Citigroup has begun its largest test to date of artificial intelligence in the workplace. This month the bank rolled out a pilot of agentic AI inside its proprietary platform, Citi Stylus Workspaces, involving 5,000 employees. The new tools allow staff to direct an AI system to complete multi-step tasks with a single prompt, such as researching a client, pulling data from internal and external sources, and translating the findings into another language.
Chief Technology Officer David Griffiths said the company believes the models are finally ready for this kind of work. “A couple of years ago, you could do agentic things with the early versions of the models that were available then. But they weren’t always very reliable. They weren’t always very good at invoking tools. But they are now,” Griffiths said to Wall Street Journal.
The pilot will run for four to six weeks. Citi is measuring how employees use the system, how much productivity it drives, and whether the costs stay manageable. Running AI agents can become expensive as tasks lengthen and involve more data. Griffiths said the company has built cost controls into the system and will monitor the balance between value and usage.
The platform is built to work with multiple models, including Google’s Gemini and Anthropic’s Claude. Citi has also integrated the tools directly with its own data and systems, such as employee directories and project management software. In practice, that means an employee can compress several separate steps into one request. Griffiths offered an example: a banker could ask the system to profile a client, aggregate internal and public information, and prepare a report in Spanish. all in a single flow.
Building on broader AI investments
The agentic AI pilot is one part of Citi’s larger technology program. The bank spent nearly $12 billion on technology in 2023 and continues to retire legacy applications while modernizing infrastructure. In the past year, it equipped about 30,000 developers with generative AI coding assistants, launched AI-driven advisory tools in its wealth division, and introduced “Agent Assist” for customer service staff. According to the bank, roughly 175,000 employees now have access to some form of AI tool across its operations.
Earlier this month Citi also named Shobhit Varshney, a longtime IBM executive, as its new head of AI. Varshney reports to Chief Operating Officer Anand Selva and works alongside Griffiths to coordinate deployment across businesses. In a memo to staff, Selva said the goal is to “responsibly build AI capabilities that enhance the client and colleague experience, strengthen internal controls and boost internal productivity.”
Citi’s wealth management arm has taken a measured approach to client-facing AI. Gunjan Bhatt, global head of data science at Citi Wealth, recently emphasized that the technology is being used to support advisors rather than replace them. He cited early gains in client satisfaction and advisor productivity from tools such as Advisor Insights, which recommends engagement opportunities, and AskWealth, which automates research.
Competitors advance their own AI tools
Other large banks are moving in the same direction, often at larger scale. JPMorgan Chase has developed an internal suite of large language models and copilots that it says are used by more than 200,000 employees. The bank reports that coding assistants have improved software developer productivity by as much as 20 percent and that AI has contributed to increased sales in wealth management.
Goldman Sachs has deployed a generative AI assistant, known as GS AI Assistant, to thousands of employees across the firm. It is designed to summarize documents, draft communications, and analyze data. Goldman executives have said the technology is already improving efficiency across deal teams and research.
Morgan Stanley has focused its AI program on wealth management. The firm uses assistants that can generate meeting summaries and draft follow-up emails, allowing advisors to spend more time on direct client engagement.
Bank of America has expanded Erica, its AI-driven customer assistant, from consumer banking into more back-office functions. The company says it now processes hundreds of millions of client requests through the system annually.
These initiatives reflect a competitive landscape in which the largest U.S. banks are embedding AI more deeply into daily operations. Each institution is experimenting with copilots and agents that handle routine research, compliance checks, customer service, and risk management tasks.
Workforce and cost considerations
For Citi, the question of workforce impact remains open. Griffiths said the tools could significantly increase capacity, but he stopped short of drawing conclusions about headcount. “Does it mean that we need less people? I don’t know. It certainly means that we would get a lot more done. And we’ll see how the workforce evolves with that massive boost of capacity that we’re getting here,” he said.
The economics are still developing. AI tasks are priced based on “tokens”, units of data processed by the model. Costs can rise quickly for complex, multi-hour tasks, though Griffiths noted that model pricing is falling and that most current tasks finish in minutes. Citi will evaluate the pilot’s cost-benefit ratio before expanding access further.
Banks are under pressure to modernize technology not just for efficiency but also for compliance. Citi’s own modernization efforts accelerated after regulators fined the firm over data management lapses. Updating infrastructure has created a foundation for AI adoption, but also underscored how much ground Citi needs to make up compared to peers.
Industry-wide, the pace of development suggests that AI will become embedded in nearly every layer of bank operations. The near-term gains are in productivity, compliance, and cost management. Longer term, the tools may shape how financial institutions interact with clients and design products.
