Lloyds Banking Group is building generative and agentic AI models using Google's technology within Google Cloud, after a yearlong migration from legacy systems.
"We want to enable the whole bank with AI," said Ranil Boteju, chief data and analytics officer at Lloyds.
The London bank had an on-premise, legacy machine learning and data science stack that was about 10 years old, Boteju said during a call on Friday.
"It was incredibly modern at the time," he said. "And then a few years ago, we realized we needed to modernize this. We wanted to move to the public cloud. And so we selected Vertex AI as our machine learning platform."
Lloyds has moved 15 modeling systems from its on-premise infrastructure to Google's Vertex AI on Google Cloud, which according to the bank has reduced the estimated lifetime carbon footprint of its machine learning platform by about 27 tons. The new AI development platform has also enabled the bank to pursue its agentic AI aspirations, Boteju said.
Vertex AI is Google's platform for building, deploying and managing generative AI models and AI agents, which it debuted in 2021. It provides access to more than 200 generative AI models, including Google Gemini, Anthropic's Claude and open source models like Meta's Llama, according to Stuart Moncada, head of product, cloud AI agents and search at Google Cloud.
"At Google Cloud, we believe in offering a wide choice of models so that customers can pick the right model for their specific use case," Moncada said.
Boteju said this was part of what drew him to Vertex.
"We can use Gemini when we want, but also the fact that we can bring in open source models, or any number of different models, that is incredibly useful and powerful for us," he said. "It means that we can select the right element for the right task."
The fact that Lloyds is using a variety of models makes sense in this rapidly changing AI technology landscape, said Steve Rubinow, a professor in the department of information technology and management at Illinois Institute of Technology and former chief information officer at the New York Stock Exchange.
"We can't bet on a single horse — it's way too early," Rubinow said. "We need to have access to all the popular models — proprietary, open source — and we have to see which one works better for us. And, by the way, next month that ranking may change. If we build a flexible architecture, we will be able to incorporate the best results."
During the first year of building generative AI models, Boteju and his team were concerned about the potential for hallucinations and privacy violations.
"What we said was, until such time as we have confidence in the guardrails, we will not expose any of the generative AI capabilities directly to customers," Boteju said. "So our initial set of use cases were very much focused on the back office, or they had a human in the loop as an additional step of risk mitigation. But over that period, we have started to strengthen the guardrails."
The bank used retrieval augmented generation to limit the data on which the gen AI models would be trained. It also set up agents to review other agents' work.
"The combination of those things gives us a lot more confidence that we can start to expose these capabilities directly to customers," Boteju said. "And that's effectively where we are in that we now have guardrails. The confidence is high, and we want to start the process of exposing this to customers and learning as we go."
Since deploying the new platform on Vertex AI, Lloyds has initiated more than 80 new machine learning use cases and launched more than 18 generative AI systems into production, spanning its entire business, the bank said. Twelve more generative AI systems are expected to go live by the end of June.
Use cases include customer advice, claims and underwriting.
"If I think back to last summer, agentic AI was just a concept for us," Boteju said. "It was just emerging and we tried to understand it." His team worked with a Google team to create a minimum viable product with agentic AI capability to provide financial tips and guidance to customers, such as advice about debt consolidation and how to save.
"Obviously, we're working through that very carefully with regulation," Boteju said. "We're now in the process of building something that we want to expose to customers. We're hoping to have something by August or September of this year."
Engineers across Lloyds Bank are now starting to use these capabilities for internal use cases, too. For example, the bank has an automated process for building data products. Recently, engineers started building on top of that an agentic approach that makes that process easier to use and more intuitive, and reduces a lot of the workload, Boteju said.
Toby Brown, head of global banking solutions for Google Cloud and previously a technology leader at Wells Fargo, said Google is seeing strong demand for generative AI from financial services.
"We're seeing a lot of external forces, whether that's continued margin compression due to changing industry economics, intensifying competition between the incumbents and the fintechs, or heightened digital expectations as customers continue to live more of their financial lives online and really on their phones," Brown said.
On Wednesday, Google unveiled a new agent development kit. The company is also releasing an agent-to-agent interoperability protocol that will make it possible for agents built on different frameworks and vendors to work together, Moncada said. Google is working with more than 50 enterprise software companies including Salesforce and ServiceNow to enable their agents to collaborate and communicate.
Rubinow, for one, gives Lloyds' latest AI moves high marks.
"They're doing everything that you would expect an organization like Lloyds to do," Rubinow said. "They're experimenting, they're running pilots, they're being careful, because the mistakes are costly."
