Kingfisher: useful shopping assistance before autonomous purchase

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A home-improvement shopper may know the problem—an unknown broken part, an unfamiliar tool or a project that needs several products—without knowing the right product name. The useful first step is helping that shopper ask a question, identify an item and understand the available options. That is different from allowing a machine to choose a seller, accept a price and place an order.

What Kingfisher has actually deployed

Kingfisher’s earlier 2022 Hello Casto was a French Android voice interface for finding and controlling 53 fixed tutorial videos. It could start a tutorial, pause it, change volume and move through steps. This was voice and content automation, not product selection or purchasing. 1

The later, generative Hello Casto launched on Castorama France’s online shopping channel in November 2023. It answered natural-language do-it-yourself questions, provided step-by-step advice and recommended products, initially around power tools. Photo-led identification and visual search were described as forthcoming. In June 2024, Kingfisher said the assistant handled 60,000 customer conversations per month and linked shoppers to product pages. The company had only just begun analysing sales impact, so the figure demonstrates usage, not return on investment or autonomous buying. 2 3

Kingfisher’s 2025/26 annual report says Hello Casto visits rose 61%, and customers who clicked from the assistant to a product converted at more than twice the Castorama France average across first- and third-party commerce. These are useful funnel observations, but the report gives no absolute conversion rates, visit base, traffic mix, click rate, cancellations, returns or control group. Higher-intent shoppers may be more likely to use the assistant and click a product. The same report says Hello B&Q launched in Q4 FY 2025/26 as an advice and product-recommendation assistant, without reporting its volume, accuracy, conversion or transaction authority. 4

The data layer is not the buying authority

Kingfisher describes Athena as an orchestration layer that selects an appropriate microservice and model for a query. It combines recommendation systems, Vertex AI Search, enterprise data and conversational AI across text, voice, images and video. One illustrated flow uses a photo of an unknown broken part, visual catalogue search and a product match. Kingfisher also says it runs 48 Vertex AI Pipelines to orchestrate, monitor, govern and retrain its recommendation workflows. 3 5

The company’s recommendation infrastructure is commercially material: it reported that B&Q’s marketplace grew from roughly 40,000 to more than 1 million distinct products in under two years from March 2022. A Google Cloud case says recommendation conversion doubled and recommendation-engine revenue exceeded 10% of B&Q online sales, with the engine live at B&Q, Castorama, Screwfix and Brico Dépôt. The source does not provide starting and ending rates, a holdout, attribution method or incrementality. These are recommendation-system claims, not proof that Hello Casto or an agent caused additional sales. 5

What the 2026 agent announcement means

On 18 March 2026, Kingfisher and Google Cloud announced a multi-year plan to bring Vertex AI Search for Commerce to B&Q, Castorama France and Poland, and Brico Dépôt France. The announcement describes conversational discovery and future agents that could help plan projects, create shopping lists and execute seamless purchases. It says early B&Q trials produced “meaningful results,” but gives no metric, baseline, sample, date or safety result. 6 7

Kingfisher’s H1 2026/27 report is the important status check: the group was preparing for agent-enabled shopping, while B&Q natural-language search was planned for H2. No retained source confirms a completed rollout by 23 October 2026. The report’s approximately £100m of AI-driven personalisation sales, up 16%, is a management disclosure about personalisation—not a disclosed agent, search or payment result. 8

The evidence therefore supports answering DIY questions, retrieving catalogue and manual information, visual part identification, recommendations and product links. It does not establish cart creation, seller selection, substitutions, delivery commitment, checkout, payment-token use, order placement, cancellation or returns. “Execute seamless purchases” remains an announced aspiration until a controlled transaction path and consent architecture are disclosed.

An illustrative next step for a retailer

For a home-improvement retailer, a sensible application is a governed project-to-parts flow, not machine buying alone. This is illustrative, not a reported Kingfisher deployment. Ground questions and photos against canonical product identities, dimensions, compatibility, pack contents, safety constraints, seller, price, stock, delivery and returns. Let deterministic rules reject incompatible or stale matches. The customer should remain authoritative for the merchant, final price, delivery and payment.

Test the flow with real queries, ambiguous photos, near-duplicate products and unavailable stock. Measure wrong-part rate, clarification rate, product-click conversion, checkout conversion, cancellations, returns, support handoff and incremental margin using a comparable holdout. Preserve the source record, catalogue snapshot, model version and reviewer correction. A verified shortlist is meaningful progress; it is not permission to purchase.

Sources and next reading

• Hello Casto historic public entry point — historic France entry point; recorded as temporarily unavailable at the cutoff and not a verified current demo.
• Hello Casto generative-assistant launch disclosure — official product description.
• Hello B&Q launch disclosure — annual-report confirmation, not a direct assistant URL.

Explore OfferSignal · Focused Business & AI Workshop · All research

References