US banking giant Citi pilots agentic AI with 5k staff (www.theregister.com)

Citi has begun a six-week pilot of an “agentic” version of its Stylus Workspaces AI platform with 5,000 staff, aiming to automate multi-stage workflows like customer research, profile construction from internal and public data, and foreign-language conversion. The upgraded platform links directly to Citi systems, project-management tools and external web searches, and orchestrates multiple LLMs (Citi says it uses models including Google’s Gemini and Anthropic’s Claude). Technical controls such as input limits have been added to curb runaway compute costs, and the pilot’s explicit goals are to measure productivity gains and calculate ROI while observing how roles evolve as employees are “freed” from manual tasks. The move is significant because it tests agentic AI at scale inside a highly regulated, risk-averse industry where accuracy, cost and controls matter. Citi’s trial highlights both the promise—faster, more complex task execution and consolidated workflows—and the risks: need for human verification, budget controls, and governance. Analysts warn many agent projects fail (Gartner expects >40% cancellations by 2027; researcher-reported task success rates are around 30–35%), and generative AI is entering a “trough of disillusionment.” Still, early adopters who manage costs and risks could capture large efficiency gains in finance, even as the industry debates headcount and oversight implications.