Boots collaborated with TCS and Microsoft to launch the first customer-facing generative AI chatbot in UK retail. The chatbot can answer customer service queries, find products to solve problems, and provide basic healthcare advice using curated Boots data sources. The chatbot was deployed in 16 weeks, approved by the HMRA, and currently serves 50% of web traffic. The chatbot also uses LLMs for analytics, PII masking, and hallucination catching, and has inspired more AI innovation across the business.
Customer satisfaction is key
Boots operates 2,000 stores and a strong online presence, but customers face challenges including:
- Accessing trusted healthcare advice on demand.
- Finding products to solve problems rather than just products they know about.
With the dawn of generative AI (genAI) the Boots data team saw an opportunity to solve these problems and tackle business challenges:
- Boots is a large and complex business but needs to innovate at pace.
- Understanding what customers really need from the website beyond shopping.
- Understanding costs associated with the implementation of running a genAI product.
GenAI provided an ideal opportunity for Boots to tackle these challenges. Usually, a project like this would take a year or more, but with Christmas trading 16 weeks away the data science team was tasked with getting a proof-of-concept chatbot in front of customers by this short deadline.
The team collaborated with TCS and Microsoft and corralled support from senior leaders in the business to free themselves from a lot of the red tape that hinders innovation and released the chatbot to real world customers on at the end of November.
The chatbot can:
- Answer customer service queries with access to over 500 pages of customer help and FAQs.
- Converse with customers and find products to help solve their problems from over 10,000 products.
- Answer basic healthcare questions using a curated collection of documents.
To control the chatbot’s responses, the team used a RAG architecture grounding all responses in these Boots datasets. The Boots chatbot stands out as:
- It was the first customer facing chatbot in UK retail.
- It handles the high stakes area of healthcare.
- It was successfully launched and ran through the UK’s busiest trading period.
LLMs are used for more than conversations. The team at Boots has integrated LLMs into every step of the chat process using it to synthesise questions for hybrid search and hallucination catching; but its analytics feature is the most impressive.
All chatbot queries are collected, embedded using LLMs, and clustered for analytics. The team at Boots then use the LLM to label and tag features in the clusters. This analytics is used for insight, alerting, and to inform the development pipeline of the chatbot and elsewhere in the customer experience. The driving force is the prioritisation of building what customer conversations state they want.
Safety was a key concern when developing this product, for customers and the brand. Boots has a 175-year legacy working with people’s health and there have been numerous examples of business being reputationally damaged with improperly tested genAI applications. Most applications have comparatively low stakes if something goes wrong, but this is not the case with Boots as the team are offering healthcare advice.
There were multiple stages of safety in the development process. The team first gathered key test questions from the business to assess the development on. After this, an internal UAT was arranged with over 1,000 Boots users able to test the chatbot. Their instructions were to use the chatbot realistically and to attempt to derail it. Through this process the data team adjusted the prompts and added additional checks to the chatbot routing. The LLM-based analytics flags any inappropriate or harmful queries and responses so they can be further investigated.
To provide healthcare advice, Boots got the chatbot approved by the Medicines and Healthcare products Regulatory Agency (MHRA).
Privacy was of utmost importance. Chat histories are useful but with an open text input users can enter anything, including PII; 1/1,000 exchanges contain PII data. The team attempted to use off the shelf PII masking solutions but found them too sensitive or too lax. The solution was to use the LLM for PII masking of data which was far more accurate and discerning of difficult subjects like a brand name from a customer name.
The Boots data science team has surpassed all expectations and are regarded as a hub of innovation. The chatbot was designed to serve 1% of web traffic for four weeks – it is currently serving 50% of traffic. In just over six months Boots has had over 100,000 exchanges with customers helping them find products, providing health advice, and answering over 30,000 queries.
Finally, the Boots team has further used this project to inspire ideas for the use of AI around the business. Its final deliverable was to highlight what can be done when there is an appetite for innovation. The data team now host regular sessions across the business to explain what can be done with this technology as well as other AI technologies and how these ideas can become reality.
Boots are part of the DataIQ membership programme – the trusted global collaboration and intelligence platform for data leaders. Find out more here: https://www.dataiq.global/membership
