Skip to main content
AmigoFish

How Starling Bank's AI serves 5M customers and boosts canceled scam payments 300%

From Google Cloud · · 2 min

The player loads from YouTube only when you click. It uses YouTube's privacy-enhanced mode.

If the video does not play, the owner may have turned off embedding. Watch it on YouTube instead.

Watch on YouTube

Not rated yet

Description

*Featured in this video:* Frédéric Laurent, deputy chief information officer at Starling Bank

*Executive summary:* Digital-only Starling Bank developed its Starling Assistant — the UK’s first agentic banking assistant — to help nearly five million customers easily manage their money, find and use app features intuitively, and set savings and budgeting goals. Built using Google’s Gemini, open-source Agent Development Kit (ADK), BigQuery, and robust, reusable evaluation pipelines, Starling Assistant combines generative AI reasoning with action-oriented agentic capabilities. Starling next took advantage of its strong data and AI infrastructure to create the Scam Intelligence tool in just four weeks to save customers money and frustration by detecting fraudulent advertisements and reducing scam payments by 300%. That’s in addition to the Spending Intelligence tool and internal-facing AI applications that help employees serve customers better. And Engine by Starling is a platform plus blueprints that other banks can use to meet digital customer needs. Starling’s team members have saved 8,000 hours per month with these Google-powered innovations, and the Gemini platform provides the agentic guardrails needed to meet regulatory and governance requirements.

*Challenge:* As a digital-only bank in a highly regulated industry, Starling knew its users needed practical financial support that was easy to access and helped them manage complex tasks easily. The bank wanted a secure, highly available way for customers to get the help they needed using conversational, multimodal AI, while also maintaining strict regulatory guardrails and deterministic transaction safety.

*Solution:* Starling Bank built Starling Assistant and its other AI tools using Google Cloud’s Gemini, BigQuery, and the ADK toolkit for creating smart AI agents. The assistant goes beyond basic chatbot functionality to include agentic qualities, and serves as an active helper to both retail and business customers. Starling integrated Scam Intelligence into the assistant as a specialized sub-agent, so customers can now evaluate advertisements or potential purchases using multimodal features before making a payment.

*Results:* The Starling Assistant helps customers with a range of financial questions, such as comparing spending across time periods, creating a budget, setting savings goals, and more. Starling customers using Scam Intelligence have canceled suspected fraudulent payments four times as often as those not using it. Customers interacting with Starling Assistant also experience shorter average handling times when contacting support, and the bank has seen a 50% reduction in customer service referrals after launching the assistant. Starling developers have saved many hours of time using the combination of ADK, templates, and trusted Google-built pipelines.

*Key takeaways and highlights from our interview with Frédéric Laurent, deputy chief information officer at Starling Bank:*

→ “Historically, banks tend to use technology to reduce costs, but not necessarily to improve customer experience. The main motivation of Starling Bank is to help customers manage their money better using technology.”

→ “A lot of chatbots out there don’t even use AI, in my opinion. They’re just a fancy decision tree with a very low-quality understanding of natural language. We don’t think about the Starling Assistant as a chatbot. We genuinely think about it as an assistant, combining the thinking capabilities of gen AI and the doing capabilities of agentic AI, so it can carry out tasks on behalf of the customer.”

→ “We never saw the regulatory and governance angle as being a challenge, because we’ve invested in what we think are critical building blocks using the technology that Gemini platform gives us: the ML pipelines and the ADK. You can prove the models perform at the right level and quickly evaluate the agents’ guardrails as needed.”

→ “The reputability of the evaluation pipelines…

Back to Google Cloud · Report a problem