AI Commerce Weekly: Week 27, 2026
Two UK studies say shoppers won't hand AI the wallet. The same week, Visa proved an agent could pay and be safely accepted at both ends of the transaction.
TL;DR
W27 was the week the rails got built at both ends of the transaction while the shopper stayed unconvinced. Two fresh UK studies, Commerce with PayPal and ACI with YouGov, found Britons keen on AI help and wary of AI autonomy, and ACI's sharpest number was a one-strike rule: sixty per cent would drop an agent after a single mistake. The supply side laid track regardless. Visa proved an agent could pay (with BBVA) and be safely accepted (by eDreams) in the same seven days. Salesforce shipped Agentforce Commerce and argued the sale stays on owned ground. New Look took AI upstream into the design studio. Capability was never the constraint this week. Trust was.
The demand side won't hand over the wallet
The week opened with the demand side restating its terms, and it did so twice, from two different vendors and two different survey houses, inside a few days of each other. Commerce and PayPal published research putting 64% of UK consumers as interested in trying AI agentic shopping tools. Read past the headline and the shape is familiar: only 21% use AI for shopping today, 70% would like to in future, and what they want it for is mostly price, finding the best one (31%), surfacing every promotion (28%), flagging when something is cheaper elsewhere (23%). The hesitation is where the wallet is. Forty-three per cent worry about AI completing a purchase without approval, 39% about bank-account security, and 83% say any AI tool has to offer payment security at least as good as what they use now. On transparency the line is firm: 55% want sponsored content clearly labelled and 46% think brands should not be able to pay for preferential treatment inside an AI shopping tool.
Then ACI Worldwide, with YouGov, put the hardest number of the week on the table. Across 2,080 UK adults, just 19% trust AI assistants to make everyday purchasing decisions, against 55% who trust a human expert. Sixty per cent would stop using an agent after a single mistake. That last figure is the one worth sitting with, because it changes what the engineering problem actually is. This is not a soft preference that improves with familiarity. It is a one-strike threshold, and it makes reliability, not model capability, the thing that decides adoption. The trust drops off a cliff the moment money and accountability enter the frame.
Sarah has made the point before that the industry keeps mistaking consumer enthusiasm for consumer trust, and this is the cleanest evidence of the gap in a while. The appetite for help is real. The willingness to delegate is not, and it collapses precisely at the point of payment and problem-handling. ACI's Adriana Iordan reads it as an instruction to prioritise control over capability: explicit approvals, hard spending limits, protected payment details, and clear accountability when something goes wrong. Notice who consumers hold responsible when it does. Fifty-four per cent say the company that built the agent should be liable for refunds, 9% blame the retailer, and only 3% blame the bank. If you are building or buying an agent, that liability question is now yours whether you asked for it or not.
The third study points the finger inward. HyperFinity found only 29% of retailers say their loyalty systems are fully ready for agentic AI, with 72% admitting they are not. Loyalty and retention is usually the easiest place to prove personalisation ROI, so if that is where the readiness gap sits, the problem is rarely the model. It is the connected, clean customer data underneath it, which is the same drag Quickfire's "Growth Tax" research surfaced the week before. Agentic loyalty is only ever as good as the data it gets to reason over.
Both ends of an agentic payment went live
Here is what makes W27 worth remembering. Two of the constraints the demand-side research keeps naming, payment security and accountability, got a working infrastructure answer in the same week, and it landed at both ends of the transaction.
On the issuer side, BBVA completed a live AI agent-initiated card transaction with Visa. Not a sandbox demo: real card credentials, an active merchant's live systems, an agent making the purchase on a cardholder's behalf. It ran on Visa Intelligent Commerce with tokenisation and real-time fraud monitoring, and, the load-bearing detail, Visa Payment Passkeys satisfied the EU's Strong Customer Authentication requirement. That is the whole point. It shows an agent can pay inside the regulation that already governs online payments, with cardholder consent and issuer oversight intact, on rails that are in production today rather than a roadmap slide. The same Agentic Ready programme and the same SCA plumbing apply in the UK.
