This initiative is the latest phase of
The expanded collaboration, which the companies made public on Tuesday, is significant because
"Leading banks garner attention for their AI efforts, and those pursuing agentic AI at scale are clearly on the leading edge," said Dan Latimore, chief research officer at The Financial Revolutionist, told American Banker. "Most banks should be watching and learning from these leaders so that they can pick the most effective large language model strategies and, over time, implement agentic AI as the bugs are worked out."
While many banks have rolled out generative AI models like OpenAI's ChatGPT or Anthropic's Claude to employees, few have launched agentic AI across the enterprise, though some have begun using agents in contained areas such as software development.
"Our goal overall is to get generative AI tools in the hands of all of our employees,"
The technology rollout is about efficiency, but it's also more than that, Kerrins said.
"We absolutely think that the tools will help our employees do their jobs better, more efficiently, but probably more importantly, provide greater insights and intelligence to the work they're doing so they can spend more time on the higher order functions," Kerrins said.
NotebookLM and Deep Research
Google's NotebookLM research assistant will let employees put all their documents, presentations and reports in one place and "have access to information at a scale they haven't before," Kerrins said. "We think this is a real game changer for our employees, but also just overall, to the business."
Deep Research lets users ask questions in natural language. The model provides a response based on information the user has allowed it to access. If a model has been given access to a specific set of internal documents, its answers will be limited to what's in those files. Or a user might give it access to that plus the whole of the internet, with citations.
"The idea here is, think like a research scientist: what are the research steps that you would want to do?" Rohit Bhat, general manager and managing director, financial services at Google Cloud, told American Banker. For instance, a researcher might want to know the impact of the latest information on tariffs based on last week's information and news, and how he should contemplate that moving forward.
"It's a pretty broad question," Bhat said. "Out of that question might come hypotheses. So the first step in the planning process for this agent would be creating hypotheses. Each hypothesis has to then get tested. You collect information. That multi-step process is what a Deep Research agent is very good at doing. It becomes this massive productivity unlock when you're trying to do things like investment research or wealth advice research internally. These are very, very complex tasks that otherwise would take a long time and would take a lot of multiple cycles, also prone to error, and it kind of brings it into one view for you."
When a generative AI research tool is crawling bank documents and policies, one challenge for some is making sure all those documents are accurate and up to date.
"That's why we're doing it slowly and making sure that we have all of the right controls in place," Kerrins said. "Policies and procedures aren't as much of a concern there; when we publish a new policy or update an existing one, it goes through a pretty rigorous process. So once it's published, we know it's published." But with other documents, the bank is putting controls and monitoring in place to ensure the software doesn't access data it shouldn't.
Deep Search and NotebookLM have been deployed to around 2,000 employees so far.
Making the tools available to all staff will take a few months, Kerrins said.
Employees could use this to make contract checks to make sure all the bank's policies were followed. This could further be used to help create credit memos.
"When a contract is 80 pages long, and you've got a person whose sole responsibility is to make sure that KYC clauses are appropriately identified, that personnel has to find that clause in the 80-plus pages and then say in the memo, yes, this was good," Kerrins said. "You still want the human in the loop, but that work of reading those 80 pages is where these agents can come into play. "
Over time, this tool might help with vendor risk management, and making sure the bank is consistent about security, data sharing, KYC compliance and other possible points of friction with vendors.
"That's the power of agentic technology, full stop," Kerrins said. "It is a 24/7, 365, digital workforce. You can have these digital agents looking at anything you want in real time, looking for changes, anomalies, patterns, differences. That's the greater insight you get, and then you can act on those insights, it makes us all better, and hopefully it provides our employees with more information so they can do their jobs better, and they can service customers and clients better as well."
Creating virtual assistants in Agentspace
Using Agentspace, the bank plans to give virtual assistants out to staff throughout the company.
"They'll log in and they'll interact with it, just like an enterprise virtual assistant, and they'll have access to any agents we decide to publish," Kerrins said.
Branch bankers and customer service reps will get agents that help them better serve customers.
Some agents even will be allowed to interact directly with customers, for instance, to handle debit card reordering.
In the corporate and investment bank,
It also plans to create an agent for contract management, which should help with the roughly quarter of a million documents related to vendor agreements that the bank handles every year. A custom agent can rapidly query these extensive documents — identifying contracts with specific clauses, payment terms, contract types and other details.
"Eventually, when we're ready and it's safe and we believe we can control it, we'll allow employees to create some of their own agents that will still be monitored and controlled. That has a higher risk profile and we're not quite ready to allow employees to start creating their own agents."
Kerrins does not expect to deploy agentic AI agents that operate autonomously like virtual employees without human intervention any time soon.
"The technology is still new, and we want to be really careful," Kerrins said. To date, agents are providing information that humans review and decide if they agree. "It'll be a hybrid, I think, for a while, especially on higher-risk case scenarios. The technology is not going to be the limiting factor. I think it'll be the complexity and the risk level of the use cases."
