AI Commerce Weekly: Week 30, 2026
Everyone spent the week building the buy-side for AI agents: a UK sandbox, fresh payment rails, an off-the-shelf assistant, a cart inside ChatGPT. Then CI&T and Morrisons showed what UK shoppers actually did with it.
TL;DR
W30 was a picks-and-shovels week. Almost nobody shipped a shopper-facing flourish, and nearly everyone laid a length of pipe underneath one. Mastercard chose the UK for Proto, a sandbox to test whether an agent can find and pay for your products before any of it goes live. Natural raised $30m to rebuild payment rails for agents, and, tellingly, to rebuild disputes. Michaels switched on Google's off-the-shelf assistant in six weeks and found 60% of the usage was still just browsing. Shoppable dropped a one-click, cross-retailer cart into ChatGPT. Then two UK stories closed the week and grounded all of it: CI&T's research says 68% of UK and Ireland shoppers have already used an AI agent, but mostly to research, not to buy; and Morrisons put Instacart's smart trolleys on a Preston shop floor. The industry laid beautiful rails for agents to buy. The UK shopper kept using agents to look.
The UK gets a sandbox for agent-readiness
The anchor story of the week was ours, which does not happen often. Mastercard has chosen the UK to launch Proto, a sandbox where retailers and payment providers can test agentic AI commerce before it goes anywhere near production. The bit worth a Head of Technology's attention is exactly what Proto lets you test: whether an AI shopping agent can discover your products at all, whether your payment journeys scale in a way a customer will trust, and whether your dispute-handling survives contact with a simulated fault. It builds on Mastercard's Agent2Agent protocol and an Agent Toolkit that uses the Model Context Protocol to expose Mastercard's own API docs to AI assistants, and it ships alongside three new agents, one for shopping, one for merchant onboarding, one for disputes. It goes live in the UK in August.
I have spent two years watching "can an agent find and buy my product" get asked as a worry. Proto is the first thing that turns it into a test you can book. Simon Forbes, Mastercard's UK and Ireland boss, leaned on the usual reasons (digitally-savvy consumers, a strong fintech scene), and the Chancellor got a quote in about Britain being the fastest AI adopter in the G7, which tells you who the announcement was really for. Keep the scepticism handy. It is pre-launch, the participating firms are named only as "in the coming weeks", and those three agents are described rather than demonstrated. This is a programme to get on the list for, not a capability you can deploy on Monday. But a card network standing up a UK-first agentic test bed, with agent identity and disputes wired in from the start, is the clearest signal yet that agent-readiness is something a British retailer can pilot rather than only fret about.
Rails for a shopper who isn't there
The venture-funded version of the same bet landed the next day. Natural raised a $30m Series A led by Forerunner, taking its total to $40m, on the argument that cards and ACH assume a human presses "pay", and an agent built to act on its own breaks that assumption. So Natural is building an orchestration layer that lets agents move and store funds, pay, collect and transact with other agents. It names Stripe as the target, has hired out of Stripe, Ramp and Square, and hedges its rails across stablecoins and bank payments.
The pitch worth clocking is not the checkout. It is the disputes. Natural is rebuilding how a disputed transaction gets handled when no human authorised it, and that is the same problem Mastercard's Proto foregrounds with its dedicated dispute agent. Two companies in one week treating "what happens when the agent buys the wrong thing" as the core problem rather than a phase-two footnote is the tell. It sounds clever right up until an agent orders forty of something at three in the morning and someone has to decide who eats the loss. The whole industry has spent months building the happy path from discovery to checkout. This was the week it started, visibly, building the sad path underneath it.
Buy the assistant off the shelf
Two more builds showed the same instinct from the retail side: stop engineering the assistant, buy it. Michaels switched on "Ask Mike", an AI shopping assistant built on Google Cloud's Gemini Enterprise for Customer Experience, and the number that matters is the turnaround: concept to production-grade AI in about six weeks. It is the packaged version of what Kingfisher and B&Q built custom on Vertex AI back in March, which quietly makes "buy versus build" a live question for anyone still scoping a bespoke assistant. Ask Mike has held nearly 75,000 conversations since a soft launch in May, and more than 60% of that usage was product discovery. Hold that 60% number. It comes back.
The checkout end got its off-the-shelf moment too. Shoppable dropped a plugin into the ChatGPT Plugin Directory that gives US users a single cart combining orders across different retailers, drawing on a catalogue it puts at 500 million products, inside a chat app it puts at 900 million weekly users. Those are the company's and the platform's own figures, so treat them as such. What is not a vendor claim is the shape: the whole thing sits on a universal-checkout MCP server, the fourth or fifth time in a month the Model Context Protocol has turned up as the standard connector under somebody's product. If your team still files MCP under research-curiosity, your vendors have already moved it onto the roadmap. Shoppable's founder, Heather Udo, put the strategy plainly.