Two days earlier, the other end. eDreams ODIGEO switched on Visa's Trusted Agent Protocol and Agentic Directory so its storefronts can recognise a verified agent and tell it apart from unknown or unverified traffic before letting it transact. Where BBVA proved an agent can pay, eDreams proved a merchant can safely accept one. This is the control most retailers have not thought about yet, and it is the one that matters most for apparel, because a wrong size burns trust faster than a mis-booked flight. Agent verification is a merchant problem, not just an issuer one. eDreams runs UK-facing brands in Opodo and Travellink, and the Trusted Agent Protocol, Agentic Directory and Payment Passkey stack it deployed is available to any UK merchant that accepts Visa. The question to add to your roadmap now is simple to state and awkward to answer: how would our storefront recognise a trusted agent, and reject a scraper wearing one's clothes?
Worth being honest about where this leaves us. The plumbing is demonstrably built at both ends. Whether shoppers trust it enough to use it is the ACI data's problem, not the network's, and that gap is going to take longer to close than the engineering did.
Salesforce bets the sale stays on home ground
If the payments networks answered the security question, Salesforce spent the week answering the strategic one. Agentforce Commerce is now generally available, with three agents live: a Shopper Agent that checks live inventory, confirms carrier cut-offs, offers store pickup and closes the sale inside one conversation; a B2B Buyer Agent that handles procurement over WhatsApp and SMS; and a Merchant Agent for the back office. Native ChatGPT integration is live now, with Google Search and the Gemini app following in the coming months. There is also an intent-aware search layer, built on technology Salesforce acquired from Cimulate, that will run on non-Salesforce storefronts.
The interesting part is the argument underneath the launch, and it is a direct answer to the quarter's trust data. Salesforce's Nitin Mangtani concedes discovery will scatter across the AI platforms and the social apps, then plants a flag on where the transaction should stay.
Keep the agent on your own surfaces, keep the wallet and the first-party data in-house, remain merchant of record, and get it live before peak. It is a coherent position, and the presence of Iceland Foods as a named reference customer (its Head of Ecommerce Technology, Luke Barber, endorsing the underlying platform) grounds an otherwise US-led vendor launch in a British grocery aisle. Whether "predominantly on owned properties" survives contact with a generation that starts every search in ChatGPT is the open question. But as a hedge against being disintermediated by the discovery layer, owning the transaction is a more defensible place to stand than owning the discovery.
The fashion win is upstream, in the design studio
The clearest UK-fashion story of the week was not a shopping agent at all, and it was not anywhere the shopper could see it. New Look put Fermat's generative AI into its buying and design teams, using it to create virtual product renders, test multiple design iterations and explore styling options without the traditional sampling round. Creative Director Anica Wislawski framed it as taking time out of the mechanical parts of design and reinvesting it in the human parts, understanding the customer and turning trends into wearable, everyday clothes.
This is the second UK fashion name to do this in five months. ASOS embedded Fermat across its design function back in February, upskilling more than 100 designers and reporting an average 75 to 80% time saving across key design processes. Treat that number as vendor-reported rather than audited, but the direction is hard to argue with, and two names in a quarter makes AI-in-the-design-studio a pattern rather than a one-off. My career started in design before it ever touched a database, so I have a soft spot for this being where the returns show up first. If your entire AI roadmap points at the product page and the checkout, you may be walking past the higher-certainty win sitting upstream in product development, where the customer never sees the model and never has to trust it. Sampling cost and trend-to-market speed are real numbers on a fashion P&L, and they move without asking a shopper to hand an agent their card.
The plumbing the shopper never sees
A cluster of supply-side launches this week shared a theme: wiring AI into places the shopper never looks. Square introduced a ChatGPT app and a Claude plugin for its sellers, plus an Alexa+ tie-up with Amazon, so a merchant can be discovered and transact inside an AI conversation without building the integration themselves. It goes live first with US food-and-beverage sellers, and, notably, Square is not charging extra marketplace commission on those orders. Block's Morgan Kuntze called managing each new AI channel by hand an "impossible game of catch-up", which is an honest description of the problem every merchant is about to have.
Two back-office launches rounded it out. Revuze shipped autonomous agents and a Model Context Protocol layer so brands can pipe consumer-intelligence data straight into their own AI systems, and Fusemachines expanded an AI collaboration with a global luxury-handbags retailer for demand planning, pricing and merchandising, with agentic automation named as the next phase. Both are US-led, B2B, and light on independent metrics, so file them as watch-items. But the common thread is worth naming. MCP turned up three times this week, in Square's app, in Revuze's integration layer, and in the 100-plus MCP integrations eDreams says it has wired into its booking engine. It is quietly becoming the default connective tissue for plugging commerce systems into external AI, which is the sort of standard you want to understand before it is a fait accompli in your own stack.
