery, agent and fulfilment have to be co-engineered. For anyone watching how this lands in our region, this is the template Takealot, Checkers Sixty60 and Pick n Pay asap eventually have to answer, and the gating factor will be back-of-house data quality, not the conversational layer.
Starbucks has discontinued its NomadGo-built AI inventory system across North American stores after a nine-month deployment, reverting to manual counts. The tablet-mounted camera and LiDAR rig — a centrepiece of CEO Brian Niccol's turnaround — couldn't reliably distinguish oat milk from dairy or detect items already on shelves. The retirement lands the same week MIT's NANDA initiative reported that 95% of enterprise gen-AI pilots deliver no measurable P&L impact against $30–40bn in spend.
Why it matters
This is the counterweight to every Sparky-style headline. Starbucks isn't gen-AI: it's computer vision, deployed nationally at one of the most capital-rich retailers on earth, and it still failed on the boring edge cases (similar SKUs, occluded shelves) that decide whether an AI system can replace a human workflow. For payments and merchant-tech operators evaluating any agentic ops pitch — pre-loss intelligence, AI-managed inventory, autonomous merchandising — the lesson is to underwrite the failure modes, not the demo. The Starbucks failure also matters because it strengthens the bear case on every back-of-house AI roadmap that hasn't been pressure-tested in real stores.
Visa's Spring 2026 Biannual Threats Report — released this week — documents a structural shift in fraud loss patterns. Network-level defences are holding: token fraud is down 9.6%, enumeration losses down 16%. But scam activity hit nearly $1bn in H2 2025, ransomware incidents are up 26%, and attack-to-payout windows have compressed from days to minutes via AI-generated voice, deepfakes and synthetic content. Loss is migrating from infrastructure attacks to behavioural manipulation of consumers and third-party partners.
Why it matters
Together with Adyen's $1.6T-dataset report from earlier this week, this is now the dominant narrative the networks want merchants to hear: the rails are getting harder to attack, but the humans and ecosystems around them are getting softer. The strategic point for any acquirer or merchant-tech vendor is that fraud spend is migrating up-stack — to identity, intent, and cross-partner coordination — which is exactly the surface area the networks are buying themselves into (BVNK, Agentic Ready, Intelligent Commerce Connect). If your fraud product still scores transactions in isolation, you're solving the part of the problem that's getting easier.
A Ravelin survey of 1,504 fraud and payments leaders at large merchants finds 44% have already integrated agentic commerce protocols (ACP, AP2, UCP and the like), with another 32% planning to within six months. Only 29% describe themselves as 'very prepared' for the fraud and security exposure. Chargeback liability when an agent transacts on a consumer's behalf — and how dispute reason codes map to that — remains undefined.
Why it matters
This puts hard numbers on a structural worry that was qualitative two weeks ago. Adoption is running well ahead of governance: more than four in ten large merchants are taking agent traffic without a clear answer on who eats the chargeback if the agent goes off-script. That's the operating environment Mastercard, Visa, Stripe, Adyen and Fireblocks are all pricing into their 'agentic-ready' products, and it's a wedge for any acquirer or PSP that can credibly own dispute logic for agent flows. For SA operators, the read-across is that we'll be importing both the protocols and the unresolved liability questions — better to engage early with PASA/PayInc on the dispute framework than inherit US-shaped defaults.
Investec set out a strategic push to add 122,000 affluent private-banking clients on top of its current 128,000 SA base by March 2030, targeting a R3bn ($182m) operating-profit lift in the segment. The same week, Investec confirmed it has applied for a full Irish banking licence to expand European corporate and wealth coverage. The two moves together signal a tilt from corporate banking margin grind to relationship-led wealth services on both ends of the franchise.
Why it matters
Investec's reposition is data that the tier-1 SA banks are actively retreating from volume-driven mass-market segments to defend higher-margin wealth, and that's an opening — not a threat — for SA fintech. As Investec, Discovery and the big four crowd the affluent tier, mid-merchant acquiring, mid-market SME banking, and lower-LSM payment acceptance are being structurally underinvested in by the incumbents. Yoco, Capitec, TymeBank, Lesaka and Ozow's various 2026 plays all read better with this context: the big banks are signalling where they don't want to compete.
Treasury and SARB extended the public-comment deadline on the draft Capital Flow Management Regulations from 18 May to 30 June 2026 after industry pushback on forced-asset-disposal language, penalty severity, and the integration of crypto assets into a 1961-era exchange-control framework. The regulations cover 59 registered CASPs and any platform handling ZAR-crypto flows. Parallel reporting positions the package as a deliberate response to corporate liquidity stress (WeBuyCars' R7bn loss is cited) and a tightening of crypto-to-fiat as a capital-export channel.
Why it matters
Treasury is signalling intent to bring digital assets fully inside exchange-control supervision, not to liberalise around them — the deadline extension is procedural breathing room, not a policy retreat. For SA payments operators with any crypto-adjacent settlement, stablecoin or remittance corridor exposure (and for global players like BVNK, Yellow Card and the Mastercard partnership announced last week), the practical question is whether the final rules treat stablecoin rails as 'crypto' for capital-control purposes — that determination changes the economics of every SA-bound stablecoin remittance product currently being pitched. 30 June is now a real diary date.
