As brands look for AI edge, B&Q retail owner Kingfisher is expanding in-house development

It’s an example of a brand launching an AI feature without the assistance of the agency world, which has been hard at work to prove its relevancy in this area. WPP and Publicis Groupe, for example, kicked off this year touting their own expertise with generative AI. In particular, they took pains to stress the idea that brand marketers would need to rely on their developers, their engineers and their links to companies such as OpenAI and Anthropic, to make sense of this brave new world. But some marketers are finding that agency partners aren’t necessarily necessary.

78% of in-house agencies are using generative AI for image, text and video generation, a recent survey of 49 British and European businesses conducted by the In-House Agency Leaders Club, an industry group for U.K. in-house marketers.

Kingfisher isn’t allergic to hiring agencies or backing high-concept ads; it works with Uncommon, Valtech and DAC Group, among others. It wasn’t long ago that the firm put out a surreal TV spot in partnership with the former for B&Q, (which carries the “DIY.com” trophy domain) soundtracked by David Bowie’s “Sound and Vision” for B&Q.

But Kingfisher developed its first on-site AI chatbot internally because it wanted to emphasize flexibility, transparency — and lower costs, said Tom Betts, Kingfisher’s chief data officer. “It’s very difficult to plan [for] what the world is going to look like in 12, 18, 24 months’ time in terms of gen AI tools,” he told Digiday.

The firm’s chatbot assistant was released in November and was built using Athena cloud software that allows users to draw from several large language models (LLMs). Mohsen Ghasempour, group AI director at Kingfisher, said it gave the company access to the best features of different models when building the bot, which is called Hello Casto (it shares its name, if not any underlying tech, with an Alexa voice assistant also released for Castorama by Kingfisher back in 2022).

“Some of them are good at reasoning, some of them are good at summarization, some of them are good at providing or generating content. If you stick with one, you’ll only have those limited capabilities,” he explained. “What our team does is try to use the best of each language model.”

Betts said that the approach meant Kingfisher could keep its options open. “We’re not pinned into one particular technology choice,” he said, adding it meant its team retained “the opportunity to test and learn which technologies will go to solve a particular problem.”

There’s also cost advantages. Building a bespoke LLM from scratch — like the custom “Brains” WPP is offering clients as part of its WPP Open platform — is a pricey endeavor. “Building a large language model from scratch, you need a lot of computing resource. It’s quite expensive,” said Ghasempour.

But pricing plans for Athena begin at $99 a month. Ghasempour didn’t share the exact amount spent on the project, and the team hadn’t estimated how much an agency might have charged, but said it took a small internal team with just a “couple” of staffers two weeks to build a proof-of-concept widget using Athena, with a further two weeks to make it ready to debut on the Castorama website.

According to a Kingfisher spokesperson, Hello Casto currently handles 60,000 customer conversations each month. Though Betts and Ghasempour said Kingfisher has only just begun analyzing how the bot might have driven sales, they’re already beginning to plot out similar applications for Kingfisher’s other retail brands, which include Brico Depôt in France, Romania and Spain, Turkish business Koçtaş, and Britain’s B&Q.

The bot itself is a similar application to others deployed by marketers in recent months; Klarna’s AI assistant, for example, processes customer queries on behalf of the buy-now-pay-later giant, while Ikea released a custom GPT. Unlike the Swedish flat pack giant’s effort, which only directed users toward blog posts and content, Hello Casto does send users to a product page to make a purchase.

Debbie Ellison, VML Commerce’s chief digital officer, told Digiday that many brands were using chatbots as an opportunity to “dip their toes” into generative AI development, because upfront costs are lower than other integrations. AI chatbots can offer more naturalistic conversation, and process a larger variety of conversations, than older chatbot assistants.

She said that when chatbots are used to mimic in-store customer experiences (such as asking a shop assistant for advice), they can provide “huge, huge value for brands and retailers — when done correctly.”

The WPP agency is primarily exploring AI deployment in aid of higher fidelity social listening techniques, more efficient content production or trend prediction, she added. Chatbot experiences were a “first foot” for brands experimenting with AI, but would likely not deliver personalized marketing experiences alone, she said. “There’s loads of opportunities to deliver that heightened and connected customer experience … I don’t think we’re seeing it in its full glory yet, if I’m honest,” she said.

The In-House Leader’s Club survey found that concerns over legal risk and brand safety were high among in-house marketers. Betts said Kingfisher’s choice to test an assistant in the market was itself a means of avoiding risk.

“We spent a lot of time in the second half of last year working on group-wide policies for AI with a cross functional group of experts, [but] the reality is, whatever you put on paper today is probably not going to be fit for purpose in 18 months’ time,” he said. “The best way of learning about these technologies is to have a go at applying them.”

Without that experience, Betts said the risk of being led astray by a third party was higher. “You may just buy something which doesn’t work, or is snake oil, if you don’t really understand it,” he said.