GEO gets priced as a representation problem
If agents are going to do the discovering, the question every brand inherits is whether the agent can find you and describe you correctly. London's geoSurge raised an oversubscribed 12 million dollar seed, led by AlbionVC with angels from Google DeepMind and Microsoft AI, on a sharper version of that question than most of the GEO market is asking. Corroborated by UKTN, the pitch is that tracking which pages an AI cites in a live answer is the wrong game, because the durable battleground is the representation layer, how a model learns, remembers and describes your brand over time.
That one line is the argument. If a model increasingly answers from what it has internalised rather than what it retrieves in the moment, then getting cited on a good day matters less than being accurately represented all the time. geoSurge has no named retail or fashion customer yet (its clients are in finance, education and hospitality), and "Corpus Engineering" is its own coinage pending outside validation, so I would not rebuild a roadmap around the label. But the underlying point is the one this publication has been circling all year, most recently through Nudge's raise the week before: machine-readability is becoming a distribution channel. Whether or not you buy geoSurge's tool, the audit it implies is free. Ask ChatGPT, Gemini and Claude what your brand sells, what it costs, and who it is for, and see whether you recognise the answer. That is what an agent reads before it decides.
Sources
The demand side won't hand over the wallet
All three are vendor- or agency-commissioned surveys; attribute figures to Commerce/PayPal, ACI Worldwide and HyperFinity respectively. The ACI figures are inline-linked to the RTIH write-up for a reliably reachable source; the Business Wire release is the primary.
- More than half of consumers interested in using agentic AI tools for shopping (Commerce/PayPal research), FashionUnited UK, 29 June 2026
- Many consumers would bin AI shopping agents after just one mistake, ACI Worldwide research finds (YouGov, 2,080 UK adults), RTIH, 29 June 2026
- Six in Ten UK Consumers Would Stop Using an AI Shopping Agent After One Mistake, ACI Survey Finds (release primary), Business Wire, 28 June 2026
- Many retailers admit their loyalty systems are not ready for the AI era, HyperFinity research finds (200 senior leaders), RTIH, 29 June 2026
Both ends of an agentic payment went live
- BBVA completes its first AI agent-initiated transaction together with Visa (issuer side), BBVA newsroom, 2 July 2026
- Visa and BBVA prove today's rails can handle agentic payments, PYMNTS, 2 July 2026
- eDO taps Visa to allow AI agents to purchase travel (Trusted Agent Protocol + Agentic Directory, merchant-acceptance side), PYMNTS, 3 July 2026
- eDreams ODIGEO and Visa power agentic commerce in travel with new secure AI agent protocols (primary), eDreams ODIGEO newsroom, 3 July 2026
Salesforce bets the sale stays on home ground
The fashion win is upstream, in the design studio
ASOS's 75-80% design-time saving is vendor-reported; treat as directional. The Retail Gazette write-up corroborates both the New Look partnership and the ASOS precedent.
- UK fashion retailer New Look announces partnership with AI-powered visualisation platform Fermat, RTIH, 30 June 2026
- ASOS partners with Fermat to upskill all designers in generative AI (February precedent, 75-80% time saving), ASOS plc, 3 February 2026
- New Look turns to AI to speed up product design (corroboration), Retail Gazette, 30 June 2026
The plumbing the shopper never sees
Revuze and Fusemachines are vendor releases with no independent metrics; the Fusemachines retailer is anonymised. Watch-items, not headline claims.
- Square Intros ChatGPT and Claude Integrations for Sellers (plus Alexa+), PYMNTS, 1 July 2026
- Revuze Launches AI Agents and Model Context Protocol to Power Next-Gen AI for CPG and Retail, PR Newswire, 24 June 2026
- Global Retailer in Luxury Handbags Expands Collaboration with Fusemachines for Demand Planning and Pricing Prediction, GlobeNewswire, 1 July 2026
GEO gets priced as a representation problem
"Corpus Engineering" is geoSurge's own terminology, pending third-party validation. No named retail/fashion customer disclosed yet.
Stay Connected
Follow LLCommerce on LinkedIn
Get the latest AI commerce insights, analysis, and industry news delivered to your feed.
Large Language CommerceAuthors