The Ken's deep-dive on Bajaj Finance shows the Indian NBFC has repositioned its two-decade-old call centre as the data and distribution layer for AI-driven lending: 31 million voice interactions converted into training data, 27 live autonomous agents, and loan-approval capacity scaled from 100k to 600k on a single day. Voice-based lending converts at 3–4x the rate of pure-digital flows. The kicker is the funding side — Bajaj's NBFC deposit base gives it materially cheaper capital than digital-native competitors that must lean on institutional debt.
Why it matters
This is the cleanest counter-example to the 'digital-native eats legacy' narrative in fintech. Legacy distribution and customer data, paired with AI, can compound faster than greenfield digital — particularly when funding economics are also structurally cheaper. The direct read-across for South Africa is the African Bank / Capitec / Discovery Bank cohort: their physical and call-centre footprints are not the liability the venture narrative says they are, provided they treat workflow data as the strategic asset and rewire it through agent layers. For anyone selling AI into SA banking, the pitch isn't 'replace your call centre' — it's 'rebuild it as the data engine for your underwriting moat'.
African identity and fraud-prevention infrastructure player Prembly disclosed it is closing in on $10m ARR, profitable at group level, processing roughly 7m verifications a month across 35 countries. CEO Lanre Ogungbe walked through the company's pivot from biometric payments to compliance-led fraud prevention after a failed merger and explicit market-demand signals. The published lessons are operational: regulatory fragmentation forces market-by-market localisation, currency volatility is a structural rather than tactical risk, and standardised cross-market ops models break.
Why it matters
Paired with this week's Chimoney shutdown post-mortem, Prembly is the positive control — a profitable African infrastructure player that survived because it picked a single defensible vertical (KYC/KYB/AML), localised aggressively per market, and let demand pull the pivot rather than founder conviction. For any merchant-tech or payments operator looking at multi-country African expansion, the Prembly playbook (deep localisation, profitability over growth, regulatory adaptation per market) is more useful than another funding-round headline. It also flags identity infrastructure as a category where SA's mature Smart ID base and pending digital ID regs create a genuine asymmetric advantage — if we don't cede it to a Lagos-based incumbent first.
Bank of Ghana Governor Johnson Asiama announced at ACI FMA World Congress 2026 that the e-Cedi has completed its retail pilot and is moving into cross-border settlement and wholesale payments, framed inside Ghana's broader regional-integration push. In parallel, the AfCFTA Secretariat and Ecobank Group signed an MoU committing $3bn in trade finance over three years across Ecobank's 34-market footprint, with focused programmes for women-led and youth-led SMEs and explicit attention to non-tariff barrier reduction. The 3i Africa Summit in Accra this week framed the same agenda: shift from innovation to integration.
Why it matters
The pan-African payments story has been mostly narrative for three years — PAPSS announcements, mobile-money interoperability briefings, AfCFTA promises. This week it picked up two operational anchors: a CBDC moving into wholesale cross-border, and a $3bn trade-finance commitment from the bank with the broadest African footprint. For SA fintech and merchant operators with continental ambitions, the practical implication is that the rails are arriving faster than the playbook for using them — Ecobank's network plus PAPSS plus an e-Cedi wholesale channel changes how you'd structure a Lagos-Johannesburg-Accra B2B settlement product today versus six months ago. The constraint, as the Africa Prosperity Network and B&FT noted this week, remains physical logistics and senior engineering talent.
At its Relate conference this week, Zendesk launched outcome-based pricing (OBP): customers pay only for verified ticket resolutions, not per-seat licences. The framing — 'agents as a unit of labour, not software' — explicitly shifts AI-performance risk to the vendor. Early-customer data points cited cost reductions of ~30% with no quality regression. The launch lands the same week Anthropic disclosed 54% of new enterprise logos via self-serve by threading Claude across Clay, LeanData, Salesforce, Gong and Ironclad, and Pattern launched its Pi engine for autonomous e-commerce ops on a 77T-datapoint moat.
Why it matters
Three signals in one week that AI is repricing B2B SaaS itself. OBP is the right structural answer to Block's Q1 lesson — if AI is genuinely productive, charge against the output, not the seat. For payments and merchant-tech vendors with AI features bolted onto per-merchant or per-terminal pricing, this is the canary: enterprise buyers will start asking 'why am I paying for capacity when you're charging me for an agent?' The 2026 OPEXEngine benchmark showing no private SaaS cohort clearing Rule of 40 is the macro forcing this — every category will see a vendor try OBP within 12 months, and the winners will be the ones whose unit economics actually survive it.
Pick n Pay upgraded FY guidance to headline losses narrowing 10–20% (49.23–55.39c per share), crediting stronger-than-expected Boxer performance and late-period margin work in the core supermarket. The new development: this guidance upgrade lands the same day as the second material Boxer disposal — ~12.5% sold via accelerated bookbuild for R4.7bn — pushing core breakeven to 2028. Food CPI hit a 13-month low of 2.8%, but NAMC flags a 121% urea spike and 69% Brent move reloading the cost stack in the coming weeks.
