A Citibank branch in the headquarters of Citigroup Inc. on Park Avenue in New York

Corbis via Getty Images

Citi operates at a scale few institutions can match, serving clients across nearly 180 countries and moving trillions of dollars daily. For Shobhit Varshney, the firm’s Global Head of Artificial Intelligence, that scale defines both the opportunity and the challenge of deploying AI effectively.

Varshney joined Citi in September 2025 to lead its enterprise-wide AI strategy. His mandate is grounded in a philosophy he summarizes simply: “doing the right AI and doing AI right.” At a global bank, that distinction proves essential.

Value First, Then Technology

Varshney emphasizes that AI at Citi does not begin with technology. It begins with value. “Doing the right AI is focusing on value creation,” he emphasized. “You have to re-engineer processes end-to-end, not just automate at the edges.”

That approach requires simplifying or eliminating unnecessary steps before introducing automation. Rather than improving legacy workflows incrementally, Citi pushes teams to rethink them entirely. The goal is not marginal efficiency gains but meaningful transformation, often targeting improvements of 30% or more. This philosophy reflects a broader shift away from experimentation-first models. Instead of launching isolated pilots and searching for use cases, Citi defines the desired business outcome upfront, quantifies the value and aligns leadership around it before building solutions.

Building AI at Scale

To support that ambition, Citi has developed a centralized AI platform designed to scale across the enterprise. The platform is modular, multi-model and multi-cloud, enabling flexibility while maintaining control. Varshney explained that many AI use cases share common patterns. “Quite often, we see a lot of use cases that have common patterns,” he noted. “We should provide those as-a-service centrally.”

This approach reduces duplication and accelerates deployment. Instead of each business unit building its own tools, Citi develops shared capabilities that can be reused across functions. As models improve, enhancements propagate across the entire organization. The result is both speed and consistency. More than 182,000 employees now have access to AI tools, with over 70% actively using them, a level of adoption that reflects the firm’s deliberate focus on usability and governance.

A Dual Approach to Adoption

Citi’s AI strategy combines two complementary approaches: bottom-up and top-down. On the bottom-up side, the firm provides productivity tools such as its

internal large language model gateway, enabling employees to integrate AI into daily workflows. These tools help with tasks ranging from summarization to software development, where Citi has already achieved significant productivity gains.

On the top-down side, leadership focuses on re-engineering the most critical processes across the organization. These include complex workflows such as client onboarding, anti-money laundering investigations and risk management. Every week, senior executives review progress on these initiatives, addressing bottlenecks and ensuring alignment. This level of executive engagement helps move AI efforts from experimentation to enterprise-wide impact.

Culture as the Multiplier

Varshney believes that technology alone is insufficient without a corresponding cultural shift. Citi has invested heavily in fostering an AI-first mindset across its workforce. The firm mandated generative AI training for employees and created a network of approximately 4,000 AI “accelerators and champions” embedded across business units. These individuals help colleagues adopt AI in practical ways, reinforcing learning through peer engagement rather than top-down instruction.

“People learn from people two desks over,” Varshney said. “That’s how you create real change.” This grassroots approach complements broader initiatives such as AI-focused events and internal showcases, helping to normalize experimentation and accelerate adoption.

Responsible AI as a Foundation

In a highly regulated industry, responsible AI is not optional. Varshney stresses that governance must be embedded from the outset. “Doing AI right starts with responsible AI,” he underscored, emphasizing the need for ethical principles, risk controls and cross-functional collaboration.

Citi has extended its existing risk management frameworks to address the complexities of generative and agentic AI. This includes coordination across technology, risk, compliance and audit functions, as well as the establishment of an AI risk committee. By involving stakeholders early in the development process, Citi reduces friction later and ensures that solutions meet regulatory and operational requirements from day one.

From Automation to Reimagination

One of Varshney’s recurring themes is the need to move beyond incremental automation toward true process reimagination. He likens the shift to the introduction of dishwashers. Rather than replicating manual dishwashing faster, the technology changed the process entirely, eliminating steps and enabling scale. “The mindset shift is from a faster scrubber to a dishwasher,” he said. This analogy captures Citi’s broader ambition: to rethink how work is done, not just to do it faster.

Looking Ahead: Agents and Beyond

Looking forward, Varshney sees significant potential in agentic AI, where systems can take on more autonomous roles within defined guardrails. While current implementations maintain human oversight, he expects autonomy to increase over time.

He also points to advancements in software development as an area of continued impact, as well as the longer-term convergence of AI and quantum computing.“We’re only at the start of this journey,” he said.

At Citi’s scale, even incremental improvements can have outsized effects. But Varshney’s focus remains on transformation rather than optimization, ensuring that AI reshapes how the organization operates at its core.

Peter High is President of Metis Strategy, a business and IT advisory firm. He has written three bestselling books, including his latest Getting to Nimble. He also moderates the Technovation podcast series and speaks at conferences around the world. Follow him on X @PeterAHigh.