Kingfisher, the group that owns DIY retailers B&Q and Screwfix, has launched a generative AI-powered virtual assistant in its French Castorama stores. The virtual assistant, which is built on top of a reusable platform called Athena, has been trained to answer customers’ DIY queries and provides step-by-step advice on a range of home improvement projects, as well as tailored product recommendations.
Along with conversing with customers via text chat, Kingfisher plans to enable the assistant to analyse photos to perform visual searches and answer visual queries. By uploading a photo, Kingfisher hopes to offer customers the ability to use the assistant to identify a particular part – for example, “I want to replace this broken part of my sink but I don’t know what it’s called,” or “I’d like to find another cushion like this one”.
Tom Betts, group data director at Kingfisher, joined the company three years ago, which marked the start of building capabilities around data. He says: “Over the last three years we’ve gradually been building and evolving those capabilities.” Earlier this year, the company made data part of its corporate strategy. According to Betts, this strategy recognises the importance of data in creating better experiences for customers.
Kingfisher’s data team developed the Athena proprietary AI orchestration framework, to support the virtual DIY assistant and other future applications of AI. Athena is being used to manage prompting and interaction with enterprise versions of a number of large language models (LLMs), as well as other AI tools developed in-house. It provides the necessary compliance and security guardrails that ensure proprietary and personally identifiable information remains within Kingfisher’s cloud infrastructure.
Betts describes Athena as technology-agnostic: “Apart from security and compliance, the reason that we created Athena is that we wanted a way to be able to test and learn quickly and without having to reinvent the wheel every time or think about how we deploy a particular piece of technology or test different large language models in a very quick way,” he says.
Discussing the main challenges of the project, Mohsen Ghasempour, group head of data science at Kingfisher, adds: “The biggest amount of time and energy was spent building a team, rather than the tech.” While the technology was relatively easy to implement, he says: “There were many compliance meetings, trying to understand what is good, what is acceptable and what is not right. There are a lot of unknowns.”
For a model training process, he says Kingfisher chose to build in-house technology that captured knowledge of the DIY experts that work in its stores: “Getting human knowledge close to this technology is important. A lot of time was spent making sure we surrounded these models with our internal information.”
Although Ghasempour admits the approach is by no means perfect, the initial version of the virtual assistant is being tested by the DIY experts in Castorama. “We still have to improve aspects of accuracy,” he adds.
The aim of the initial project is to replicate online some of the expertise that occurs on the shop floor in a DIY store, where a customer is able to get advice from knowledgeable staff.
A ranking process is used to retrieve the most relevant information held in internal documents containing DIY tips, as Ghasempour explains: “The whole idea is grounded in a large language model. So rather than just coming with the general answer to the question of how you paint your bathroom, we can say, ‘This is how all our experts think you have to paint your bathroom.’ So you’ll get the general information, but also the rank mechanism is used to make the online shopping experience a bit closer to shopping in-store.”
Among the risks associated with LLMs is that they can generate nonsensical information, often referred to as “hallucinations”, that are inaccurate. To reduce the risk of this occurring, Ghasempour says the virtual assistant uses a consensus checking layer, where information is drawn from multiple LLM providers that operate using completely different data models: “We use a foundation model and a task-specific model to answer the specific question and we try to have another large language model that simply checks for contradictory information.” While this does not guarantee that hallucinations will not occur, he adds: “The chance of five or six different large language models providing wrong information is significantly reduced.”
The DIY virtual assistant project builds on a range of initiatives including an AI-powered product recommendation and personalisation engine at B&Q and Screwfix, which are already generating up to 10% of e-commerce sales, and AI-driven tools to optimise markdowns and clearance.