According to Fortune, Citigroup serves clients in over 180 countries with physical presence in more than 90 markets, and they’re leveraging this global footprint as their AI advantage. CFO Mark Mason revealed that every Citi business has identified five specific AI use cases to develop, with teams meeting weekly to share progress and review risks. The bank has already rolled out AI tools like Citi Assist and Citi Stylist to approximately 180,000 of its 229,000 employees, saving thousands of hours on tasks from code review to document summarization. Mason emphasized that robust controls and human oversight remain essential for critical financial reporting, with the bank conducting regular stress testing as part of planning. The CFO personally spent over 30 minutes on AI prompt training, noting that “almost all the power is in the prompt” when it comes to quality output.
Sponsored content — provided for informational and promotional purposes.
The global moat
Here’s the thing about Citi’s strategy – they’re betting that their physical presence across 90+ countries creates a data advantage that pure digital players can’t easily replicate. Mason basically said multinationals aren’t retrenching despite geopolitical tensions, which means they’ll still need cross-border banking support. And having local leadership from those markets? That’s been “incredibly valuable” for maintaining credibility. It’s a classic case of an old-school advantage becoming newly relevant in the AI era.
The messy middle of AI implementation
What’s really interesting is how Citi is approaching this rollout. They’re not just throwing AI at everything and hoping it sticks. Every business unit has exactly five use cases – that kind of discipline prevents the “shiny object” syndrome that plagues so many corporate AI initiatives. And the weekly progress meetings? That’s where the real work happens. It’s one thing to identify use cases, but actually making them work across different regulatory environments and business cultures? That’s the hard part.
I’m struck by Mason’s balanced approach. He’s “excited but cautious” about AI in finance, which honestly feels like the right tone. They’re moving fast – 180,000 employees already using AI tools is no small feat – but maintaining dual processes and human oversight for critical reporting. That’s the kind of responsible innovation you want to see from a systemically important bank.
The prompt training reality check
When the CFO himself spends 30+ minutes on prompt training, that sends a message. Mason’s observation that “almost all the power is in the prompt” is something more companies need to internalize. We’re seeing this across industries – the companies getting real value from AI aren’t just buying fancy tools, they’re investing in teaching people how to use them effectively. And his point about subject matter expertise being essential? That’s the hidden cost many organizations underestimate.
The returns conversation is equally realistic. Some use cases drive revenue, others just save time – and that’s okay. Not every AI investment needs to be transformative. Sometimes eliminating manual work is valuable enough. This pragmatic approach might not make for exciting headlines, but it’s probably why their AI initiatives are gaining traction without the usual hype cycle burnout.
What this means for everyone else
Look, if a global bank with nearly 10,000 finance professionals can move this deliberately with AI, what’s stopping smaller organizations? The playbook is becoming clear: start with specific use cases, maintain human oversight, invest in training, and measure returns realistically. It’s not about replacing people – it’s about amplifying what they can do.
This aligns with broader trends we’re seeing in corporate AI adoption where companies that combine technology with strong process discipline are pulling ahead. The real test will be whether Citi can maintain this disciplined approach as AI capabilities advance and competitive pressures increase. But for now, they’re showing what responsible AI transformation looks like at scale.
