look
← Insights

Applied AI · 8 min read

Agentic commerce in Europe: why checkout is the wrong place to start

Walmart pulled out of AI checkout after it converted three times worse than their own site. And the EU deadline European retailers are bracing for just moved 16 months.

In November 2025, Walmart put roughly 200,000 products into OpenAI's Instant Checkout. By March 2026, they had pulled out.

The stated reason was not compliance and not a technical failure: in-chat checkout converted about three times worse than sending the same shopper to walmart.com. Walmart replaced it with Sparky, their own assistant, running inside ChatGPT. Shoppers authenticate with a Walmart account, carts sync, and the purchase completes on Walmart's own secure infrastructure.

That is the best-resourced retailer in the world, with direct platform support, finding that taking checkout out of their own domain made conversion three times worse. It is worth sitting with before you scope an agentic checkout project.

The regulatory timeline moved (and why it creates an opportunity)

Many European retailers are bracing for an EU AI Act high-risk compliance deadline in August 2026. But the regulatory timeline has shifted: the Digital Omnibus on AI (approved by the European Parliament on 16 June 2026 and Council on 29 June 2026) pushed Annex III high-risk standalone obligations to 2 December 2027, and Annex I embedded systems to 2028.

What remains live from 2 August 2026 is narrower: Article 50 transparency and AI-content labelling rules, GPAI provider obligations, and Article 5 prohibited practices.

This changes the thesis for European teams. You have 16 more months of regulatory runway than consensus suggests. The reason to defer autonomous checkout is not because Brussels forbids it; it is because in-chat checkout does not convert yet and liability remains unallocated.

RegulationStatus in 2026Practical Impact for Commerce
EU AI ActHigh-risk deferred to Dec 2027Only Article 50 transparency and AI labelling apply in 2026. More runway to build.
GDPRActiveConversational memory blurs session context vs retained user profiles. Sizing preferences stored in chat require clear deletion paths.
PSD2Active (PSD3 expected 2027/28)Strong Customer Authentication (SCA) assumes active human presence at payment (e.g., banking app 2FA).
Does the agent move money?Is liability contractually resolved?Build it nowDefer the transaction layerdiscovery · guided selling · order supportAI Act · GDPR · PSD2low exposure, measurable ROIkeep checkout on your own stackyesnono
The decision most European teams should be making in 2026.

The Nordic factor: parcel lockers over doorsteps

Building agentic discovery in the Nordics requires accounting for local delivery habits that global AI demos ignore. According to PostNord's Spring 2026 report, parcel lockers account for close to half of all Finnish e-commerce deliveries, making Finland the highest locker-preference market in the Nordics (with the fastest growth among shoppers aged 65 to 79).

If your AI agent recommends products and prepares a basket, it must also query live locker location capacity and pickup preferences. An agent that defaults to doorstep delivery in Finland is an agent that fails at checkout handoff.

What is actually buildable today

The winning pattern keeps discovery, qualification, and support with the agent, while checkout remains on your proven infrastructure.

LayerWho owns itBuild in 2026?
Discovery & comparisonAgent (yours or platform's)Yes. 56% of European shoppers have tried AI-assisted shopping at least once.
Catalogue, inventory, pricingYou (exposed as tools)Yes. Essential substrate for all AI discovery.
Guided selling & supportYour agent (read-heavy)Yes. High value, low compliance risk.
Autonomous checkoutUnresolved liabilityNo. Defer transaction layer, keep on merchant site.

Where the protocols actually sit

Understanding the protocol stack prevents wasted engineering effort across MCP, UCP, and ACP:

MCP (Model Context Protocol) is the open transport substrate. Anthropic released it in late 2024, and by May 2026 its official registry listed roughly 9,650 servers. It allows models to safely query catalogue, stock, and order status as typed tools. Building an MCP server for your inventory is a durable investment regardless of which chat platform dominates.

UCP (Universal Commerce Protocol) launched at NRF on 11 January 2026, co-developed by Google with Shopify, Etsy, Wayfair, Target and Walmart, to standardise cart and discovery schemas across platforms. ACP (Agentic Commerce Protocol) from OpenAI and Stripe is coupled to ChatGPT's surface and, since the Instant Checkout retreat, functions mainly as a discovery layer that hands the transaction back to the merchant.

Agent surfacesACP · OpenAI + StripeUCP · Google + ShopifyMCP · the transport you controlChatGPT · Gemini · your own assistantcheckout, coupled to one surfacefull lifecycle, protocol-agnosticcatalogue · inventory · orders · support
MCP is the substrate. ACP and UCP are distribution channels above it.

A 90-day execution plan

  1. Weeks 1 to 3: Make data legible. Expose product catalogues, stock, and fulfillment options (including locker locations) via structured MCP tools.
  2. Weeks 4 to 8: Ship one read-heavy agent. Launch order tracking, returns support, and guided product discovery with clear human fallback.
  3. Weeks 9 to 12: Instrument conversion. Track containment rate and assisted conversion lifts before deciding on transactional experiments.

The short version

Agentic commerce in Europe right now is a discovery and support problem wearing a checkout costume. The teams that win are those building catalogue and tool infrastructure this year, focusing on read-heavy use cases with high ROI and low regulatory exposure.

If you are running commerce in the Nordics and want a second opinion on structuring your catalogue for agentic discovery, reach out to start a conversation.

Sources