Responding faster to customers with GenAI

Jaja Finance (Jaja) is a leading digital lender in the UK, focused on providing simple and fair credit through its own brand credit card, Jaja Vanta, and its Asda Money Credit Cards. By focusing on simplicity, functionality, service, and security, Jaja has built a strong and rapidly growing customer base. With this growth, the company recognised a need to boost its customer service offering while maintaining its commitment to helping customers use credit better.

Jaja partnered with Slalom to explore the potential of using generative AI (GenAI) to automate up to 30% of routine customer requests. The goal was to reduce the need for manual intervention by Jaja’s customer service team on more general enquiries while freeing up customer service colleagues to handle more complex enquiries, enhancing the overall customer experience.

The project began with a rapid feasibility assessment of automating knowledge-based responses to customer queries using large language models (LLMs) and semantic searching. From there, the team moved quickly to productionising the solution, which went live to a select customer segment in March 2024, with a full launch in May 2024. 

Jaja became one of the first UK fintechs to launch a GenAI chat assistant using Anthropic’s state-of-the-art Claude 3 model family, which it has now upgraded to Anthropic’s more advanced language model, Claude 3.5. Known as Airi, the GenAI chat assistant focuses on complete accuracy in customer chat responses and understanding the reasons why conversations end to determine the next steps required in business processes.

“As a digital lender we’re already employing sophisticated data models and advanced technologies to ensure we’re giving our customers the tools they need to manage their credit well,” says Francesco Di Costanzo, interim CEO and CFO of Jaja Finance. “Adding an intelligent generative AI assistant like Airi helps us unlock the power of advanced AI for our customers, while ensuring their financial health and well-being.”

Airi is enhancing the customer experience by providing a 24/7 service, offering increased intelligence with near-human levels of responsiveness and improved data accuracy. Jaja customers can use Airi to access a broad range of self-serve options with 24/7 availability, whether it is managing general statement enquiries or discussing payment options.  Airi is also allowing customers to reply to conversations at their own convenience, while providing them with a full transcript of their conversation.    

At its core, Airi uses Jaja’s customer and colleague knowledge base to automatically respond to customer queries, but this is just the start of its power. The solution was designed to be extensible, allowing for the integration of additional context document inputs in the future. Key elements of the solution include:

• Amazon Bedrock and Claude LLMs to power Airi’s GenAI capabilities
• Retrieval-augmented generation (RAG) to ensure high accuracy in AI-generated responses
• Seamless integration with Jaja’s existing tech stack via the new REST API  

Jaja’s operations team and quality assurance specialists conducted rigorous and iterative testing, which quickly identified areas for further improvement. The solution’s extensibility was enhanced by the development of test frameworks and a testing UI, reusable Terraform scripts, and easy-to-maintain context files.

Integrating Airi into Jaja’s existing technology stack allows for robust monitoring and reporting capabilities, as well as relatively low run costs. Jaja continues to regularly monitor and assess Airi’s progress ensuring its continued accuracy and intelligent near-human responsiveness. 

After implementing Anthropic’s more advanced large language model, Claude 3.5, Jaja saw a marked improvement in Airi’s grasp of nuance, humour, and complex instructions, while delivering exceptionally high-quality content with near-human responsiveness.

Additional guardrails for hallucinations have also been implemented to ensure customers receive the highest level of accuracy when interacting with the chat assistant. The improved process gives Airi access to more information so even more filters can be added and used, increasing its overall efficiency for both customers and colleagues, including its ability to take more account actions.

During testing, Airi achieved high levels of accuracy. Response times ranged from 7 to 14 seconds, with 80% accuracy across three measures (correctness, completeness, and readability).

When Airi launched, Jaja predicted that the chat assistant would handle 30% of routine customer interactions, freeing its customer service teams to devote 30-40% more time to complex, high-value customer interactions that require human expertise. 

The company also predicted a 65% reduction in customer response times, but three months later, Jaja announced that Airi had in fact reduced customer response times by 90%. The digital lender can now answer customer enquiries in just under 15 seconds, compared to three minutes before Airi.

In just under six months, Airi has already engaged in over 61,000 unique customer conversations and independently resolved over 25% of general enquiries.  

“Airi continues to massively exceed our expectations in terms of delivering the superior experience our customers expect,” says Di Costanzo. “Our colleagues are also seeing the benefits of its advanced capabilities by being able to spend more time dealing with customers’ complex enquiries, but also by helping our customers use their credit better.” 

 “This is only the start of our journey with GenAI,” Di Costanzo adds. “We have already made, and will continue to make, many more advancements by optimizing its technical expertise, for both our customers and our colleagues.”