The worked examples Shoppable leads with are telling for this desk. "What to wear to a wedding" and a "multi-brand beauty routine" are not electronics problems. A cross-retailer cart is, functionally, an outfit-assembly engine, and fashion is the one category where the multi-brand basket is the norm rather than the exception. Which only sharpens the real question, and it is not whether Shoppable's cart works. It is whether your product data is clean and structured enough for the plugin layer to recommend you at all.
What UK shoppers actually do with an agent
Here is where the week turns, and where I think the more interesting story lives. All that buy-side plumbing assumes shoppers want to delegate the buying. CI&T's new research into the UK and Ireland path to purchase says they mostly don't, or at least not yet. Sixty-eight per cent of UK and Ireland shoppers have already used an AI agent when shopping, and 86% have used one or are open to it, which is a genuinely high adoption number. But look at what they use it for.
Every top task is research. Compare, price, locate. Not one of them is "complete the purchase for me". That is the discovery-versus-delegation gap the feed has tracked all quarter, and this is the cleanest UK reading of it we have had. It sits right next to Michaels' own 60%-is-discovery telemetry from the other side of the Atlantic, two independent data points saying the same thing: give a shopper a capable agent and they reach for it to find things, not to hand over the wallet. CI&T's Melissa Minkow framed it in a way worth pinning up.
The uncomfortable read is that the industry spent this week widening the gap, not closing it. More checkout rails, more one-click carts, more payment orchestration, all aimed at a delegation behaviour the demand data does not yet support. I don't think that is a failure, exactly. It is a sequencing bet, laying the buy-side track now so it is there when trust catches up. But it is a bet, and if your customer is like CI&T's sample, the thing they will judge you on first is whether the agent found you and described you correctly, not whether it could check you out. Worth being honest that the figures are CI&T's own commissioned research rather than an ONS-grade sample. The direction, though, is corroborated everywhere you look.
The one thing that shipped to a shop floor
For all the infrastructure, the one development that actually landed in a British shop this week was a trolley. Morrisons became the first UK retailer to deploy Instacart's AI-powered Caper smart trolleys, in a Preston store, and it is Instacart's first UK retail partnership too. The carts recognise items as you drop them in, weigh loose produce on built-in scales, show a running basket total, and let you scan a Morrisons More Card for loyalty prices as you shop. One store, for now, and the accompanying shopper-satisfaction figures are Instacart's own, so hold them lightly.
What makes it more than a gadget is the lineage. This is the same Instacart that bought shelf-vision firm Arpalus a fortnight ago, and it follows Morrisons' own Simbe shelf-robot trial and Tesco's. Three capture models for the same unglamorous thing: knowing what is actually on the shelf and in the basket. The robot walks the aisle, the phone camera scans it, and now the trolley puts the sensing in the shopper's own hands. The choice a UK grocer now faces is which of those three it can run at scale, and whether the availability data feeds the order and the fulfilment model or just a store-ops dashboard nobody reads.
So the week ends where it is most honest. The clever money went into rails, carts, sandboxes and orchestration layers, all built for a shopper who will one day let an agent buy for them. The one thing a British customer could touch this week was a trolley in Preston that adds up their shopping as they go. Capability was never the constraint here. What was missing is duller and harder to buy off a shelf: a catalogue an agent can trust, a dispute you can actually settle, a store worth turning up to, and a shopper who, for now, still wants to look before she delegates. Everyone shipped the rails. Almost nobody shipped the reason to use them.
What to do this week
Sources
The UK gets a sandbox for agent-readiness
Proto is pre-launch; participating firms are unnamed and the shopping/onboarding/dispute agents are described rather than demonstrated. Live in the UK August 2026.
Rails for a shopper who isn't there
Natural's transaction-growth projection is the founder's framing; the $30m raise and $40m total are confirmed.
Buy the assistant off the shelf
The ~75,000-conversation figure is on the Retail TouchPoints page; the 60%-is-discovery figure is on the Google Cloud primary and PYMNTS, not Retail TouchPoints. Shoppable's 500m-product and 900m-weekly-user figures are the vendor's and platform's own (a vendor-submitted release).
- Michaels launches 'Ask Mike' AI-powered shopping assistant, Retail TouchPoints, 21 July 2026
- Michaels Unveils Ask Mike, Built With Google Cloud's Gemini Enterprise for CX (primary, source of the 60% figure), Google Cloud press corner, 20 July 2026
- Shoppable's Universal Checkout Plugin Is Now Available in ChatGPT, PYMNTS newswire, 22 July 2026
What UK shoppers actually do with an agent
CI&T-commissioned research (survey of 2,000+ UK&I participants); attribute to CI&T. Distinct from CI&T's December 2025 'Retail Tech Reality Check'.
The one thing that shipped to a shop floor
One Preston store for now; the accompanying Researchscape International shopper figures are Instacart-commissioned. Store name, cart count and trial duration from secondary summaries are not asserted.
Stay Connected
Follow LLCommerce on LinkedIn
Get the latest AI commerce insights, analysis, and industry news delivered to your feed.
Large Language CommerceAuthors
