Dunelm's 'Ask Dunelm' Is the UK's Agentic Retail Moment - But Can It Survive Contact With a Real Customer?

There's a version of the future where your customer doesn't type "grey curtains 90 drop eyelet" into a search bar anymore. They just say what they want, in the words they actually use, and the agent finds it. That future landed in the Dunelm app on iOS on 6 August.

Dunelm, the UK's leading homewares retailer, has launched 'Ask Dunelm': an AI-powered shopping assistant designed to help customers discover products through natural conversation, enabling them to describe what they're looking for in their own words, ask questions about products, visualise items in their homes, and receive tailored recommendations. No Boolean search strings. No category rabbit holes. Just talk.

I've been watching retailers try this for fifteen years. Faceted search. Guided selling wizards. "Virtual stylists" that were really just decision trees wearing a chatbot mask. Most of them were expensive failures dressed up in press releases. So I was cynical. Then I looked at what's actually powering it.

The tech stack matters more than the demo

Dunelm is one of the first retailers in the UK to launch the new generation multi-modal shopping agent utilising Google Cloud's Gemini Enterprise for Customer Experience product suite. That's not a trivial claim. Gemini Enterprise for CX is Google's serious commercial offering, not a hobbyist API wrapper. The implementation partner is Valtech, who worked alongside Google Cloud and Dunelm to upgrade the retailer's search engine to AI Commerce Search on Gemini Enterprise, integrating with Dunelm's engineering, search and analytics teams.

This isn't a startup bolting a thin UI onto a generic model. It's a considered stack build with a named enterprise partner, on a named enterprise product tier. That changes the product-market fit conversation considerably.

Customers can describe what they are looking for in their own words and receive suggestions tailored to their style, budget and requirements. The tool can also help shoppers plan an entire room, find a specific item or explore homeware inspiration without relying on traditional keyword searches. Critically: Ask Dunelm is already live on the customer app, which the company introduced earlier this year and which has already been downloaded by more than 500,000 customers. The distribution is real. It's not a microsite waiting for traffic.

The Ask Dunelm launch follows Dunelm reporting a solid finish to FY26, with total-year sales rising 3.1% to £1.83 billion. That's the commercial context. They're not burning venture money on a vibe. They're a profitable retailer investing from a position of strength.

The payment layer problem nobody's talking about

Here's the bit I find genuinely fascinating, and it connects Dunelm's launch to something bigger happening right now in UK fintech.

An agentic shopping experience only works end-to-end if the payment layer is also agent-ready. And right now, it isn't. Not really. The agent can find your grey eyelet curtains. But when it tries to check out, it hits the same friction-filled card form that's been annoying people since 1998.

That's exactly why Mastercard's timing is interesting. Mastercard selected the United Kingdom as the launch market for Proto, a new sandbox environment designed to let banks and retailers test agentic AI use cases before deploying them in live operations, as part of a broader set of AI commitments aimed at helping UK businesses prepare for a shift toward software agents that autonomously search, decide, and pay on customers' behalf.

Proto will go live in the UK in August and will allow businesses to test and scale new forms of AI-enabled commerce. That's this month. Simultaneously with the Dunelm launch. Probably not coordinated, but the timing tells you where the industry is.

Mastercard also announced three new agents designed to support key enterprise functions and make retail AI-ready: a shopping agent, an onboarding agent and a dispute agent. The dispute agent is the one I'd watch. In agentic commerce, when the agent buys the wrong thing or the wrong size, the chargeback question gets complicated fast. Who authorised the transaction? The consumer? The agent? The model? I sat in enough identity verification and KYC meetings in a previous life to know that "the AI did it" is not a defence that survives contact with a financial regulator.

Forbes said: "Before agents are trusted to act on behalf of consumers and businesses and be deployed at scale, retailers and their partners need the tools and environments where they can evaluate performance, identify failure points and refine controls." That's the honest positioning. Proto isn't a product. It's a professional qualification for your stack.

The trademark signal hiding in the noise

There's a quieter data point that's been nagging at me. According to AIBD analysis of IPO trademark data, Class 42 UK filings - the class that covers software, SaaS, IT services, and now explicitly AI-as-a-Service following the 13th Edition of the Nice Classification entering force in January - hit just 3,912 in Q3 2026, down 47.2% on the prior period.

That's a significant contraction. Artificial intelligence as a service is now a recognised service in Class 42 under the new classification, so you'd expect filings to increase as founders race to protect their AI product brands. Instead they're collapsing.

A few possibilities. One: founders are skipping formal IP protection altogether, gambling on speed-to-market over defensibility. Two: the consolidation dynamic in UK AI funding is filtering down. New findings show that capital is returning at scale, although the recovery is being shaped by fewer deals and a small group of unusually large rounds. Fewer companies forming means fewer trademarks filed. Three: the early AI branding goldrush of 2023-24 simply burned itself out. Founders who registered every combination of "AI" and a noun have stopped.

My read: it's mostly option two. AI-related companies attracted 74% of all venture capital deployed in the UK in H1 2026, but the structure of that investment is megarounds going to proven names. The long tail of early-stage AI SaaS founders who'd normally be filing Class 42 trademarks as one of their first acts? They're not getting funded. So they're not filing.

When you've been around long enough to remember the post-dotcom trademark graveyards of 2001-2003, this pattern looks familiar. Concentration at the top, silence everywhere else.

Who is Ask Dunelm actually for?

Back to the product. The product-market fit question that matters: is this for Dunelm's existing app users, or does it pull in new customers?

The launch follows the introduction of Dunelm's customer app earlier this year, which has now been downloaded by more than 500,000 shoppers. The app includes personalised product recommendations and Click and Collect integration. Dunelm said Ask Dunelm was the first of several new enhancements planned for the platform. The positioning is clear: this deepens the existing app relationship. Retention and basket-size, not acquisition.

That's the right call for a retailer at this stage. I've made the mistake of building customer-facing AI features as acquisition tools before. They're not. People don't download an app because it has a chatbot. They use a chatbot because they already trust the brand. Ask Dunelm works because Dunelm has 50-odd years of homewares credibility and 500,000 people already in the app. A startup launching the identical feature tomorrow would get 40 users and a write-up on Product Hunt.

Positioning: sensible. Timing: well-judged. Tech stack: enterprise-grade, not prototype. The onboarding experience was, from what I can gather, deliberately understated. No splashy launch event. Just there, in the app, working.

I built a version of this in 2008 for a furniture client. Natural language product search, powered by a decision tree the size of a small forest, costing approximately what a Series A would cost today. It worked about 30% of the time. The other 70%, it returned a beanbag.

August 2026 is different. The models are real now.

Founder's verdict

Ask Dunelm is the most credible agentic retail deployment to land in the UK this year: grounded in a real distribution base, a mature tech partnership, and commercial results that justify the investment. The Mastercard Proto sandbox arriving simultaneously shows the payments infrastructure will eventually catch up with the discovery layer.

The falling Class 42 trademark numbers are the dark matter of all this excitement. The funding is concentrating, the ecosystem is thinning, and the founders who were going to build the agentic middleware layer are quietly running out of runway. Watch that gap. The enterprise players filling it with proprietary sandboxes will own a lot of what comes next.

And if you're a UK SaaS founder who hasn't checked whether your AI product brand is protected under the new Nice Class 42 AIaaS definition? That's homework.