How Citi, Citizens, TD deploy AI agents without upsetting humans

• Key insight: Banks are grappling with how to get employees enthused about agentic AI.
  
• What's at stake: Workforce roles, vendor relationships and compliance expectations are being redefined.
  
• Expert quote: "You must take people along" — Nikhil Joshi, Citi CIO, on reskilling and adoption.

Source: Bullets generated by AI with editorial review

When companies roll out agentic AI — artificial intelligence models orchestrated to not only spit out answers but also execute tasks — it can stir up a range of emotions. People who enjoy using AI in their personal lives may be delighted to use it more at work. People who see their core work being given to bots may fear losing their jobs, sometimes with good reason.

"The cultural element of what this technology means is huge," said Jo Jagadish, head of digital banking and contact centers at TD, at the Reuters AI Momentum conference this week in New York. "The fact that people are using it more and more in their personal lives is creating an ecosystem in your work life where you're really questioning, well, why can't AI do this for me? If I can get the answer a lot more seamlessly in my personal life, my work should mimic that. It's creating a lot of robust dialogue on use cases and innovation, which is great."

On the flip side, some will wonder if AI will replace what they do today.

"That's where as an organization we really need to be a lot more transparent, and redefine what does productivity look like for a specific team?" Jagadish said. "How does AI enable teams to be more effective by reducing the redundant tasks and allowing humans to really focus on what they like, which is human interactions, creativity, decision making? Those are some of the things that we're spending a lot of time defining."

At TD, Citizens, Citi and other banks, leaders are providing AI training, providing virtual assistants employees can use as they like and trying to redeploy employees to more interesting work.

Reskilling

Citizens Bank in Providence, Rhode Island, has rolled out agentic AI in the form of Github Copilot's coding agent to about half its software developers – 800 people. It's also working with fintechs that are using AI to rewrite legacy systems to be moved to the cloud. For instance, the bank's mortgage system, which has millions of lines of code, was rewritten and moved to the cloud in six weeks.

"We're finding that gen AI in general is really good for code conversion," Michael Ruttledge, chief information officer, told American Banker in an interview.

Citizens also uses agentic AI to automate complaint handling. AI agents classify incoming complaints and respond to customers. And the bank uses agentic AI to automate quality control. Where typically humans review about 5% of call center transcripts for accuracy and tone, Citizens is currently piloting agentic AI to check answers and conduct sentiment analysis for all calls, instantly.

For these and other use cases, Citizens Bank has been educating and reskilling employees to help them adjust to these new ways of working. It uses generative AI to create video tutorials.

TD Bank has been using agentic and generative AI to provide better knowledge management to contact center agents, according to Jagadish. Most agents — 80% — started using the technology the first month it was available. TD is now rolling it out in branches and working to make it more predictive, so that when a customer calls, the agent knows what the customer was trying to do in online banking and in other channels that resulted in a call to the contact center.

Before TD rolled out generative and agentic AI in its contact centers, it started communicating with and teaching people about the new tools months in advance, "creating excitement of how this enables you to do your job better and faster and more effectively," Jagadish said. "That has been a really great way to set the stage and the foundation of our colleagues actually embracing this technology."

Nikhil Joshi, chief information officer, markets technology at Citi, made a similar point at the Reuters AI Momentum conference.

"Whenever you look at large scale automation in general, especially when you look at something like AI, the most important element is the human element," he said. "So you have to make sure that you take the people along with you on the journey. And that involves reskilling, unlearning, relearning."

Citi has rolled out AI training for more than 175,000 employees. "This is not just some one-size-fits-all, here's an LLM, here's AI type thing," Joshi said. "This is an adaptive course that is applicable to pretty much the entirety of the population of the bank. And it assesses, as you go through, what you know, what you don't, and it adapts the complexity depending on what needs teaching. This is one of many things that we're doing, but we as an industry need to think about these kinds of things to make sure we meet our people where they are."

Reducing third-party reliance rather than reducing staff

So far, Citizens Bank has been cutting back its use of outsourcing providers in software engineering and contact centers, rather than laying off software engineers and customer service agents, Ruttledge said.

"If you look at our contact center today, we have a very large vendor population that supports our contact center," he said. "So it's those that we're phasing out, rather than our own colleagues. So we message some of that."

Presenting AI as a helpful copilot

When he's implementing AI in a bank, Daniel Jameel, founder and CEO of Saris AI, said employees are not always ready for it.

"They're unsure: What is this? Is it going to impact my work? Is it going to impact my job?" Jameel said.

He presents Saris AI a copilot. "We make you look like a rock star, and we take away the work that you don't enjoy yourself doing," Jameel said. "I think that resonates really well with the rest of the team, because they can focus on the high-impact work."

At Citizens Bank, "we're trying to introduce [AI] into the colleague base so they get excited about it," Ruttledge said. That bank has given all employees the use of Microsoft Copilot.

Giving AI the boring work?

A common AI trope is that artificial intelligence is going to take care of all the boring work and leave humans with creative, fun and interesting things to do.

At Citi, which has rolled agentic AI out to its 40,000 developers, the AI agents, created using Cognition's Devin software, handle tasks that are deterministic and easy to test, such as code reviews and middleware upgrades.

The agents have conducted more than a million code reviews, he said. They have also generated documentation of the bank's legacy code base, a task that surely no human likes to do. The AI agents have been fed the bank's coding and engineering standards and they are always overseen by a human in the loop, he said.

"The agents go do what they have to do, and the human developer oversees once the agents are done with their task, as they would with any other human developer," Joshi said. "This obviously makes things very efficient, but also it saves a lot of time for developers to focus on what we consider business priorities."