Why it matters
The guidance upgrade is real but structurally hollow — Boxer's outperformance is funding the core's survival while the ownership stake compounds downward. The reader already knows the second Boxer disposal and the 2028 breakeven; what's new here is that management is now simultaneously guiding for improvement and accelerating the asset sell-down on the same day, which confirms the dynamic is demand-driven (bookbuild oversubscription) not planned sequencing. The food-inflation reading adds a second wrinkle: a brief CPI window is closing before input-cost pressures from fuel and fertiliser reload. For merchant-tech and payments operators, the implications are unchanged — volume is migrating to discount and informal formats — but the timeline to Pick n Pay's structural resolution is clearer: Boxer proceeds are the only fuel, and the tank is shrinking.
Three Korean internet banks launched under identical regulatory licences have diverged sharply through 2025-2026. KakaoBank posted a record ₩480.3bn ($315m) net profit by shifting to platform-based non-interest revenue. K Bank's net profit fell 12% as its NIM collapsed from 2.4% to 1.4% after July 2024 crypto-deposit rate regulation lifted funding costs. Toss Bank doubled earnings to ₩96.8bn riding zero-CAC distribution through the Toss super-app, with a planned $10bn+ US listing in Q2 2026.
Why it matters
This is the cleanest controlled experiment we have on neobank survival: same licence, same market, three radically different outcomes. The pattern that drops out is that pure-UX neobanks lose, and neobanks embedded in a high-engagement distribution surface (messaging, super-app, telco) win — which is the exact thesis behind Discovery Bank-via-Discovery, Capitec-via-branch density, TymeBank-via-PnP/Boxer kiosks, and GoTyme's targeted $15bn IPO. K Bank's specific failure mode — a single rate regulation collapsing NIM by 100bps — is also a reminder for any SA neobank or digital lender whose funding model relies on a specific deposit-spread arbitrage: that arbitrage can be legislated away on one parliamentary cycle.
Agentic commerce: shipping is not working Walmart's Sparky drives a 35% AOV lift and 4x unit growth, while Starbucks pulls its NomadGo inventory AI after nine months because it can't reliably tell oat milk from dairy. Ravelin finds 44% of large merchants have already integrated agentic protocols but only 29% feel ready for the fraud and liability exposure. Adoption is outrunning reliability and governance — and the gap is where the next category of vendors will be built.
Fraud has crossed the indistinguishability line — and the networks are repositioning around trust, not rails Adyen's $1.6T-dataset fraud report and Visa's Spring 2026 Threats Report land within a day of each other with the same conclusion: attacks now look legitimate transaction-by-transaction. Token fraud is down 9.6%, but scam losses approached $1B in H2 2025. False declines now cost merchants roughly what the fraud itself does. The product implication: identity-and-intent layers, not transaction flagging, become the moat.
Outcome-based and AI-native pricing starts eating per-seat SaaS Zendesk's Relate launch ties pricing to verified resolutions — vendor eats the risk. Anthropic's go-to-motion pushes 54% of new enterprise logos through self-serve by making Claude connective tissue across the stack. Pattern's Pi engine treats merchant ops as autonomous execution on a 77T-datapoint moat. The 2026 SaaS benchmark data confirms why this is happening: no private cohort clears Rule of 40, so vendors are repricing around outcomes to defend value.
Africa's payment rails keep consolidating top-down while the talent layer hollows out Ghana takes the e-Cedi cross-border; AfCFTA-Ecobank sign a $3bn trade-finance MoU; the 3i Africa Summit explicitly reframes the agenda from innovation to integration. Meanwhile The B&FT documents a senior-engineer scaling ratio of 5% seed-to-Series A against a 30% global benchmark, with 15–30% of senior talent leaving within 2–3 years. The infrastructure is arriving faster than the operators who can run it.
SA consumer is bifurcating — Boxer up, core supermarkets still bleeding Pick n Pay narrows its loss guidance by 10–20% on the back of Boxer outperformance, then sells another 12.5% of Boxer for R4.7bn to fund the core turnaround. Food inflation eases to a 13-month low at 2.8% but NAMC flags a 121% urea spike and 69% Brent move that will reload the cost stack within weeks. The pattern is unambiguous: discount formats and informal acceptance rails are where the volume is heading.
What to Expect
2026-05-30—COLIBRIX ONE × BitGN Agentic Ecommerce Challenge opens — first structured sandbox stress-test for autonomous payment agents in live commerce scenarios.
2026-06-01—Cape Town minibus taxis go fully cashless; SARS Traveller Management System for foreign-registered vehicles activates the same day.
2026-06-07—London Tech Week begins (7–10 June); first formal UK-Africa Ecosystem Week kicks off alongside, with structured founder support for SA, Nigerian, Kenyan, Egyptian, Ghanaian operators.
2026-06-30—Extended public comment deadline for SA's draft Capital Flow Management Regulations — the file that absorbs crypto into 1961-era exchange controls.
2026-08-31—Postbank's hard cutoff for SASSA Gold-to-Black card migration.
— The Merchant Desk
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste