Lloyds Bank and Coinbase: Inside Their Generative AI Playbooks

Lloyds Banking Group, one of the U.K.’s largest financial services firms, is planning to launch a consumer-facing AI agent as early as August.

That’s according to Lloyds Chief Data and Analytics Officer Ranil Boteju, who said during a recent Google Cloud roundtable discussion on AI in finance that it represents a fast turnaround for a company at which agentic artificial intelligence (AI) was still a concept last summer.

“The really amazing thing for me is agentic has gone from literally PowerPoint slides 12 months ago to MVPs [minimum viable products]” today, Boteju said. “That for me, is quite amazing, the pace of how this happened.”

How was Lloyds able to pull it off, in a highly regulated industry, no less?

Boteju said Lloyds — which owns Lloyds Bank, Bank of Scotland and Halifax — used to have an on-premises machine learning platform that was about a decade old, until it decided to move to the cloud with Google. This let it tap AI models and tools to power a “significant step up” in capability, he said.

Starting last summer, Lloyds decided to focus on AI agents and build an underlying architecture that can be deployed for different use cases, whether it’s for financial advice, software development, claims or underwriting.

Boteju believes there are at least 50 use cases for AI agents they could develop. “We want to enable the whole bank with AI.”

Read more: 72% of Finance Leaders Use AI in Their Operations

Lloyds worked with Google’s engineering team in a 12-week sprint, the result of which was a working prototype — an intuitive, intelligent AI agent that interacted directly with customers to give financial tips and guidance.

But Lloyds wanted to mitigate risks, so it started with a less risky use case: giving advice on topics like debt consolidation and how to save. The next step would be to gradually build in more advice capabilities.

This AI agent architecture is also being tapped for other uses. Lloyds’ engineers, for example, used AI agents to make the process of building data products easier to use and more intuitive, Boteju said.

Lloyds’ strategy around generative and agentic AI consists of the following:

• Lloyds is building a bank-wide workbench to provide agent-building capabilities to employees across the bank. It means everyone will have the same set of capabilities.
• Employees will have a central set of guardrails for those AI agents.
• Lloyds is upskilling a core group of 350 senior business leaders as well as technicians on AI. The goal is to motivate them to reimagine their business units, not just stick with the same old procedures and sprinkle a bit of agentic AI on top.

Boteju said the question for business leaders is, “if you have these tools and a blank piece of paper, what would [your unit] look like?”

Rajarshi Gupta, head of AI at cryptocurrency exchange Coinbase and part of the Google Cloud roundtable, said the company decided to “invest heavily” in generative AI as a platform around two years ago, soon after OpenAI released GPT-3.5.

Coinbase has since launched the following initiatives:

• Search, help and chat powered by generative AI for customers. It can answer questions, access their account information and route calls to a human agent. Coinbase uses Gemini for this capability.
• An AI-powered employee assistant that can exist as their own avatar on Slack. It has all their information and access to all their documents. Employees just have to ask their avatar questions. Gupta said this AI persona capability is “heavily used” at Coinbase.
• Built an underlying platform called Coinbase-GPT for both the customer- and employee-facing AI assistants that facilitates security, infrastructure and billing across multi-cloud environments. This platform also lets employees access different capabilities as APIs.

Gupta said he’s particularly happy that the quality of “cheap and small models” from Gemini have “gone up a lot.” That’s because “as you start building these more complex use cases, you realize that sending everything to the top-end LLM is very expensive and has a lot of latency,” he said.

What Coinbase does is send simpler queries to cheaper, smaller models and only what’s necessary to the top-end large language model (LLM), such as complex queries. “Gemini in particular has been phenomenal in in making the smaller models faster and cheaper,” Gupta said. (Coinbase also uses AWS and Microsoft Azure.)

The result of generative AI becoming mainstream is that AI used to be the domain of Gupta’s tech team. Now, every Coinbase department including HR and marketing are thinking about how they can use AI in their operations.

“It’s wonderful to see that every part of the company is trying to use AI to automate their journeys and make it more efficient,” Gupta said. “I think that is going to be the lasting effect of this revolution.”

Read more: Why AI Is Becoming the ‘Pacemaker’ of Company Finances

Both Lloyds and Coinbase said they proceeded cautiously with generative AI because of its problems with hallucinations — the tendency to make things up and state these things with confidence.

However, Boteju and Gupta said they are much more comfortable about their guardrails today than they were a year or so ago.

“That was something we were quite concerned about, probably for the first 12 or 18 months,” Boteju said. It was decided that “until such time as we have confidence in the guardrails, we will not expose any of the generative AI capabilities directly to customers.”

Initially, Lloyds focused on back-office efficiencies as the first use cases, or they had a human in the loop to keep an eye on activities.

Meanwhile, the engineering teams learned to use different techniques to ensure accuracy. These include fine-tuning the data, using RAG and creating a “reviewer” agent that double-checks the answers before they go to the customer.

“The combination of those things gives us a lot more confidence that we can start to expose these capabilities directly to customers,” Boteju said. “The confidence is high, and we want to start the process of exposing this to customers and learning as we go.”

Coinbase addresses hallucination and privacy concerns also by implementing a phased approach.

“You start with internal-facing use cases, then you go to assist mode where there is a user being assisted … with AI, and then you go to AI directly talking to consumers,” Gupta said. “That’s what we did.”

What a difference a year makes. At this time last year, hallucinations “would have been my biggest worry because we were three months away from releasing this to customers,” Gupta said. “Then we managed to get to the point where we were comfortable enough with the guardrails that we did it.”

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.

The post Lloyds Bank and Coinbase: Inside Their Generative AI Playbooks appeared first on PYMNTS.com.