From storefronts to shopping agents: the rise of AI commerce

For two decades, online retail has run on the same loop: a customer lands on a page, browses a grid, filters, compares and clicks through to checkout. That loop is now being rewritten. People are increasingly starting their shopping in a conversation, asking an AI assistant for a recommendation rather than typing a search term and scrolling. At Vercel Ship London, two sessions framed this shift: Tomas Jansson on ecommerce at Currys and Elkjøp, and Jeanne Grosser, Vercel's COO, on business apps and agents. The common thread is that the storefront is no longer the only place a sale begins.

The storefront is no longer the starting line

Browse-and-click was designed for humans navigating a screen. It assumes the customer will do the work of discovery, narrowing hundreds of products down to a shortlist through filters and category pages. Conversational shopping inverts that. The customer states an intent in plain language, and an agent does the narrowing. Instead of filtering a fridge by capacity, then energy rating, then price, they say what they actually want: a fridge that fits a galley kitchen, is quiet enough for an open-plan flat, and costs under a set budget.

For a retailer like Currys or Elkjøp, with deep, technical catalogues, this is an opportunity rather than a threat. The hard part of selling a washing machine has always been matching specifications to a customer who does not speak in specifications. An agent that understands both the catalogue and the question can do that translation well, provided the catalogue is built to support it.

Agentic commerce: from recommending to transacting

The first wave of AI in retail was about recommendation, a smarter way to suggest the next product. Agentic commerce goes further. An agent does not just point at a product; it can act on the customer's behalf, assembling a basket, applying the right options, checking availability, and completing the purchase. The conversation becomes the transaction.

That changes what a commerce platform has to expose. A recommendation only needs a product to be readable. A transaction needs it to be actionable: the agent must read live price and stock, add the correct variant, and trigger a payment under clear constraints set by the customer. The systems behind the storefront have to be safe to operate without a human clicking each button, which raises the bar on data quality, permissions and auditability.

A readiness checklist for retailers

Most of the work of preparing for agentic commerce is unglamorous and happens below the interface. Before an agent can shop well on a customer's behalf, the foundations have to be in place:

• Clean, structured product data and feeds. Consistent attributes, accurate specifications, and well-formed feeds so an agent can reason about products rather than guess from marketing copy.
• Agent-ready APIs. Endpoints for search, product detail, live pricing, stock, basket, and checkout that an agent can call reliably, with predictable shapes and sensible rate limits.
• Identity and payments. A way to know who the agent is acting for and to take payment within limits the customer has set, without exposing credentials to the model.
• Trust and safety. Guardrails on what an agent may do, confirmation steps for high-value actions, and clear logging so every agent-driven decision can be reviewed after the fact.

None of this is exotic, but it demands discipline. A catalogue that is good enough for a human who can tolerate a missing field is not automatically good enough for an agent that takes those values at face value.

Being found by agents, not just search engines

There is also a discovery question. For years, retailers optimised pages so search engines would rank them. As more shopping starts inside AI assistants, the goal shifts from ranking in a results list to being surfaced in an answer. If a customer asks an assistant for the best mid-range dishwasher, you want your products to be the ones the agent can confidently describe and recommend.

Practically, the same structured data that powers transactions also powers discovery. Machine-readable specifications, honest descriptions and stable identifiers make a catalogue easy for an agent to interpret and cite. Answer optimisation is less about keywords and more about being legible to a model: the easier your products are to understand, the more likely they are to appear in an AI shopping experience.

How Inteeka helps retailers get ready

Inteeka is a UK-based product engineering studio that builds AI-native software, and this is the kind of work we do. We help retailers turn the AI conversation into production-grade systems across four stages. We structure product data so it is consistent and machine-readable. We build agent-ready commerce APIs for search, pricing, stock, basket and checkout. We build the shopping agent itself on the same modern stack as Vercel Ship, Next.js and Vercel with the AI SDK and frontier models such as Anthropic's Claude. And we measure conversion, because an agent that recommends well but does not sell is not finished.

The same foundations pay off well beyond the storefront. Vercel's own session at Ship framed business apps and agents as a way to absorb routine load, citing the potential to handle up to 90% of support tickets with agents. The structured data and clean APIs that make an agent a good salesperson also make it a good support and operations assistant.

Where this leaves retailers

Shopping agents will not replace storefronts overnight, but they are already changing where a sale begins and how it completes. The retailers who do well will be the ones whose data, APIs and trust controls are ready before customers expect to shop by conversation. The groundwork is concrete: structure the catalogue, make it actionable, keep it safe and measure what works. Get those right and your products will be found, understood and bought, whether the customer arrives through a page or a prompt.